igneum-network/igneum
1
0
Fork 0
Code Issues Pull requests Releases Activity

F10 rows, I05 epoch tag, I06 pack geometry rule across the three hosts

Browse source 
Rows: docs/analysis/review-2026-10-08-b/f10/rows.md with the pod rows (5090: steady
70.4 MH/s select against 63.7 full, 86.2 against 88.7 J per 2^24 batch; 3090 at the
card's clock 30.1 against 27.6, 184.6 against 197.9 J; the 3090 pod's power cap at
210 MHz named per run; ptxas 92 and 87 registers, no spills) beside the kernel timer,
the emulation and loader logs, the mini's geometry and serve logs. serve-bench.sh reads
a kit pack's seeds from program.h, execs the worker in its coproc (the recorded pid
and the CPU seconds are the worker's).

I05: every worker's found line carries a fifth field, the epoch the hash was computed
under (16 hex): CUDA and OpenCL from the pair, Metal from the job; select-check.sh and
serve-check.sh assert it, the swap case with pack B's epoch. The miner reads four or
more fields.

I06 (gate INT-10): one normative geometry rule. packfile.h reads the word count at
64 bits and refuses a count above 2^32 - 1 and a log2 that is not the floor of log2 of
the count; proto-metal/main.swift gains packGeometry with the same rules and refusal
words, sizes the day's dataset from the exact word count before any Metal work, and
--geometry-check <vectors> [pack dir...]. emu/geometry-vectors.txt is the shared
conformance file (ds55, ds85, ds115, the power-of-two sizes, wrong count or log2,
overflowing counts, missing leaves, leaves on a v2 pack): 19 vectors green in C on
build-7 (packfile-test.sh) and in Swift on the mini, with the two checked-in packs.

Record: cells harness:cuda-select-check (ROT-01, POW-01 partial) and
harness:pack-geometry (POW-02 partial); batch f10-worker-20261008.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-08 19:53:24 +00:00
parent bd8223cfb5
commit ed93b5426f
38 changed files with 892 additions and 37 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
docs/analysis/review-2026-10-08-b/f10/3090-full-r1.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.710 prod_wall_s=58.497 prod_mhs=9.18 prod_mean_w=164.9 prod_j_per_batch=301.4 prod_mh_per_w=0.0557 samples=234 steady_wall_s=53.787 steady_mhs=9.98 steady_mean_w=168.3 steady_j_per_batch=282.9 steady_mh_per_w=0.0593 samples=215 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 1630.34 ms, select 0.000 ms, read-back 24.723 ms, scan 24.119 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/3090-full-r2.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.231 prod_wall_s=90.096 prod_mhs=5.96 prod_mean_w=134.4 prod_j_per_batch=378.4 prod_mh_per_w=0.0443 samples=360 steady_wall_s=85.865 steady_mhs=6.25 steady_mean_w=136.9 steady_j_per_batch=367.3 steady_mh_per_w=0.0457 samples=343 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 2637.86 ms, select 0.000 ms, read-back 22.784 ms, scan 20.625 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/3090-full-r3.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.110 prod_wall_s=23.537 prod_mhs=22.81 prod_mean_w=282.4 prod_j_per_batch=207.7 prod_mh_per_w=0.0808 samples=94 steady_wall_s=19.428 steady_mhs=27.63 steady_mean_w=325.9 steady_j_per_batch=197.9 steady_mh_per_w=0.0848 samples=77 worker_cpu_s=5.61 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 557.15 ms, select 0.000 ms, read-back 24.225 ms, scan 23.753 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/3090-full-r4.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.397 prod_wall_s=90.051 prod_mhs=5.96 prod_mean_w=134.3 prod_j_per_batch=377.9 prod_mh_per_w=0.0444 samples=360 steady_wall_s=85.654 steady_mhs=6.27 steady_mean_w=136.7 steady_j_per_batch=365.9 steady_mh_per_w=0.0459 samples=342 worker_cpu_s=5.81 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 2624.25 ms, select 0.000 ms, read-back 27.536 ms, scan 22.950 ms; 32 jobs 32 chunks

4
docs/analysis/review-2026-10-08-b/f10/3090-kernel-bench.log Normal file
View file

@ -0,0 +1,4 @@
info igneum-worker-cuda 1.0 (4 October 2026): device 0 NVIDIA_GeForce_RTX_3090 (sm_86, 82 SMs), driver 13.2 from libcuda.so.1, NVRTC 12.8 from libnvrtc.so.12, target sm_86 (the device's architecture, listed by NVRTC)
pack /root/p02/packs/hl-v6-all on NVIDIA_GeForce_RTX_3090: nvrtc 2558 cache 3 dataset 37 hot 0 check 1190 race 0 ms variant base class v5 (state leaves 93, uploaded for the build and freed); self-test PASS (cache head, last line and FNV-1a 64 448274a57f508cbc; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes)
warm-up dispatch (base 0): 554.01 ms; 5 timed dispatches of 16777216 nonces: mean 557.06 ms
RESULT pack=/root/p02/packs/hl-v6-all class=mx8-erad810f22d+sh256x27+state+reg64c+fold+rw device=NVIDIA_GeForce_RTX_3090 arch=sm_86 regs=87 blocks_per_sm=16 warps=0 resident=1312 arena_mib=0 hot_mib=0 hot_slots=0 hot_fill_ms=0.00 nonces=16777216 batches=5 check=PASS fingerprint=59e6708e46f1e87c mhs=30.117 loads=128 bytes=512 scratch_ops=0 time=wall

10
docs/analysis/review-2026-10-08-b/f10/3090-run-more.log Normal file
View file

@ -0,0 +1,10 @@
== 3090-full-r3 2026-10-08T19:44:29Z sm 210 MHz, Not Active
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.110 prod_wall_s=23.537 prod_mhs=22.81 prod_mean_w=282.4 prod_j_per_batch=207.7 prod_mh_per_w=0.0808 samples=94 steady_wall_s=19.428 steady_mhs=27.63 steady_mean_w=325.9 steady_j_per_batch=197.9 steady_mh_per_w=0.0848 samples=77 worker_cpu_s=5.61 worker_sha=9b36b523d0812c3c
3090-full-r3 sm MHz mean 1769 min 210 max 1965 over 103 samples
== 3090-select-r3 2026-10-08T19:44:55Z sm 1950 MHz, Not Active
card=NVIDIA_GeForce_RTX_3090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.297 prod_wall_s=83.427 prod_mhs=6.44 prod_mean_w=138.7 prod_j_per_batch=361.6 prod_mh_per_w=0.0464 samples=334 steady_wall_s=79.130 steady_mhs=6.78 steady_mean_w=140.1 steady_j_per_batch=346.4 steady_mh_per_w=0.0484 samples=316 worker_cpu_s=4.23 worker_sha=9b36b523d0812c3c
3090-select-r3 sm MHz mean 318 min 210 max 1950 over 343 samples
== 3090-full-r4 2026-10-08T19:46:24Z sm 210 MHz, Active
card=NVIDIA_GeForce_RTX_3090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.397 prod_wall_s=90.051 prod_mhs=5.96 prod_mean_w=134.3 prod_j_per_batch=377.9 prod_mh_per_w=0.0444 samples=360 steady_wall_s=85.654 steady_mhs=6.27 steady_mean_w=136.7 steady_j_per_batch=365.9 steady_mh_per_w=0.0459 samples=342 worker_cpu_s=5.81 worker_sha=9b36b523d0812c3c
3090-full-r4 sm MHz mean 210 min 210 max 210 over 369 samples
== 3090-select-r4 2026-10-08T19:47:59Z sm 210 MHz, Active

3
docs/analysis/review-2026-10-08-b/f10/3090-select-r1.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=7.093 prod_wall_s=91.584 prod_mhs=5.86 prod_mean_w=131.7 prod_j_per_batch=376.9 prod_mh_per_w=0.0445 samples=366 steady_wall_s=84.491 steady_mhs=6.35 steady_mean_w=136.9 steady_j_per_batch=361.5 steady_mh_per_w=0.0464 samples=337 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 2637.57 ms, select 0.532 ms, read-back 0.296 ms, scan 0.019 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/3090-select-r2.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.138 prod_wall_s=21.996 prod_mhs=24.41 prod_mean_w=286.1 prod_j_per_batch=196.7 prod_mh_per_w=0.0853 samples=88 steady_wall_s=17.858 steady_mhs=30.06 steady_mean_w=330.7 steady_j_per_batch=184.6 steady_mh_per_w=0.0909 samples=71 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 555.70 ms, select 0.268 ms, read-back 0.217 ms, scan 0.013 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/3090-select-r3.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_3090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=4.297 prod_wall_s=83.427 prod_mhs=6.44 prod_mean_w=138.7 prod_j_per_batch=361.6 prod_mh_per_w=0.0464 samples=334 steady_wall_s=79.130 steady_mhs=6.78 steady_mean_w=140.1 steady_j_per_batch=346.4 steady_mh_per_w=0.0484 samples=316 worker_cpu_s=4.23 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_3090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 87 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_86 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 2470.39 ms, select 0.503 ms, read-back 0.206 ms, scan 0.017 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-full-r1.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.275 prod_wall_s=10.715 prod_mhs=50.10 prod_mean_w=278.0 prod_j_per_batch=93.1 prod_mh_per_w=0.1802 samples=43 steady_wall_s=8.440 steady_mhs=63.61 steady_mean_w=328.3 steady_j_per_batch=86.6 steady_mh_per_w=0.1938 samples=34 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 236.64 ms, select 0.000 ms, read-back 12.794 ms, scan 13.245 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-full-r2.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.373 prod_wall_s=11.405 prod_mhs=47.07 prod_mean_w=270.7 prod_j_per_batch=96.5 prod_mh_per_w=0.1739 samples=46 steady_wall_s=9.031 steady_mhs=59.45 steady_mean_w=326.9 steady_j_per_batch=92.3 steady_mh_per_w=0.1819 samples=36 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 236.65 ms, select 0.000 ms, read-back 25.097 ms, scan 19.536 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-full-r3.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.400 prod_wall_s=10.582 prod_mhs=50.73 prod_mean_w=267.7 prod_j_per_batch=88.5 prod_mh_per_w=0.1895 samples=43 steady_wall_s=8.182 steady_mhs=65.62 steady_mean_w=335.4 steady_j_per_batch=85.8 steady_mh_per_w=0.1956 samples=33 worker_cpu_s=2.97 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 236.73 ms, select 0.000 ms, read-back 7.679 ms, scan 10.341 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-full-r4.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.334 prod_wall_s=10.566 prod_mhs=50.81 prod_mean_w=285.8 prod_j_per_batch=94.4 prod_mh_per_w=0.1778 samples=43 steady_wall_s=8.233 steady_mhs=65.21 steady_mean_w=351.5 steady_j_per_batch=90.4 steady_mh_per_w=0.1855 samples=33 worker_cpu_s=2.92 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 237.01 ms, select 0.000 ms, read-back 8.639 ms, scan 10.488 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-full-r5.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.463 prod_wall_s=10.767 prod_mhs=49.86 prod_mean_w=276.8 prod_j_per_batch=93.1 prod_mh_per_w=0.1801 samples=43 steady_wall_s=8.304 steady_mhs=64.65 steady_mean_w=340.3 steady_j_per_batch=88.3 steady_mh_per_w=0.1900 samples=33 worker_cpu_s=3.13 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback full worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 134217728 B (full, 0 full fallbacks); mean per chunk: kernel+sync 236.70 ms, select 0.000 ms, read-back 9.028 ms, scan 12.796 ms; 32 jobs 32 chunks

4
docs/analysis/review-2026-10-08-b/f10/5090-kernel-bench.log Normal file
View file

@ -0,0 +1,4 @@
info igneum-worker-cuda 1.0 (4 October 2026): device 0 NVIDIA_GeForce_RTX_5090 (sm_120, 170 SMs), driver 13.2 from libcuda.so.1, NVRTC 12.8 from libnvrtc.so.12, target sm_120 (the device's architecture, listed by NVRTC)
pack /root/p02/packs/hl-v6-all on NVIDIA_GeForce_RTX_5090: nvrtc 1476 cache 1 dataset 13 hot 0 check 532 race 0 ms variant base class v5 (state leaves 93, uploaded for the build and freed); self-test PASS (cache head, last line and FNV-1a 64 448274a57f508cbc; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes)
warm-up dispatch (base 0): 236.80 ms; 5 timed dispatches of 16777216 nonces: mean 236.68 ms
RESULT pack=/root/p02/packs/hl-v6-all class=mx8-erad810f22d+sh256x27+state+reg64c+fold+rw device=NVIDIA_GeForce_RTX_5090 arch=sm_120 regs=92 blocks_per_sm=20 warps=0 resident=3400 arena_mib=0 hot_mib=0 hot_slots=0 hot_fill_ms=0.00 nonces=16777216 batches=5 check=PASS fingerprint=59e6708e46f1e87c mhs=70.885 loads=128 bytes=512 scratch_ops=0 time=wall

19
docs/analysis/review-2026-10-08-b/f10/5090-run-more.log Normal file
View file

@ -0,0 +1,19 @@
== 5090-full-r3 2026-10-08T19:44:33Z sm 180 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.400 prod_wall_s=10.582 prod_mhs=50.73 prod_mean_w=267.7 prod_j_per_batch=88.5 prod_mh_per_w=0.1895 samples=43 steady_wall_s=8.182 steady_mhs=65.62 steady_mean_w=335.4 steady_j_per_batch=85.8 steady_mh_per_w=0.1956 samples=33 worker_cpu_s=2.97 worker_sha=9b36b523d0812c3c
5090-full-r3 sm MHz mean 2387 min 180 max 2895 over 51 samples
== 5090-select-r3 2026-10-08T19:44:46Z sm 2895 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.327 prod_wall_s=9.951 prod_mhs=53.95 prod_mean_w=290.5 prod_j_per_batch=90.3 prod_mh_per_w=0.1857 samples=40 steady_wall_s=7.624 steady_mhs=70.42 steady_mean_w=365.4 steady_j_per_batch=87.1 steady_mh_per_w=0.1927 samples=30 worker_cpu_s=2.33 worker_sha=9b36b523d0812c3c
5090-select-r3 sm MHz mean 2636 min 472 max 2895 over 49 samples
== 5090-full-r4 2026-10-08T19:44:58Z sm 2895 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.334 prod_wall_s=10.566 prod_mhs=50.81 prod_mean_w=285.8 prod_j_per_batch=94.4 prod_mh_per_w=0.1778 samples=43 steady_wall_s=8.233 steady_mhs=65.21 steady_mean_w=351.5 steady_j_per_batch=90.4 steady_mh_per_w=0.1855 samples=33 worker_cpu_s=2.92 worker_sha=9b36b523d0812c3c
5090-full-r4 sm MHz mean 2666 min 457 max 2895 over 51 samples
== 5090-select-r4 2026-10-08T19:45:12Z sm 2895 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.590 prod_wall_s=10.216 prod_mhs=52.55 prod_mean_w=285.4 prod_j_per_batch=91.1 prod_mh_per_w=0.1841 samples=41 steady_wall_s=7.626 steady_mhs=70.40 steady_mean_w=355.9 steady_j_per_batch=84.8 steady_mh_per_w=0.1978 samples=31 worker_cpu_s=2.55 worker_sha=9b36b523d0812c3c
5090-select-r4 sm MHz mean 2608 min 555 max 2895 over 50 samples
== 5090-full-r5 2026-10-08T19:45:24Z sm 2880 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=full jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.463 prod_wall_s=10.767 prod_mhs=49.86 prod_mean_w=276.8 prod_j_per_batch=93.1 prod_mh_per_w=0.1801 samples=43 steady_wall_s=8.304 steady_mhs=64.65 steady_mean_w=340.3 steady_j_per_batch=88.3 steady_mh_per_w=0.1900 samples=33 worker_cpu_s=3.13 worker_sha=9b36b523d0812c3c
5090-full-r5 sm MHz mean 2679 min 795 max 2880 over 52 samples
== 5090-select-r5 2026-10-08T19:45:38Z sm 2880 MHz, Not Active
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.393 prod_wall_s=10.019 prod_mhs=53.59 prod_mean_w=287.5 prod_j_per_batch=90.0 prod_mh_per_w=0.1864 samples=40 steady_wall_s=7.626 steady_mhs=70.40 steady_mean_w=360.8 steady_j_per_batch=86.0 steady_mh_per_w=0.1951 samples=30 worker_cpu_s=2.39 worker_sha=9b36b523d0812c3c
5090-select-r5 sm MHz mean 2771 min 2212 max 2880 over 49 samples
== end 2026-10-08T19:45:50Z

3
docs/analysis/review-2026-10-08-b/f10/5090-select-r1.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.447 prod_wall_s=10.070 prod_mhs=53.31 prod_mean_w=290.5 prod_j_per_batch=91.4 prod_mh_per_w=0.1835 samples=41 steady_wall_s=7.624 steady_mhs=70.42 steady_mean_w=362.5 steady_j_per_batch=86.4 steady_mh_per_w=0.1943 samples=31 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 236.71 ms, select 0.234 ms, read-back 0.178 ms, scan 0.005 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-select-r2.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.408 prod_wall_s=10.032 prod_mhs=53.52 prod_mean_w=289.1 prod_j_per_batch=90.6 prod_mh_per_w=0.1851 samples=40 steady_wall_s=7.624 steady_mhs=70.42 steady_mean_w=363.0 steady_j_per_batch=86.5 steady_mh_per_w=0.1940 samples=30 worker_cpu_s=0.00 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 236.69 ms, select 0.245 ms, read-back 0.270 ms, scan 0.007 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-select-r3.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.327 prod_wall_s=9.951 prod_mhs=53.95 prod_mean_w=290.5 prod_j_per_batch=90.3 prod_mh_per_w=0.1857 samples=40 steady_wall_s=7.624 steady_mhs=70.42 steady_mean_w=365.4 steady_j_per_batch=87.1 steady_mh_per_w=0.1927 samples=30 worker_cpu_s=2.33 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 236.71 ms, select 0.262 ms, read-back 0.353 ms, scan 0.006 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-select-r4.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.590 prod_wall_s=10.216 prod_mhs=52.55 prod_mean_w=285.4 prod_j_per_batch=91.1 prod_mh_per_w=0.1841 samples=41 steady_wall_s=7.626 steady_mhs=70.40 steady_mean_w=355.9 steady_j_per_batch=84.8 steady_mh_per_w=0.1978 samples=31 worker_cpu_s=2.55 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 236.63 ms, select 0.239 ms, read-back 0.320 ms, scan 0.010 ms; 32 jobs 32 chunks

3
docs/analysis/review-2026-10-08-b/f10/5090-select-r5.row Normal file
View file

@ -0,0 +1,3 @@
card=NVIDIA_GeForce_RTX_5090 mode=select jobs=32 batch_log2=24 hashes=536870912 found=519 setup_s=2.393 prod_wall_s=10.019 prod_mhs=53.59 prod_mean_w=287.5 prod_j_per_batch=90.0 prod_mh_per_w=0.1864 samples=40 steady_wall_s=7.626 steady_mhs=70.40 steady_mean_w=360.8 steady_j_per_batch=86.0 steady_mh_per_w=0.1951 samples=30 worker_cpu_s=2.39 worker_sha=9b36b523d0812c3c
ready: ready cuda NVIDIA_GeForce_RTX_5090 pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 16777216 regs 92 prepare 1 path nvrtc 12.8 driver 13.2 arch sm_120 variant base race off readback select worker 1.0 (4 October 2026)
transfers: transfers per chunk: down 536 B (select, 0 full fallbacks); mean per chunk: kernel+sync 236.78 ms, select 0.296 ms, read-back 0.244 ms, scan 0.007 ms; 32 jobs 32 chunks

283
docs/analysis/review-2026-10-08-b/f10/emu-test.log Normal file
View file

@ -0,0 +1,283 @@
igneum-pow export /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-b
seed "igneum-epoch/5e5d0c3b2a19f8e7d6c5b4a3928170f6e5d4c3b2a1908f7e6d5c4b3a2918f7e6/day/69676e65756d2d6461792ffa51000000000000", day "bytes:69676e65756d2d6461792ffa51000000000000", dataset 2^28 words (memory-hard), generator v2 attempt 0 program id 36e7763af4220724, loads/hash 128; epoch built in 376.1 ms
/home/build/wt/worker-select-pass/proto-cuda/nvrtc/worker.cpp:131:14: warning: ‘void* libSym(void*, const char*)’ defined but not used [-Wunused-function]
131 | static void* libSym(void* lib, const char* name) {
| ^~~~~~
/home/build/wt/worker-select-pass/proto-cuda/nvrtc/worker.cpp:124:14: warning: ‘void* libOpen(const std::string&)’ defined but not used [-Wunused-function]
124 | static void* libOpen(const std::string& name) {
| ^~~~~~~
/home/build/wt/worker-select-pass/proto-cuda/nvrtc/worker.cpp:109:20: warning: ‘std::string exeDir()’ defined but not used [-Wunused-function]
109 | static std::string exeDir() {
| ^~~~~~
built /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/igneum-worker-cuda-emu (CPU emulation, not a GPU build)
== --check pack A
info igneum-worker-cuda 1.0 (4 October 2026): device 0 CPU_emulation_shim_(not_a_GPU) (sm_120, 0 SMs), driver 12.8 from emulation (host threads, no GPU), NVRTC 12.8 from emulation (source recorded and checked, nothing compiled), target sm_120 (the device's architecture, listed by NVRTC)
info emu-nvrtc: kernel.cu source check PASS for pack 1: 7385 bytes handed over = the pack file's first 7385 of 8893 bytes; dropped tail = host launch wrappers only (1508 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info emu-nvrtc: kernel_bound.cu source check PASS for pack 1: 6003 bytes handed over = the pack file's first 6003 of 6887 bytes; dropped tail = host launch wrappers only (884 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
check PASS /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-a in 2679 ms: nvrtc 0 cache 16 dataset 2193 hot 0 check 470 race 0 ms variant base class v2; self-test PASS (cache head, last line and FNV-1a 64 448274a57f508cbc; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes)
epoch edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07 day 69676e65756d2d6461792ffa50000000000000, dataset 2^28 words, cache 2^26 words in 65536 segments, 0 registers, 0 blocks/SM at 1 warp(s)/block, target sm_120
== the annotation rule: without -default-device the stand-in must reject program.h's igneum_launch_* declarations as host code (what the RTX 5090 reported on 4 October 2026)
flagged as NVRTC does: program.h(46): cudaError_t igneum_launch_cache_fill(uint32_t* cache, uint32_t nSegments);
== --serve: two jobs on A, prepare B, a job on B (swap), a job on A after the swap (error), quit
info igneum-worker-cuda 1.0 (4 October 2026): device 0 CPU_emulation_shim_(not_a_GPU) (sm_120, 0 SMs), driver 12.8 from emulation (host threads, no GPU), NVRTC 12.8 from emulation (source recorded and checked, nothing compiled), target sm_120 (the device's architecture, listed by NVRTC)
info emu-nvrtc: kernel.cu source check PASS for pack 1: 7385 bytes handed over = the pack file's first 7385 of 8893 bytes; dropped tail = host launch wrappers only (1508 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info emu-nvrtc: kernel_bound.cu source check PASS for pack 1: 6003 bytes handed over = the pack file's first 6003 of 6887 bytes; dropped tail = host launch wrappers only (884 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info emu-nvrtc: igneum_select.cu source check PASS: 763 bytes, the select pass runs as a host loop
ready cuda CPU_emulation_shim_(not_a_GPU) pack igneum-epoch/edc4fa844da9dc98d37e965176f6558a31560e40502ab3ae5491b21aaaabfb07/day/69676e65756d2d6461792ffa50000000000000 dataset-log2 28 batch 8192 regs 0 prepare 1 path nvrtc 12.8 driver 12.8 arch sm_120 variant base race on readback select worker 1.0 (4 October 2026)
info readback select (per dispatch of 8192 nonces: the hits and 34 sentinel words come back; the GPU scans)
info first pack /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-a: nvrtc 0 cache 15 dataset 2686 hot 0 check 472 race 0 ms variant base class v2; self-test PASS (cache head, last line and FNV-1a 64 448274a57f508cbc; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes)
race edc4fa844da9dc98 device CPU_emulation_shim_(not_a_GPU) variants 1 base only, no race (emulation)
found 1 0 85f8352a2701deaa edc4fa844da9dc98
found 1 1 287bddfe85a71a6c edc4fa844da9dc98
found 1 2 fd65144d28e0e2b1 edc4fa844da9dc98
found 1 3 4c451942cdccec1a edc4fa844da9dc98
found 1 4 b974b6a86b1cf0ed edc4fa844da9dc98
found 1 5 c470d05d126a1a61 edc4fa844da9dc98
found 1 6 824f53a0342e593d edc4fa844da9dc98
found 1 7 ab83476b2d6aa8d0 edc4fa844da9dc98
found 1 8 bb1a171d76ebe772 edc4fa844da9dc98
found 1 9 952788c85aa6ba6a edc4fa844da9dc98
found 1 10 eab425f84eeb2a81 edc4fa844da9dc98
found 1 11 6279385502e9b189 edc4fa844da9dc98
found 1 12 92bb6782d9337a70 edc4fa844da9dc98
found 1 13 bf69940f96c7607b edc4fa844da9dc98
found 1 14 4ec115ccd6700418 edc4fa844da9dc98
found 1 15 aee365ce8b71d4fd edc4fa844da9dc98
found 1 16 3b1074a344f48d75 edc4fa844da9dc98
found 1 17 7bbba30cdeee2318 edc4fa844da9dc98
found 1 18 85917afa2fb98002 edc4fa844da9dc98
found 1 19 a57d924a33c9eb47 edc4fa844da9dc98
found 1 20 ddf5d02d5440bed5 edc4fa844da9dc98
found 1 21 e21722ea35922f3d edc4fa844da9dc98
found 1 22 7c71ff55e0202ea0 edc4fa844da9dc98
found 1 23 2e264b7a3db0424e edc4fa844da9dc98
found 1 24 737435f635d5cd13 edc4fa844da9dc98
found 1 25 26a94c56d64fa788 edc4fa844da9dc98
found 1 26 c5a88d468a03eebb edc4fa844da9dc98
found 1 27 31ca283df921550d edc4fa844da9dc98
found 1 28 06d3199ce7b1adf4 edc4fa844da9dc98
found 1 29 0c8f5f7605deecd1 edc4fa844da9dc98
found 1 30 c43156d1e4a5cbec edc4fa844da9dc98
found 1 31 78ba75b0e4c908d5 edc4fa844da9dc98
found 1 32 e40b3c4ce16e0674 edc4fa844da9dc98
found 1 33 9a6b697360a70406 edc4fa844da9dc98
found 1 34 e28eb3b8c845ef22 edc4fa844da9dc98
found 1 35 a8db66a07fa26a94 edc4fa844da9dc98
found 1 36 81d256fa67d8bed3 edc4fa844da9dc98
found 1 37 f681ad4e0426b0e5 edc4fa844da9dc98
found 1 38 ad49770893af9688 edc4fa844da9dc98
found 1 39 e830f95c3e013a41 edc4fa844da9dc98
found 1 40 7f050e695bd46847 edc4fa844da9dc98
found 1 41 b6c2ae03f8ddf3ae edc4fa844da9dc98
found 1 42 772509fa6f742f94 edc4fa844da9dc98
found 1 43 3d0317e357e2381d edc4fa844da9dc98
found 1 44 94f2baf0d5af5a6b edc4fa844da9dc98
found 1 45 3e2ba39651256a56 edc4fa844da9dc98
found 1 46 588295beedbc0aff edc4fa844da9dc98
found 1 47 edc862c227a7f27d edc4fa844da9dc98
found 1 48 a540ec001fd3fe33 edc4fa844da9dc98
found 1 49 3f3a2aa56b364df1 edc4fa844da9dc98
found 1 50 b395503049cfefa2 edc4fa844da9dc98
found 1 51 bf7a4473bf7be326 edc4fa844da9dc98
found 1 52 8af17d040c654b2a edc4fa844da9dc98
found 1 53 d2670d56b9848c44 edc4fa844da9dc98
found 1 54 ac5f448acfb856d5 edc4fa844da9dc98
found 1 55 89d2248ea07b109a edc4fa844da9dc98
found 1 56 dd4c5e4c6bb7c2e6 edc4fa844da9dc98
found 1 57 2529ceca58759d77 edc4fa844da9dc98
found 1 58 0bdff5ded9bc65fb edc4fa844da9dc98
found 1 59 6ea20ccedaf3e18c edc4fa844da9dc98
found 1 60 bedd6f14e5fa465d edc4fa844da9dc98
found 1 61 08d7d815f5786ced edc4fa844da9dc98
found 1 62 7d289bf14d911c5f edc4fa844da9dc98
found 1 63 e2f4237e31cbd041 edc4fa844da9dc98
done 1 64 9.66
found 2 4294967264 7b6273e5bd32d0b5 edc4fa844da9dc98
found 2 4294967265 38ca4cce09bfb623 edc4fa844da9dc98
found 2 4294967266 d8b0ed94f45cdbf7 edc4fa844da9dc98
found 2 4294967267 9d8faff62d0efa15 edc4fa844da9dc98
found 2 4294967268 1df806f32926defe edc4fa844da9dc98
found 2 4294967269 e3654b2d705786bb edc4fa844da9dc98
found 2 4294967270 ab6741286c7c88dd edc4fa844da9dc98
found 2 4294967271 a15815bf414e9320 edc4fa844da9dc98
found 2 4294967272 6f1cc5a7a2fcc7c0 edc4fa844da9dc98
found 2 4294967273 6e98c101317928c6 edc4fa844da9dc98
found 2 4294967274 d87e3fefd2a7b381 edc4fa844da9dc98
found 2 4294967275 f7f30848fd476fb5 edc4fa844da9dc98
found 2 4294967276 257fc4d2e3603beb edc4fa844da9dc98
found 2 4294967277 1aab6f77c1365d4c edc4fa844da9dc98
found 2 4294967278 1c9d005f8f4ac463 edc4fa844da9dc98
found 2 4294967279 161427ad7f81f2f9 edc4fa844da9dc98
found 2 4294967280 e8611bc03fb827c1 edc4fa844da9dc98
found 2 4294967281 42454a79caee8c69 edc4fa844da9dc98
found 2 4294967282 190aa2b3529b589f edc4fa844da9dc98
found 2 4294967283 83d56d283df7f1e0 edc4fa844da9dc98
found 2 4294967284 fb355b9c51ca8988 edc4fa844da9dc98
found 2 4294967285 f5b626e35d01424a edc4fa844da9dc98
found 2 4294967286 8a0b81e4053fe9b0 edc4fa844da9dc98
found 2 4294967287 c1aa110e3db24ed7 edc4fa844da9dc98
found 2 4294967288 fc53be92ceb49a8e edc4fa844da9dc98
found 2 4294967289 c3296889fb569a0a edc4fa844da9dc98
found 2 4294967290 e38850154c2dea6c edc4fa844da9dc98
found 2 4294967291 d99fda96ee76c555 edc4fa844da9dc98
found 2 4294967292 fac8684a83d19072 edc4fa844da9dc98
found 2 4294967293 51085b8a57f3720d edc4fa844da9dc98
found 2 4294967294 7311e9864d96be76 edc4fa844da9dc98
found 2 4294967295 2c52ef30b1f52cbf edc4fa844da9dc98
found 2 4294967296 aaee181f46a06d70 edc4fa844da9dc98
found 2 4294967297 aa45e128a7ab498b edc4fa844da9dc98
found 2 4294967298 eb0f5f050141457c edc4fa844da9dc98
found 2 4294967299 7be5e2d3e944d761 edc4fa844da9dc98
found 2 4294967300 804964e4293faa67 edc4fa844da9dc98
found 2 4294967301 7567f837977d77da edc4fa844da9dc98
found 2 4294967302 1ff1311a643d3b43 edc4fa844da9dc98
found 2 4294967303 60cda5dadaaa941c edc4fa844da9dc98
found 2 4294967304 63542e063db3fb34 edc4fa844da9dc98
found 2 4294967305 e0ac767c5fab1ab4 edc4fa844da9dc98
found 2 4294967306 e2423cd93bfb5aa0 edc4fa844da9dc98
found 2 4294967307 c3210435edd81d5a edc4fa844da9dc98
found 2 4294967308 a0226e9f79d09099 edc4fa844da9dc98
found 2 4294967309 fc143bbd8555fdb0 edc4fa844da9dc98
found 2 4294967310 cc9008673f527726 edc4fa844da9dc98
found 2 4294967311 bb329b9bcf8e66fd edc4fa844da9dc98
found 2 4294967312 164442e9130c7b7a edc4fa844da9dc98
found 2 4294967313 48e43b690541132d edc4fa844da9dc98
found 2 4294967314 f4ba8f420da3337e edc4fa844da9dc98
found 2 4294967315 050dff2f1846ab61 edc4fa844da9dc98
found 2 4294967316 74972ac9ab8fb364 edc4fa844da9dc98
found 2 4294967317 b7ef1e3155e9a7f3 edc4fa844da9dc98
found 2 4294967318 3ecb2b74489fc8a4 edc4fa844da9dc98
found 2 4294967319 64fa8c8296a3d662 edc4fa844da9dc98
found 2 4294967320 6c4411933e7f690c edc4fa844da9dc98
found 2 4294967321 23234ffc83690177 edc4fa844da9dc98
found 2 4294967322 01dd1d050f87e26e edc4fa844da9dc98
found 2 4294967323 e3a05b34dde32af0 edc4fa844da9dc98
found 2 4294967324 0bcf244d4c73562d edc4fa844da9dc98
found 2 4294967325 375800d4039b5c1f edc4fa844da9dc98
found 2 4294967326 42b9a4c28e42c996 edc4fa844da9dc98
found 2 4294967327 f56d052d60a7338e edc4fa844da9dc98
done 2 64 10.89
info prepare started for epoch 5e5d0c3b2a19f8e7 day 69676e65756d2d6461792ffa51000000000000 from /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-b (NVRTC sm_120 in the background)
info emu-nvrtc: kernel.cu source check PASS for pack 2: 7206 bytes handed over = the pack file's first 7206 of 8714 bytes; dropped tail = host launch wrappers only (1508 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info emu-nvrtc: kernel_bound.cu source check PASS for pack 2: 5824 bytes handed over = the pack file's first 5824 of 6708 bytes; dropped tail = host launch wrappers only (884 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
race 5e5d0c3b2a19f8e7 device CPU_emulation_shim_(not_a_GPU) variants 1 base only, no race (emulation)
prepared 5e5d0c3b2a19f8e7d6c5b4a3928170f6e5d4c3b2a1908f7e6d5c4b3a2918f7e6 69676e65756d2d6461792ffa51000000000000 36806.6 nvrtc 0 cache 15 dataset 1535 hot 0 check 462 race 0 ms variant base class v2; self-test PASS (cache head, last line and FNV-1a 64 b5b0660c9db7ce59; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes) resident 2 programs 2 datasets
found 3 128 72b866f3367df254 edc4fa844da9dc98
found 3 129 a9772a8b6a793277 edc4fa844da9dc98
found 3 130 d07426082b75402d edc4fa844da9dc98
found 3 131 408083ae76ee518b edc4fa844da9dc98
found 3 132 c45e395f38ad22d8 edc4fa844da9dc98
found 3 133 3dbee02765058ade edc4fa844da9dc98
found 3 134 cd5f25857807cfa9 edc4fa844da9dc98
found 3 135 879e1e3fe942214a edc4fa844da9dc98
found 3 136 d2133aca0cc47d7a edc4fa844da9dc98
found 3 137 b143a77f8b19741c edc4fa844da9dc98
found 3 138 d3ed2304cc49a6c7 edc4fa844da9dc98
found 3 139 75cf3d6584e42e65 edc4fa844da9dc98
found 3 140 86db0130d062a7b7 edc4fa844da9dc98
found 3 141 05f62574da3354c5 edc4fa844da9dc98
found 3 142 7c961bec00e0cc4b edc4fa844da9dc98
found 3 143 db8c0f4e55c387ad edc4fa844da9dc98
found 3 144 7f1454c90435afbd edc4fa844da9dc98
found 3 145 002b3ca8c7bfd154 edc4fa844da9dc98
found 3 146 edae0b061a210205 edc4fa844da9dc98
found 3 147 e478cf93a9878f51 edc4fa844da9dc98
found 3 148 32c7dfd0fe55a026 edc4fa844da9dc98
found 3 149 06b93d72fb16404c edc4fa844da9dc98
found 3 150 4b6cace41d98e7fd edc4fa844da9dc98
found 3 151 41e2649e041acc06 edc4fa844da9dc98
found 3 152 0453f57fae7f514a edc4fa844da9dc98
found 3 153 ff01e0015a3f39e9 edc4fa844da9dc98
found 3 154 491ae49020261797 edc4fa844da9dc98
found 3 155 ef1bd996a361128a edc4fa844da9dc98
found 3 156 794082f3f114767d edc4fa844da9dc98
found 3 157 5822e90044ea782a edc4fa844da9dc98
found 3 158 259700c74b5cc149 edc4fa844da9dc98
found 3 159 235a57ae0bb5a2ed edc4fa844da9dc98
done 3 32 9.91
info switched to the prepared pair epoch 5e5d0c3b2a19f8e7 day 69676e65756d2d6461792ffa51000000000000 (class v2) in 0.00 ms
found 4 0 992153a3ef2b07fc 5e5d0c3b2a19f8e7
found 4 1 55e25d6fb49f4351 5e5d0c3b2a19f8e7
found 4 2 54a9131bc3066ee4 5e5d0c3b2a19f8e7
found 4 3 1c027229270c7d33 5e5d0c3b2a19f8e7
found 4 4 3806eac5cad07581 5e5d0c3b2a19f8e7
found 4 5 6a87c5563912348c 5e5d0c3b2a19f8e7
found 4 6 0b802423721bc6af 5e5d0c3b2a19f8e7
found 4 7 27ec87f124a88548 5e5d0c3b2a19f8e7
found 4 8 474bf5241c3a8ec3 5e5d0c3b2a19f8e7
found 4 9 532dbcf0caa93f96 5e5d0c3b2a19f8e7
found 4 10 259186fa7df9be63 5e5d0c3b2a19f8e7
found 4 11 f3b8c0df8787ff76 5e5d0c3b2a19f8e7
found 4 12 0e25eb13e9ddb276 5e5d0c3b2a19f8e7
found 4 13 a28c32bc4aeabf61 5e5d0c3b2a19f8e7
found 4 14 1e9c1a2be3f3f110 5e5d0c3b2a19f8e7
found 4 15 9d4a1cfde84d3aba 5e5d0c3b2a19f8e7
found 4 16 095727897ed26f5a 5e5d0c3b2a19f8e7
found 4 17 fdcbce6e7d66f667 5e5d0c3b2a19f8e7
found 4 18 aefcdcf46d9e6a4a 5e5d0c3b2a19f8e7
found 4 19 c6ec5a44a97ca398 5e5d0c3b2a19f8e7
found 4 20 50fce9ed2020b450 5e5d0c3b2a19f8e7
found 4 21 8c2fe5af5ee1e9f3 5e5d0c3b2a19f8e7
found 4 22 500ad998c15f42b6 5e5d0c3b2a19f8e7
found 4 23 95965bedaeaa462d 5e5d0c3b2a19f8e7
found 4 24 df645bcb6d1cc3ad 5e5d0c3b2a19f8e7
found 4 25 716b7f5dcc5c30dc 5e5d0c3b2a19f8e7
found 4 26 84fcbe9758749444 5e5d0c3b2a19f8e7
found 4 27 29511da3232f2dff 5e5d0c3b2a19f8e7
found 4 28 b976fd6a159db063 5e5d0c3b2a19f8e7
found 4 29 69889d22e3cb6709 5e5d0c3b2a19f8e7
found 4 30 151cc4a784c80add 5e5d0c3b2a19f8e7
found 4 31 27b10106d08793a5 5e5d0c3b2a19f8e7
done 4 32 8.22
info dropped the previous pair (its program, cache and dataset)
info job 5 is for epoch edc4fa844da9dc98 day 69676e65756d2d6461792ffa50000000000000, which is not resident; building its pack /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-a now (foreground)
info emu-nvrtc: kernel.cu source check PASS for pack 1: 7385 bytes handed over = the pack file's first 7385 of 8893 bytes; dropped tail = host launch wrappers only (1508 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info emu-nvrtc: kernel_bound.cu source check PASS for pack 1: 6003 bytes handed over = the pack file's first 6003 of 6887 bytes; dropped tail = host launch wrappers only (884 bytes); program.h (3024 bytes) and memhard.h (5848 bytes) byte-identical
info built /home/build/wt/worker-select-pass/proto-cuda/nvrtc/emu/build/pack-a: nvrtc 0 cache 15 dataset 1295 hot 0 check 481 race 0 ms variant base class v2; self-test PASS (cache head, last line and FNV-1a 64 448274a57f508cbc; dataset head, word [268435455] and 64 samples; 96 of 96 vector lanes)
race edc4fa844da9dc98 device CPU_emulation_shim_(not_a_GPU) variants 1 base only, no race (emulation)
info switched to the prepared pair epoch edc4fa844da9dc98 day 69676e65756d2d6461792ffa50000000000000 (class v2) in 1791.36 ms
found 5 96 cb13242c3a1a4ffa edc4fa844da9dc98
found 5 97 f3071494c5382b79 edc4fa844da9dc98
found 5 98 27d7c3cf7cec0c32 edc4fa844da9dc98
found 5 99 4137aaf4e3ec29e1 edc4fa844da9dc98
found 5 100 3e18fe3b1190d59f edc4fa844da9dc98
found 5 101 c1d940643197c68e edc4fa844da9dc98
found 5 102 005901b495cc73be edc4fa844da9dc98
found 5 103 2d18e57bfa59ce02 edc4fa844da9dc98
found 5 104 326bc2b1e5eec1d3 edc4fa844da9dc98
found 5 105 84b754dc15b156b0 edc4fa844da9dc98
found 5 106 bdb2c9a6d88c78ce edc4fa844da9dc98
found 5 107 89d57198d7f17874 edc4fa844da9dc98
found 5 108 3c96f431799d9d97 edc4fa844da9dc98
found 5 109 2d98f14c5b8153a4 edc4fa844da9dc98
found 5 110 a6d1bf98ddda41fb edc4fa844da9dc98
found 5 111 cd14d9e675dd60c0 edc4fa844da9dc98
found 5 112 349128ba141dd961 edc4fa844da9dc98
found 5 113 af2496d2e09bce14 edc4fa844da9dc98
found 5 114 59ca36336962df25 edc4fa844da9dc98
found 5 115 f44a3eab722df34b edc4fa844da9dc98
found 5 116 f7c766cb70017bb1 edc4fa844da9dc98
found 5 117 4fcabac398cca930 edc4fa844da9dc98
found 5 118 f2e1cb0a1478ed76 edc4fa844da9dc98
found 5 119 5fc703e08569cf0d edc4fa844da9dc98
found 5 120 4aa1b3c1b264ff4e edc4fa844da9dc98
found 5 121 008922eb5f1b0558 edc4fa844da9dc98
found 5 122 7a00d435fa182ea4 edc4fa844da9dc98
found 5 123 347fd45b06777e05 edc4fa844da9dc98
found 5 124 86e87b980be70737 edc4fa844da9dc98
found 5 125 7cdbe6e309e01af7 edc4fa844da9dc98
found 5 126 22de307670d9ee71 edc4fa844da9dc98
found 5 127 d582b681bf406666 edc4fa844da9dc98
done 5 32 1800.01
info dropped the previous pair (its program, cache and dataset)
info transfers per chunk: down 873 B (select, 0 full fallbacks); mean per chunk: kernel+sync 7.87 ms, select 0.001 ms, read-back 0.001 ms, scan 0.008 ms; 5 jobs 6 chunks
== verdict
PASS: ready + prepare 1; 64 + 64 + 32 found on pack A, prepared B with self-test PASS, swapped, 32 found on B, job 5 served after the worker rebuilt pack A by itself (self-heal); 17 sampled hashes (both packs, both sides of the 32-bit nonce boundary) equal igneum-pow hash-bound
NVRTC source check PASS for 7 files in --serve, 2 in --check
== the select pass (8 October 2026): --readback select equals the full read-back, known-failed first (emu/select-check.sh)
== select_fault_detected (known-failed first: the counter is not reset from the second chunk on)
flagged: 84 stale or extra lines in the fault run
== found_set_equals_full_* (every case, line for line)
PASS: select = full over 2255 found lines and 10 jobs (zero hits, partial, all-hit overflow with 2 full fallbacks, a 96-nonce tail, the high-32 rollover, two queued jobs, hits then zero, a mismatch error then a job); the fault run was flagged first
transfers: transfers per chunk: down 1234 B (select, 2 full fallbacks); mean per chunk: kernel+sync 149.78 ms, select 0.003 ms, read-back 0.001 ms, scan 0.052 ms; 10 jobs 20 chunks
transfers: transfers per chunk: down 7411 B (full, 0 full fallbacks); mean per chunk: kernel+sync 142.04 ms, select 0.000 ms, read-back 0.002 ms, scan 0.050 ms; 10 jobs 20 chunks

31
docs/analysis/review-2026-10-08-b/f10/geometry-check-mini.log Normal file
View file

@ -0,0 +1,31 @@
ok geometry ds55: loads as 1476395008 words, 92274688 items, mulshift 1, 5905580032 bytes
ok geometry ds85: loads as 2281701376 words, 142606336 items, mulshift 1, 9126805504 bytes
ok geometry ds115: loads as 3087007744 words, 192937984 items, mulshift 1, 12348030976 bytes
ok geometry legacy-2^28: loads as 268435456 words, 16777216 items, mulshift 0, 1073741824 bytes
ok geometry legacy-2^30: loads as 1073741824 words, 67108864 items, mulshift 0, 4294967296 bytes
ok geometry legacy-2^31-explicit-count: loads as 2147483648 words, 134217728 items, mulshift 0, 8589934592 bytes
ok geometry wrong-count-not-a-65536-multiple: refused with "multiple of 65,536"
ok geometry wrong-log2-for-the-count: refused with "IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count"
ok geometry wrong-log2-for-a-power-of-two-count: refused with "IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count"
ok geometry wrong-items: refused with "IGNEUM_DATASET_ITEMS is not IGNEUM_DATASET_WORDS / 16"
ok geometry wrong-mulshift-on-a-non-power-of-two: refused with "IGNEUM_DATASET_MULSHIFT does not match the word count"
ok geometry log2-32: refused with "IGNEUM_DATASET_LOG2 above 31"
ok geometry overflowing-count-2^32: refused with "IGNEUM_DATASET_WORDS above 2^32 - 1"
ok geometry overflowing-count-2^40: refused with "IGNEUM_DATASET_WORDS above 2^32 - 1"
ok geometry count-below-the-minimum: refused with "multiple of 65,536"
ok geometry log2-below-20: refused with "sizes out of range"
ok geometry missing-log2: refused with "has no IGNEUM_DATASET_LOG2"
ok geometry missing-leaves-on-a-class-v5-pack: refused with "without IGNEUM_STATE_LEAVES"
ok geometry leaves-on-a-class-v2-pack: refused with "state leaves belong to class v5"
ok the vectors file carries every case (18 or more): 19
ok pack /Users/igneum/Projects/igneum/proto-cuda/packs/igneum-devnet-v4-epoch0: 268435456 words (2^28), 16777216 items, 1073741824 bytes
ok pack /Users/igneum/Projects/igneum/proto-cuda/packs-ca3-v5/v5-dn3-epoch0: 268435456 words (2^28), 16777216 items, 1073741824 bytes
geometry-check: 19 vectors, 2 packs, 0 failure(s)
== Metal serve smoke, class v5 pack (packs-ca3-v5/v5-dn3-epoch0), the exact-geometry pack path
ready metal Apple_M6 dataset-log2 28 batch 4194304 prepare 1 race off
prepared 4020cb4382e3fe4b281c817c02582e147d8f851f566ae9172b28912b8e68b925 69676e65756d2d6461792ffd50000000000000 241.1 program 81.3 dataset 159.8 race 0.0 variant base class v5 loads/hash 128 cache-fill 4.0 build 95.8 resident 1 programs 1 datasets
found 1 0 c6e114e3d317ee3e 4020cb4382e3fe4b
found 1 63 77fe95d9ed898030 4020cb4382e3fe4b
done 1 64 54.50
== Metal serve smoke, the v4 pack (igneum-devnet-v4-epoch0, class v2 generator path): ready metal Apple_M6 ...; found 1 0 65413452d4d05a3c edc4fa844da9dc98; found 1 63 4b8c6966db6ac21f edc4fa844da9dc98; done 1 64 61.69
== build: swiftc -O main.swift on the mini (Apple M6, CLT swift 6.4), Thu Oct 8 20:51:52 BST 2026

75
docs/analysis/review-2026-10-08-b/f10/packfile-test.log Normal file
View file

@ -0,0 +1,75 @@
ok the fixture seeds unhex (32 epoch bytes, 19 day bytes)
ok attempt 0 of epoch 34 = the bare seed words (06af2a61 4d67274e ...)
ok attempt 1 of epoch 34 = words of seed || 01000000 (0dcff56b 6b1beb0d ...)
ok the two attempts differ
ok known-good: the attempt-1 pack loads
ok known-good: attempt, words and epoch hex read back
ok known-good: the checked-in attempt-0 pack loads from program.h's own bytes
ok the checked-in pack is attempt 0
ok known-mismatched: bare words under attempt 1 are refused
ok the refusal names the attempt and the epoch in plain words
(program pack and its seeds disagree: IGNEUM_SEEDW_INIT is not attempt 1 of the epoch seed 009858237e118f69 (attempt 1 gives 0dcff56b 6b1beb0d ..., the pack has 06af2a61 4d67274e ...); run igneum-miner export-pack again)
ok known-mismatched: seeds.txt of epoch 33 with program.h of epoch 34 is refused
ok the refusal says the files disagree
ok no attempt line reads as attempt 0 and the bare words load
ok a generator 2 pack without a class line is class v2 with no era
ok known-failed: a pack with IGNEUM_DATASET_WORDS is sized by the word count (1476395008 words, 92274688 items, multiply-shift), not by 1 << 30
ok the count is not the floor's power of two and the items are words / 16
ok known-good: a pack without the line keeps 1 << IGNEUM_DATASET_LOG2 (2^28 words, 2^24 items, the mask path)
ok known-good: the checked-in pack is sized by its power of two
ok generator 9 is refused in plain words
ok a generator 3 pack loads as class v3 with its era seed
ok the v3 pack matches a job naming class v3 and its era
ok a job naming no class accepts the pack
ok a job naming class v2 refuses the v3 pack
ok a job naming another era refuses the v3 pack
ok a job naming class v3 refuses a v2 pack
ok an era named on a v2 job is ignored
ok a generator 4 pack loads as class v4 with its era seed
ok a generator 3 pack with a shadow block (a v4 program stamped v3) is refused
ok a generator 4 pack without the shadow block is refused
ok the v4 pack with its block loads again
ok the v4 pack matches a job naming class v4 and its era
ok a job naming class v3 refuses the v4 pack
ok a job naming another era refuses the v4 pack
ok a job naming class v4 refuses a v3 pack
ok the class=v4 token is read
ok a v3 class line on a generator 4 pack is refused
ok a class line that contradicts the generator is refused
ok a pack with no generator line (generator 1) is refused
ok known-failed: a generator 5 pack without the leaf count is refused in plain words
ok a generator 5 pack with the leaf count loads as class v5 (count, FNV and file name read back)
ok the v5 pack matches a job naming class v5 and its era
ok a job naming class v4 refuses the v5 pack
ok known-failed: the leaves file missing refuses the build in plain words
ok known-failed: a short leaves file is refused with both sizes named
ok known-failed: leaves of another root (one bit) are refused by the FNV
ok known-good: the right leaves load as 2 x 16 words
ok known-failed: a generator 4 pack carrying IGNEUM_STATE_LEAVES is refused
ok a class v4 pack has no leaves and the leaf loader hands back none
ok known-good: the checked-in class v5 pack loads with its leaf count and vectors
ok known-good: the checked-in pack's leaves.bin loads under the pack's FNV-1a 64
v5 pack: 11 leaves, FNV-1a 64 f141bfee2a8b11b0, state root 7e37a9fb19b154d32daf5bf30a50d339a75029fbc9eec9ea20e95439dba5a311
ok the geometry vectors file opens
ok geometry ds55: loads as 1476395008 words, 92274688 items, mulshift 1, 5905580032 bytes
ok geometry ds85: loads as 2281701376 words, 142606336 items, mulshift 1, 9126805504 bytes
ok geometry ds115: loads as 3087007744 words, 192937984 items, mulshift 1, 12348030976 bytes
ok geometry legacy-2^28: loads as 268435456 words, 16777216 items, mulshift 0, 1073741824 bytes
ok geometry legacy-2^30: loads as 1073741824 words, 67108864 items, mulshift 0, 4294967296 bytes
ok geometry legacy-2^31-explicit-count: loads as 2147483648 words, 134217728 items, mulshift 0, 8589934592 bytes
ok geometry wrong-count-not-a-65536-multiple: refused with "multiple of 65,536"
ok geometry wrong-log2-for-the-count: refused with "IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count"
ok geometry wrong-log2-for-a-power-of-two-count: refused with "IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count"
ok geometry wrong-items: refused with "IGNEUM_DATASET_ITEMS is not IGNEUM_DATASET_WORDS / 16"
ok geometry wrong-mulshift-on-a-non-power-of-two: refused with "IGNEUM_DATASET_MULSHIFT does not match the word count"
ok geometry log2-32: refused with "IGNEUM_DATASET_LOG2 above 31"
ok geometry overflowing-count-2^32: refused with "IGNEUM_DATASET_WORDS above 2^32 - 1"
ok geometry overflowing-count-2^40: refused with "IGNEUM_DATASET_WORDS above 2^32 - 1"
ok geometry count-below-the-minimum: refused with "multiple of 65,536"
ok geometry log2-below-20: refused with "sizes out of range"
ok geometry missing-log2: refused with "has no IGNEUM_DATASET_LOG2"
ok geometry missing-leaves-on-a-class-v5-pack: refused with "without IGNEUM_STATE_LEAVES"
ok geometry leaves-on-a-class-v2-pack: refused with "state leaves belong to class v5"
ok the vectors file carries every case (18 or more)
19 geometry vectors
/tmp/igneum-packfile-test/packfile-test: 0 failure(s)

109
docs/analysis/review-2026-10-08-b/f10/rows.md Normal file
View file

@ -0,0 +1,109 @@
# F10 worker lane: the CUDA select pass, measured (8 October 2026)
Review B finding F10 (docs/analysis/review-2026-10-08-b/findings.json): the CUDA serving loop read 8 bytes per nonce back
over the bus (128 MiB per 2^24-nonce dispatch) and scanned them on the host under the GPU mutex. The select pass the
OpenCL worker has carried since 5 October 2026 is now on CUDA (proto-cuda/nvrtc/worker.cpp, landed on master as
964cdf6c, commits 12cb65cc and 4b0de631): a tiny `igneum_select` kernel built once by NVRTC writes the hits behind an
atomic counter plus 34 sentinel words, the host reads back a few hundred bytes, a chunk with more than 256 hits falls
back to the full read, the found lines come in nonce order from the sorted hits and carry the kernel's own hash word.
Byte-preserving: the bound kernel and its result layout are untouched; no valid hash changes.
Two numbers per card, never interchangeable: the **kernel timer** (`--bench`, the dispatch alone) and the
**serving-mode numbers** (`proto-cuda/nvrtc/serve-bench.sh`: the worker in `--serve` driven by a stub pool exactly as
the miner drives it, one job queued behind the one in flight, every found line read at the miner's submit point). The
**production** number runs from the worker process start (NVRTC compile, cache, dataset, self-test included) to the last
done line; the **steady** number from the ready line. Power is nvidia-smi's board reading at 4 Hz over the same window
(a RunPod container exposes no host rail; the worker's own CPU seconds stand in for the host-side cost of the read-back
and scan). Target 00000fffffffffff (one hit per 2^20 nonces, 16 found lines per job), 32 jobs of 2^24 nonces, `--race off`,
variant base, pack hl-v6-all (class v6 kit packs-class-v6-20261008T162324Z, class string
mx8-erad810f22d+sh256x27+state+reg64c+fold+rw). Worker binary sha256 9b36b523d0812c3c (Ubuntu 24.04 build of 12cb65cc).
The worker has no stale-job cancellation (a job runs to its done line; the next line waits), so the stub's queued job is
the stale case the miner produces; nothing is cancelled mid-dispatch on any backend today.
## Equivalence (build-7, CPU emulation, proto-cuda/nvrtc/emu/test.sh, green 20:20 and again 20:55 UK)
`emu/select-check.sh`, known-failed first: the fault run (IGNEUM_READBACK_FAULT_TEST skips the counter reset from the
second chunk on) is flagged with 84 stale lines; then `--readback select` equals the full read line for line over 2255
found lines and 10 jobs: zero hits, partial hits, the all-hit overflow (every chunk over 256 hits, 2 full fallbacks), a
96-nonce tail with a 32-lane launch (POW-01 in the record), the high-32 rollover (ROT-01 in the record), two queued jobs,
hits then a zero chunk, a mismatch error then a job, and every found line's fifth field equal to the pack's epoch (I05).
`emu/serve-check.sh`: 17 sampled hashes on both packs and both sides of the 32-bit boundary equal igneum-pow hash-bound;
the swap case's found lines carry pack B's epoch. Logs: emu-test.log beside this file.
## Kernel timer
| card | arch | ptxas (nvcc 12.8, the pack's kernel_bound.cu) | --bench 5 x 2^24 | kernel MH/s |
|---|---|---|---|---|
| RTX 5090 (pod 4jpym0xwr1jnf3, driver 595.91.07) | sm_120 | 92 registers, 0 bytes spill stores, 0 bytes spill loads, 0 bytes stack; worker: 20 blocks/SM, 3400 resident warps | 236.75 ms | 70.86 |
| RTX 3090 (pod xq411hlgu4qzri, driver 595.71.05) | sm_86 | 87 registers, 0 bytes spill stores, 0 bytes spill loads, 0 bytes stack, 408 bytes cmem; worker: 16 blocks/SM, 1312 resident warps | warm-up 554.73 ms, timed mean 2590.06 ms (throttled, see below) | 30.2 at clock, 6.5 throttled |
## Serving mode
The pod rows (the .row files beside this file carry every field; the clocks are nvidia-smi clocks.sm means over each run):
### 5090
Mean SM clock 2387 to 2823 MHz over the runs (2895 max); no throttle reason active.
| run | mode | setup s | production MH/s | production J per batch | steady MH/s | steady W | steady J per batch | steady MH per W | worker CPU s | per chunk: kernel+sync, select, read-back, scan ms |
|---|---|---|---|---|---|---|---|---|---|---|
| 5090-full-r1 | full | 2.275 | 50.10 | 93.1 | 63.61 | 328.3 | 86.6 | 0.1938 | (subshell pid, not read) | 236.64, 0.000, 12.794, 13.245 |
| 5090-full-r2 | full | 2.373 | 47.07 | 96.5 | 59.45 | 326.9 | 92.3 | 0.1819 | (subshell pid, not read) | 236.65, 0.000, 25.097, 19.536 |
| 5090-full-r3 | full | 2.400 | 50.73 | 88.5 | 65.62 | 335.4 | 85.8 | 0.1956 | 2.97 | 236.73, 0.000, 7.679, 10.341 |
| 5090-full-r4 | full | 2.334 | 50.81 | 94.4 | 65.21 | 351.5 | 90.4 | 0.1855 | 2.92 | 237.01, 0.000, 8.639, 10.488 |
| 5090-full-r5 | full | 2.463 | 49.86 | 93.1 | 64.65 | 340.3 | 88.3 | 0.1900 | 3.13 | 236.70, 0.000, 9.028, 12.796 |
| 5090-select-r1 | select | 2.447 | 53.31 | 91.4 | 70.42 | 362.5 | 86.4 | 0.1943 | (subshell pid, not read) | 236.71, 0.234, 0.178, 0.005 |
| 5090-select-r2 | select | 2.408 | 53.52 | 90.6 | 70.42 | 363.0 | 86.5 | 0.1940 | (subshell pid, not read) | 236.69, 0.245, 0.270, 0.007 |
| 5090-select-r3 | select | 2.327 | 53.95 | 90.3 | 70.42 | 365.4 | 87.1 | 0.1927 | 2.33 | 236.71, 0.262, 0.353, 0.006 |
| 5090-select-r4 | select | 2.590 | 52.55 | 91.1 | 70.40 | 355.9 | 84.8 | 0.1978 | 2.55 | 236.63, 0.239, 0.320, 0.010 |
| 5090-select-r5 | select | 2.393 | 53.59 | 90.0 | 70.40 | 360.8 | 86.0 | 0.1951 | 2.39 | 236.78, 0.296, 0.244, 0.007 |
5090 reading (five pairs): steady 70.41 MH/s select against 63.71 full (+10.5 percent; select sits at the kernel timer's
70.86, the read-back cost is gone); production 53.38 against 49.71 (+7.4 percent; the rest of the gap to the kernel is
the 2.3 to 2.6 s setup of each 7.6 s run); energy per 2^24 batch 86.2 J select against 88.7 J full (-2.8 percent; the card
draws more while it is never idle, 361 W against 337 W, and finishes sooner); 0.195 against 0.188 MH per W; worker CPU
2.3 to 2.6 s against 2.9 to 3.1 s per 32 jobs. The full path pays 7.7 to 25.1 ms of read-back plus 10.3 to 19.5 ms of host
scan per 236.7 ms chunk (8 to 19 percent); the select pass costs 0.23 to 0.30 ms plus 0.18 to 0.35 ms of read-back.
### 3090
The pod's host holds the card at 210 MHz under `clocks_event_reasons.sw_power_cap` for whole runs (Active at 21 W idle;
`-lgc` and `-pl` refused in the container), so the rows split by the mean SM clock of each run.
| run | mode | mean SM MHz | setup s | production MH/s | production J per batch | steady MH/s | steady W | steady J per batch | steady MH per W | worker CPU s | per chunk: kernel+sync, select, read-back, scan ms |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 3090-full-r3 | full | 1769 | 4.110 | 22.81 | 207.7 | 27.63 | 325.9 | 197.9 | 0.0848 | 5.61 | 557.15, 0.000, 24.225, 23.753 |
| 3090-select-r2 | select | 1733 | 4.138 | 24.41 | 196.7 | 30.06 | 330.7 | 184.6 | 0.0909 | (subshell pid, not read) | 555.70, 0.268, 0.217, 0.013 |
| 3090-full-r1 | full | 632 | 4.710 | 9.18 | 301.4 | 9.98 | 168.3 | 282.9 | 0.0593 | (subshell pid, not read) | 1630.34, 0.000, 24.723, 24.119 |
| 3090-full-r2 | full | 210 | 4.231 | 5.96 | 378.4 | 6.25 | 136.9 | 367.3 | 0.0457 | (subshell pid, not read) | 2637.86, 0.000, 22.784, 20.625 |
| 3090-full-r4 | full | 210 | 4.397 | 5.96 | 377.9 | 6.27 | 136.7 | 365.9 | 0.0459 | 5.81 | 2624.25, 0.000, 27.536, 22.950 |
| 3090-select-r1 | select | 210 | 7.093 | 5.86 | 376.9 | 6.35 | 136.9 | 361.5 | 0.0464 | (subshell pid, not read) | 2637.57, 0.532, 0.296, 0.019 |
| 3090-select-r3 | select | 318 | 4.297 | 6.44 | 361.6 | 6.78 | 140.1 | 346.4 | 0.0484 | 4.23 | 2470.39, 0.503, 0.206, 0.017 |
3090 reading (the one pair at the card's clock): steady 30.06 MH/s select against 27.63 full (+8.8 percent); production
24.41 against 22.81 (+7.0 percent); 184.6 J per batch against 197.9 (-6.7 percent); the full path's read-back plus scan is
48 ms of a 557 ms chunk (8.6 percent). The 6.5 MH/s kernel read of the evening is the 210 MHz pod state, not the class:
ptxas on sm_86 shows 87 registers and no spill or stack bytes (sm_120: 92 registers, none); the worker's occupancy is 16
blocks of one warp per SM on sm_86 against 20 on sm_120, register-limited on both. At clock the 5090 to 3090 gap is 2.34x
(70.4 to 30.1), beside the memory gap (1,792 to 936 GB/s, 96 MiB to 6 MiB of L2); the 10.9x read was the power cap. A
consumer-tier number for the site needs a 3090 whose clock the host does not cap; the fleet lane holds the pod id.
## Record
Harness cells (tools/ci/test-map.json): `harness:cuda-select-check` (ROT-01 partial: the high-32 rollover case; POW-01
partial: the select pass equals the full read on every case named above, the hash sample against igneum-pow) and
`harness:pack-geometry` (POW-02 partial: the geometry rule on 19 conformance vectors in C on build-7 and in Swift on the
mini, and the two checked-in packs). Batch tools/ci/batches/f10-worker-20261008.json.
## I06 (master review, gate INT-10) and I05, landed with the rows
One normative geometry rule: packfile.h pf_load (CUDA and the OpenCL generic worker) gained the 64-bit count read with
the 2^32 refusal and the floor-log2 rule; proto-metal/main.swift gained `packGeometry` (the same rules, the same refusal
words) and sizes the day's dataset from the exact word count, refusing before any Metal work; `--geometry-check
<vectors> [pack dir...]` runs the vectors. proto-cuda/nvrtc/emu/geometry-vectors.txt is the shared file (ds55, ds85,
ds115, the power-of-two sizes, wrong count or log2, overflowing counts, missing leaves, leaves on a v2 pack), read by
emu/packfile-test.c on build-7 (19 vectors, 0 failures) and by the Metal host on the mini (19 vectors, 2 packs, 0
failures; geometry-check-mini.log). I05: every worker's found line carries a fifth field, the epoch the hash was computed
under (16 hex): CUDA and OpenCL from the pair, Metal from the job (metal-serve-smoke.log on the mini: the v4 pack
through the Swift path, the class v5 pack through the exact-geometry pack path). The miner reads a found line with four or
more fields (igneum/miner/src/main.rs:778).

17
docs/plans/igneum-2.0-test-harness-map.md
View file

@ -268,6 +268,23 @@ Rule: a case maps to a cell only where the cell's tests visibly answer it; cover
- UX-05 Keep voting keys with the miner through pooling: partial: the vote-key commitment on the wire (verify::tests::a_share_under_another_key_is_refused_before_the_hash: known-pass the member's key, known-fail another key on the share and a job naming another key, refused with code vote_key before the hash), the open pool's sidechain::check_key (the claimed key, the header's key and the reveal are one key), the same-nonce-other-template wrong_hash test, the TLS binding test; the malicious-pool run and the leave-and-retain run on the devnet-4 pair are the UX-05 batch's other evidence
- UX-04 Pay small operators without hidden custody: partial: PPLNS distribution tests (fee first, never overpays, late joiner), the payout key file round trip, the signed transfer decode; the live payout path on the devnet-4 pair is the UX-04 batch's evidence
### harness:cuda-select-check
- Command: `proto-cuda/nvrtc/emu/test.sh on a box (the CUDA worker under the CPU emulation backend): emu/select-check.sh (--readback select against the full read, known-failed first) and emu/serve-check.sh (the hash sample against igneum-pow hash-bound, both sides of the 32-bit boundary, the swap)`
- Box class: suite (CPU emulation, any build box)
- Fixtures: F0, F1
- Cases:
- ROT-01 Agree across every hourly boundary: partial: the worker's high-32 nonce rollover inside a job (a job from 2^32 - 32) finds the same set on both read-back paths and the boundary nonces hash as igneum-pow; the hourly program boundary itself is the fast-time harness
- POW-01 Match independent execution across every backend: partial: the CUDA select pass equals the full read line for line (zero hits, partial, the all-hit overflow fallback, a 96-nonce tail with a 32-lane launch, queued jobs, hits then zero, the epoch tag) and 17 sampled hashes equal igneum-pow; the million-vector campaign is harness:p01-vectors
### harness:pack-geometry
- Command: `proto-cuda/nvrtc/emu/packfile-test.sh on a box (packfile.h pf_load over emu/geometry-vectors.txt and the checked-in packs) and igneum-bench --geometry-check proto-cuda/nvrtc/emu/geometry-vectors.txt <pack dirs> on the Mac mini (proto-metal/main.swift packGeometry over the same file)`
- Box class: suite (a build box for C; the mini for Swift)
- Fixtures: F0
- Cases:
- POW-02 Validate generated programs and index folding: partial: the dataset geometry rule (exact word count, items, multiply-shift, 64-bit bytes, the floor-log2 and 2^32 refusals, missing leaves) agrees across the CUDA and OpenCL loader and the Metal host on 19 conformance vectors and the two checked-in packs (master review I06, gate INT-10); the rest of POW-02 is suite:pow
## Automated cases with no harness in the matrix (NOT RUN, the reason)
- GOV-02 Approve thresholds before results: the approval is recorded in the registry's approval field; the automated half (thresholds frozen before any run_status) is the gate rule landing by 21:00

50
docs/plans/igneum-2.0-test-registry.json
View file

@ -708,18 +708,17 @@
"manual_page": 24,
"owner_lane": "hash lane (a690540514aa453d7)",
"run_status": "RUNNING",
"evidence_path": "build-1:/srv/artefacts/tas/201-7cfa422a-aa0e0f45/node-7cfa422a/box4-miner.log",
"run_id": "201-7cfa422a-aa0e0f45",
"updated": "2026-10-08T19:32:50.856Z",
"evidence_path": "docs/analysis/review-2026-10-08-b/f10/emu-test.log",
"run_id": "f10-worker-20261008",
"updated": "2026-10-08T19:52:55.240Z",
"evidence_record": {
"cell": "suite:miner",
"manifest_sha": "7cfa422a",
"cell": "harness:cuda-select-check",
"manifest_sha": "964cdf6c1",
"coverage": {
"UX-06": "partial: the miner's own template selection tests",
"UX-04": "partial: payout label and share accounting tests; custody is the pool lane's row",
"POW-01": "partial: the pack re-check seam against the pack's 96 vectors (recheck_pack); the million-vector campaign is harness:p01-vectors"
"ROT-01": "partial: the worker's high-32 nonce rollover inside a job (a job from 2^32 - 32) finds the same set on both read-back paths and the boundary nonces hash as igneum-pow; the hourly program boundary itself is the fast-time harness",
"POW-01": "partial: the CUDA select pass equals the full read line for line (zero hits, partial, the all-hit overflow fallback, a 96-nonce tail with a 32-lane launch, queued jobs, hits then zero, the epoch tag) and 17 sampled hashes equal igneum-pow; the million-vector campaign is harness:p01-vectors"
},
"at": "2026-10-08T19:32:50.856Z"
"at": "2026-10-08T19:52:55.240Z"
}
},
{
@ -749,18 +748,16 @@
"manual_page": 24,
"owner_lane": "hash lane (a690540514aa453d7)",
"run_status": "RUNNING",
"evidence_path": "build-1:/srv/artefacts/tas/201-7cfa422a-aa0e0f45/miner-9c844503/box2-pow.log",
"run_id": "201-7cfa422a-aa0e0f45",
"updated": "2026-10-08T19:32:50.856Z",
"evidence_path": "docs/analysis/review-2026-10-08-b/f10/packfile-test.log;docs/analysis/review-2026-10-08-b/f10/geometry-check-mini.log",
"run_id": "f10-worker-20261008",
"updated": "2026-10-08T19:52:55.240Z",
"evidence_record": {
"cell": "suite:pow",
"manifest_sha": "7cfa422a",
"cell": "harness:pack-geometry",
"manifest_sha": "964cdf6c1",
"coverage": {
"POW-02": "partial: program derivation, ds55 geometry, the v6 fold, the packs and the spec read-back tests; the independent re-implementation and the index-fold census are the hash lane's harnesses",
"POW-08": "partial: the freeze list and the generator version pin; the public-claim scrub is the site gate's",
"POW-07": "the mixed FP32 branch is excluded and unreachable: no class flag reaches an FP op in igneum-pow on master (the mixedfp lane's branch is not on master); evidence the grep of the emitters on master, recorded by the hash lane"
"POW-02": "partial: the dataset geometry rule (exact word count, items, multiply-shift, 64-bit bytes, the floor-log2 and 2^32 refusals, missing leaves) agrees across the CUDA and OpenCL loader and the Metal host on 19 conformance vectors and the two checked-in packs (master review I06, gate INT-10); the rest of POW-02 is suite:pow"
},
"at": "2026-10-08T19:32:50.856Z"
"at": "2026-10-08T19:52:55.240Z"
}
},
{
@ -1353,14 +1350,17 @@
"manual_page": 30,
"owner_lane": "fast-time lane (a8be71a0db962911c)",
"run_status": "RUNNING",
"evidence_path": "v5-fasttime 92bf6a7f",
"run_id": "team-2026-10-08",
"updated": "2026-10-08 18:3x UK",
"evidence_path": "docs/analysis/review-2026-10-08-b/f10/emu-test.log",
"run_id": "f10-worker-20261008",
"updated": "2026-10-08T19:52:55.240Z",
"evidence_record": {
"what_was_run": "the fast-time crossings PASS on 4cdcc488 (17:29:58) and 617cb441 (17:30:44) with the cold restart",
"run_by": "fast-time lane (a8be71a0db962911c)",
"under_the_standard": "no: team-run before the standard's procedure; the status is RUNNING until the case is re-run under its steps with the profile's numbers and an independent run where the profile asks one",
"pass_or_fail_today": "not judged under the standard yet"
"cell": "harness:cuda-select-check",
"manifest_sha": "964cdf6c1",
"coverage": {
"ROT-01": "partial: the worker's high-32 nonce rollover inside a job (a job from 2^32 - 32) finds the same set on both read-back paths and the boundary nonces hash as igneum-pow; the hourly program boundary itself is the fast-time harness",
"POW-01": "partial: the CUDA select pass equals the full read line for line (zero hits, partial, the all-hit overflow fallback, a 96-nonce tail with a 32-lane launch, queued jobs, hits then zero, the epoch tag) and 17 sampled hashes equal igneum-pow; the million-vector campaign is harness:p01-vectors"
},
"at": "2026-10-08T19:52:55.240Z"
}
},
{

124
proto-cuda/nvrtc/emu/geometry-vectors.txt Normal file
View file

@ -0,0 +1,124 @@
# Pack geometry conformance vectors (8 October 2026, master review I06, gate INT-10): one normative rule set for the
# dataset geometry of a program pack, read by every host the same way: packfile.h pf_load (the CUDA worker and the
# OpenCL generic worker; checked by emu/packfile-test.c) and proto-metal/main.swift packGeometry (checked by
# igneum-bench --geometry-check). A host allocates, builds and hashes from the exact word count, never from the log2
# alone, with 64-bit byte sizes, and refuses an unsupported pack before any GPU launch.
#
# Each case: `case <name>`, then `define <NAME> <value>` lines appended to a base program.h that carries everything
# but the geometry (seeds, generator 2, attempt, the init words, the cache sizes, IGNEUM_DATASET_MODE 1), then one
# `expect ok words=<n> items=<n> mulshift=<0|1> bytes=<n>` or `expect refuse <a substring of the refusal>`.
# A `define IGNEUM_GENERATOR` or `define IGNEUM_STATE_LEAVES` line overrides the base's class fields.
case ds55
define IGNEUM_DATASET_LOG2 30
define IGNEUM_DATASET_WORDS 1476395008u
define IGNEUM_DATASET_ITEMS 92274688u
define IGNEUM_DATASET_BYTES 5905580032ull
define IGNEUM_DATASET_MULSHIFT 1
expect ok words=1476395008 items=92274688 mulshift=1 bytes=5905580032
case ds85
define IGNEUM_DATASET_LOG2 31
define IGNEUM_DATASET_WORDS 2281701376u
define IGNEUM_DATASET_ITEMS 142606336u
define IGNEUM_DATASET_MULSHIFT 1
expect ok words=2281701376 items=142606336 mulshift=1 bytes=9126805504
case ds115
define IGNEUM_DATASET_LOG2 31
define IGNEUM_DATASET_WORDS 3087007744u
define IGNEUM_DATASET_ITEMS 192937984u
define IGNEUM_DATASET_MULSHIFT 1
expect ok words=3087007744 items=192937984 mulshift=1 bytes=12348030976
case legacy-2^28
define IGNEUM_DATASET_LOG2 28
expect ok words=268435456 items=16777216 mulshift=0 bytes=1073741824
case legacy-2^30
define IGNEUM_DATASET_LOG2 30
expect ok words=1073741824 items=67108864 mulshift=0 bytes=4294967296
case legacy-2^31-explicit-count
define IGNEUM_DATASET_LOG2 31
define IGNEUM_DATASET_WORDS 2147483648u
define IGNEUM_DATASET_ITEMS 134217728u
define IGNEUM_DATASET_MULSHIFT 0
expect ok words=2147483648 items=134217728 mulshift=0 bytes=8589934592
case wrong-count-not-a-65536-multiple
define IGNEUM_DATASET_LOG2 30
define IGNEUM_DATASET_WORDS 1476395009u
define IGNEUM_DATASET_ITEMS 92274688u
define IGNEUM_DATASET_MULSHIFT 1
expect refuse multiple of 65,536
case wrong-log2-for-the-count
define IGNEUM_DATASET_LOG2 29
define IGNEUM_DATASET_WORDS 1476395008u
define IGNEUM_DATASET_ITEMS 92274688u
define IGNEUM_DATASET_MULSHIFT 1
expect refuse IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count
case wrong-log2-for-a-power-of-two-count
define IGNEUM_DATASET_LOG2 28
define IGNEUM_DATASET_WORDS 1073741824u
define IGNEUM_DATASET_ITEMS 67108864u
define IGNEUM_DATASET_MULSHIFT 0
expect refuse IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count
case wrong-items
define IGNEUM_DATASET_LOG2 30
define IGNEUM_DATASET_WORDS 1476395008u
define IGNEUM_DATASET_ITEMS 92274689u
define IGNEUM_DATASET_MULSHIFT 1
expect refuse IGNEUM_DATASET_ITEMS is not IGNEUM_DATASET_WORDS / 16
case wrong-mulshift-on-a-non-power-of-two
define IGNEUM_DATASET_LOG2 30
define IGNEUM_DATASET_WORDS 1476395008u
define IGNEUM_DATASET_ITEMS 92274688u
define IGNEUM_DATASET_MULSHIFT 0
expect refuse IGNEUM_DATASET_MULSHIFT does not match the word count
case log2-32
define IGNEUM_DATASET_LOG2 32
expect refuse IGNEUM_DATASET_LOG2 above 31
case overflowing-count-2^32
define IGNEUM_DATASET_LOG2 31
define IGNEUM_DATASET_WORDS 4294967296ull
define IGNEUM_DATASET_MULSHIFT 0
expect refuse IGNEUM_DATASET_WORDS above 2^32 - 1
case overflowing-count-2^40
define IGNEUM_DATASET_LOG2 31
define IGNEUM_DATASET_WORDS 1099511627776ull
expect refuse IGNEUM_DATASET_WORDS above 2^32 - 1
case count-below-the-minimum
define IGNEUM_DATASET_LOG2 15
define IGNEUM_DATASET_WORDS 32768u
expect refuse multiple of 65,536
case log2-below-20
define IGNEUM_DATASET_LOG2 19
expect refuse sizes out of range
case missing-log2
define IGNEUM_DATASET_WORDS 1476395008u
expect refuse has no IGNEUM_DATASET_LOG2
case missing-leaves-on-a-class-v5-pack
define IGNEUM_GENERATOR 5
define IGNEUM_PROGRAM_CLASS "v5"
define IGNEUM_ERA_SEED_HEX "bed7ab62cbece66cf791485336d81d90fa1452ffed28ecd8a7416960ef64164c"
define IGNEUM_SHADOW_INSTRS 256
define IGNEUM_SHADOW_REPS 27
define IGNEUM_DATASET_LOG2 28
expect refuse without IGNEUM_STATE_LEAVES
case leaves-on-a-class-v2-pack
define IGNEUM_STATE_LEAVES 93
define IGNEUM_DATASET_LOG2 28
expect refuse state leaves belong to class v5

50
proto-cuda/nvrtc/emu/packfile-test.c
View file

@ -269,6 +269,56 @@ int main(int argc, char** argv) {
}
}
// 6. The geometry conformance vectors (8 October 2026, master review I06, gate INT-10): emu/geometry-vectors.txt, one
// normative rule set shared with the Metal host (igneum-bench --geometry-check reads the same file). Each case's defines
// are appended to a base program.h with everything but the geometry; pf_load must size the pack exactly as the vector
// says (words, items, mulshift, 64-bit bytes) or refuse it with the named words. Known-failed first: the vectors file
// itself carries the refusals (an overflowing count, a log2 that is not the floor, a wrong item count).
if (argc > 3) {
FILE* vf = fopen(argv[3], "rb");
char line[512], name[128] = {0}, defs[4096] = {0}, text[8192], sw[200], kw[200];
int cases = 0, inCase = 0;
words_hex(att1, sw); words_hex(keyw, kw);
CHECK(vf != NULL, "the geometry vectors file opens");
while (vf && fgets(line, sizeof(line), vf)) {
line[strcspn(line, "\r\n")] = 0;
if (line[0] == '#' || line[0] == 0) continue;
if (strncmp(line, "case ", 5) == 0) { snprintf(name, sizeof(name), "%s", line + 5); defs[0] = 0; inCase = 1; continue; }
if (!inCase) continue;
if (strncmp(line, "define ", 7) == 0) { snprintf(defs + strlen(defs), sizeof(defs) - strlen(defs), "#define %s\n", line + 7); continue; }
if (strncmp(line, "expect ", 7) == 0) {
const char* ex = line + 7;
int hasGen = strstr(defs, "#define IGNEUM_GENERATOR ") != NULL;
char what[400];
snprintf(text, sizeof(text),
"#define IGNEUM_SEED_BYTES_HEX \"%s\"\n#define IGNEUM_DAY_BYTES_HEX \"%s\"\n%s#define IGNEUM_PROGRAM_ATTEMPT 1\n#define IGNEUM_DATASET_MODE 1\n"
"#define IGNEUM_SEEDW_INIT { %s }\n#define IGNEUM_KEY_INIT { %s }\n#define IGNEUM_CACHE_LOG2_WORDS 26\n#define IGNEUM_CACHE_SEGMENTS 4096u\n%s",
EPOCH_34, DAY_20731, hasGen ? "" : "#define IGNEUM_GENERATOR 2\n", sw, kw, defs);
write_file(dir, "program.h", text);
write_file(dir, "seeds.txt", "epoch_seed_hex " EPOCH_34 "\nday_seed_hex " DAY_20731 "\n");
err[0] = 0;
++cases;
if (strncmp(ex, "ok ", 3) == 0) {
unsigned long long w = 0, it = 0, ms = 0, by = 0;
int ok = pf_load(dir, &pk, err, sizeof(err));
sscanf(ex, "ok words=%llu items=%llu mulshift=%llu bytes=%llu", &w, &it, &ms, &by);
snprintf(what, sizeof(what), "geometry %s: loads as %llu words, %llu items, mulshift %llu, %llu bytes", name, w, it, ms, by);
CHECK(ok == 1 && pk.datasetWords == (uint32_t)w && pk.datasetItems == (uint32_t)it && pk.datasetMulshift == (uint32_t)ms && (uint64_t)pk.datasetWords * 4ull == by, what);
if (!ok) printf(" (%s)\n", err);
} else if (strncmp(ex, "refuse ", 7) == 0) {
int ok = pf_load(dir, &pk, err, sizeof(err));
snprintf(what, sizeof(what), "geometry %s: refused with \"%s\"", name, ex + 7);
CHECK(ok == 0 && strstr(err, ex + 7) != NULL, what);
if (ok) printf(" (loaded: %u words)\n", (unsigned)pk.datasetWords); else if (strstr(err, ex + 7) == NULL) printf(" (said: %s)\n", err);
} else { CHECK(0, "a vector's expect line is ok or refuse"); }
inCase = 0;
}
}
if (vf) fclose(vf);
CHECK(cases >= 18, "the vectors file carries every case (18 or more)");
printf(" %d geometry vectors\n", cases);
}
printf("%s: %d failure(s)\n", argv[0], failures);
return failures ? 1 : 0;
}

5
proto-cuda/nvrtc/emu/packfile-test.sh
View file

@ -1,6 +1,7 @@
#!/usr/bin/env bash
# The pack loader's seed rule (packfile.h) on a known-good and a known-mismatched pack, and the class v5 leaf rule on the
# checked-in v5 pack (7 October 2026): emu/packfile-test.c, C99, no GPU. Runs on the Mac in a second and in CI. Usage: emu/packfile-test.sh
# checked-in v5 pack (7 October 2026), and the geometry conformance vectors (emu/geometry-vectors.txt, 8 October 2026, shared with
# the Metal host): emu/packfile-test.c, C99, no GPU. Runs on a box in a second and in CI. Usage: emu/packfile-test.sh
set -euo pipefail
HERE="$(cd "$(dirname "$0")" && pwd)"
ROOT="$(cd "$HERE/../../.." && pwd)"
@ -8,4 +9,4 @@ OUT="${TMPDIR:-/tmp}/igneum-packfile-test"
mkdir -p "$OUT"
CC="${CC:-cc}"
"$CC" -std=c99 -Wall -Wextra -Wno-unused-function -O1 -o "$OUT/packfile-test" "$HERE/packfile-test.c"
"$OUT/packfile-test" "$ROOT/proto-cuda/packs/igneum-devnet-v4-epoch0" "$ROOT/proto-cuda/packs-ca3-v5/v5-dn3-epoch0"
"$OUT/packfile-test" "$ROOT/proto-cuda/packs/igneum-devnet-v4-epoch0" "$ROOT/proto-cuda/packs-ca3-v5/v5-dn3-epoch0" "$HERE/geometry-vectors.txt"

2
proto-cuda/nvrtc/emu/select-check.sh
View file

@ -13,6 +13,7 @@
# found_set_equals_full_stale_job_queued s1, s2: two job lines queued at once, the second served after the first
# buffer_reuse_hits_then_zero b1 then b2: a hit chunk followed by a zero-target chunk prints nothing stale
# mismatch_error_then_job m1 (no pair for its seeds) errors the same way, c1 after it is served
# epoch_tag every found line carries the pack's epoch (16 hex) as its fifth field, both modes (I05)
# Usage: select-check.sh <pack A dir> <out dir> <worker command...> (the command carries --serve --pack <pack A> etc.)
set -euo pipefail
PACK_A="$1"; OUT="$2"; shift 2
@ -74,6 +75,7 @@ grep -q '^done r1 2048' "$OUT/rb-select.norm" || { echo "FAIL: high32_rollover:
awk '$1 == "found" && $2 == "r1" && $3 + 0 >= 4294967296 { n++ } END { exit n > 0 ? 0 : 1 }' "$OUT/rb-select.norm" || { echo "FAIL: high32_rollover: no found nonce at or past 2^32"; exit 1; }
grep -q '^done s1 2048' "$OUT/rb-select.norm" && grep -q '^done s2 2048' "$OUT/rb-select.norm" || { echo "FAIL: stale_job_queued: s1 or s2 did not finish"; exit 1; }
[ "$(count b1 "$OUT/rb-select.norm")" -gt 0 ] && [ "$(count b2 "$OUT/rb-select.norm")" = 0 ] || { echo "FAIL: buffer_reuse_hits_then_zero: b1 $(count b1 "$OUT/rb-select.norm") found, b2 $(count b2 "$OUT/rb-select.norm")"; exit 1; }
awk -v e="${A_EPOCH:0:16}" '$1 == "found" && $5 != e { bad++ } END { exit bad > 0 ? 1 : 0 }' "$OUT/rb-select.norm" && awk -v e="${A_EPOCH:0:16}" '$1 == "found" && $5 != e { bad++ } END { exit bad > 0 ? 1 : 0 }' "$OUT/rb-full.norm" || { echo "FAIL: epoch_tag: a found line lacks the pack's epoch ${A_EPOCH:0:16} as its fifth field"; exit 1; }
grep -q '^error m1 ' "$OUT/rb-select.norm" && grep -q '^done c1 1024' "$OUT/rb-select.norm" || { echo "FAIL: mismatch_error_then_job: m1 did not error or c1 was not served"; exit 1; }
grep -q 'full fallbacks' "$OUT/rb-select.log" || { echo "FAIL: no transfers line in the select run"; exit 1; }
fb="$(sed -n 's/.*(select, \([0-9]*\) full fallbacks).*/\1/p' "$OUT/rb-select.log" | tail -1)"

2
proto-cuda/nvrtc/emu/serve-check.sh
View file

@ -53,6 +53,8 @@ elif grep -q '^info job 5 is for epoch .* building its pack' "$LOG" && [ "$(grep
else
echo "FAIL: job 5 was neither refused nor self-healed"; exit 1
fi
[ "$(awk -v e="${A_EPOCH:0:16}" '$1 == "found" && ($2 == 1 || $2 == 2 || $2 == 3) && $5 != e' "$LOG" | wc -l | tr -d ' ')" = 0 ] || { echo "FAIL: a found line of a pack A job does not carry epoch ${A_EPOCH:0:16} as its fifth field"; exit 1; }
[ "$(awk -v e="${B_EPOCH:0:16}" '$1 == "found" && $2 == 4 && $5 != e' "$LOG" | wc -l | tr -d ' ')" = 0 ] || { echo "FAIL: a found line of job 4 (pack B after the swap) does not carry epoch ${B_EPOCH:0:16} as its fifth field"; exit 1; }
grep -q '^found 2 4294967295 ' "$LOG" && grep -q '^found 2 4294967296 ' "$LOG" || { echo "FAIL: the 32-bit boundary nonces are missing"; exit 1; }
for x in 0 1 2 31 32 63; do check 1 $x "$A_EPOCH" "$A_DAY"; done
for x in 4294967264 4294967295 4294967296 4294967327; do check 2 $x "$A_EPOCH" "$A_DAY"; done

17
proto-cuda/nvrtc/packfile.h
View file

@ -302,8 +302,23 @@ static int pf_load(const char* dir, PfPack* pk, char* err, size_t cap) {
if (!pf_define_u32(prog, "IGNEUM_DATASET_LOG2", &pk->datasetLog2)) { free(prog); return pf_fail(err, cap, "program.h has no IGNEUM_DATASET_LOG2"); }
if (!pf_define_u32(prog, "IGNEUM_DATASET_MODE", &pk->datasetMode)) pk->datasetMode = 0;
if (pk->datasetLog2 > 31) { free(prog); return pf_fail(err, cap, "program.h IGNEUM_DATASET_LOG2 above 31: the word index is 32-bit"); }
if (!pf_define_u32(prog, "IGNEUM_DATASET_WORDS", &pk->datasetWords)) pk->datasetWords = 1u << pk->datasetLog2;
{
// the word count is parsed at 64 bits first: a count at or above 2^32 (ds115 is 3,087,007,744 words, under it) would wrap
// in the 32-bit field and size the dataset at a fraction of the pack (the geometry vectors, emu/geometry-vectors.txt, 8 October 2026)
uint64_t w64 = 0;
if (pf_define_u64(prog, "IGNEUM_DATASET_WORDS", &w64)) {
if (w64 > 0xffffffffull) { free(prog); return pf_fail(err, cap, "program.h IGNEUM_DATASET_WORDS above 2^32 - 1: the word index is 32-bit"); }
pk->datasetWords = (uint32_t)w64;
} else pk->datasetWords = 1u << pk->datasetLog2;
}
if (pk->datasetWords < (1u << 16) || (pk->datasetWords & 0xffffu) != 0) { free(prog); return pf_fail(err, cap, "program.h IGNEUM_DATASET_WORDS is not a multiple of 65,536 words"); }
{
// IGNEUM_DATASET_LOG2 is the floor of log2 of the count (the exporter writes it so; a reader that sizes by the log2 alone
// is then at most a factor of two short, never wrong by a different pack's geometry): a count that disagrees is refused
uint32_t fl = 0, w = pk->datasetWords;
while (w > 1u) { w >>= 1; ++fl; }
if (fl != pk->datasetLog2) { free(prog); return pf_fail(err, cap, "program.h IGNEUM_DATASET_LOG2 is not the floor of log2 of the word count"); }
}
if (!pf_define_u32(prog, "IGNEUM_DATASET_ITEMS", &pk->datasetItems)) pk->datasetItems = pk->datasetWords / 16u;
if (pk->datasetItems != pk->datasetWords / 16u) { free(prog); return pf_fail(err, cap, "program.h IGNEUM_DATASET_ITEMS is not IGNEUM_DATASET_WORDS / 16"); }
if (!pf_define_u32(prog, "IGNEUM_DATASET_MULSHIFT", &pk->datasetMulshift)) pk->datasetMulshift = 0;

10
proto-cuda/nvrtc/serve-bench.sh
View file

@ -21,9 +21,13 @@ LOG="$OUT/$RUN.worker.log"; TL="$OUT/$RUN.timeline"; SMI="$OUT/$RUN.smi"; ROW="$
now() { date +%s.%N; }
count=$((1 << B))
target="00000fffffffffff" # one hit per 2^20 nonces: 16 found lines per 2^24-nonce job, the serving shape (a block target is far tighter)
seedline() { sed -n "s/^$2 //p" "$1/seeds.txt"; }
# the pack's seeds: seeds.txt as igneum-miner export-pack writes it, else the byte seeds program.h carries (a kit pack)
seedline() { sed -n "s/^$2 //p" "$1/seeds.txt" 2>/dev/null || true; } # a missing file is not a failure under set -e
define() { sed -n "s/^#define $2 \"\(.*\)\"/\1/p" "$1/program.h" 2>/dev/null || true; }
EPOCH="$(seedline "$PACK" epoch_seed_hex)"; DAY="$(seedline "$PACK" day_seed_hex)"
[ -n "$EPOCH" ] && [ -n "$DAY" ] || { echo "no seeds.txt in $PACK" >&2; exit 2; }
[ -n "$EPOCH" ] || EPOCH="$(define "$PACK" IGNEUM_SEED_BYTES_HEX)"
[ -n "$DAY" ] || DAY="$(define "$PACK" IGNEUM_DAY_BYTES_HEX)"
[ -n "$EPOCH" ] && [ -n "$DAY" ] || { echo "no seeds.txt and no IGNEUM_SEED_BYTES_HEX / IGNEUM_DAY_BYTES_HEX in $PACK/program.h" >&2; exit 2; }
jobline() { printf 'job %d %064x %s %d %d %s %s\n' "$1" "$1" "$target" $(( $1 * count )) "$count" "$EPOCH" "$DAY"; }
# power sampler first (its own pid file)
@ -31,7 +35,7 @@ nvidia-smi -i "$DEV" --query-gpu=timestamp,power.draw,clocks.sm,temperature.gpu
echo $! > "$PIDDIR/$RUN.smi.pid"
sleep 1
t0=$(now)
coproc WK { IGNEUM_READBACK="$MODE" "$W" --serve --pack "$PACK" --device "$DEV" --batch-log2 "$B" --race off --readback "$MODE" 2>&1; }
coproc WK { exec env IGNEUM_READBACK="$MODE" "$W" --serve --pack "$PACK" --device "$DEV" --batch-log2 "$B" --race off --readback "$MODE" 2>&1; } # exec: the pid is the worker's, not a subshell's
wkpid=$WK_PID # bash clears WK_PID when the coproc ends; the copy outlives it
echo "$wkpid" > "$PIDDIR/$RUN.worker.pid"
sent=0; done_n=0; found=0; tready=""; tlast=""; cpu=""

9
proto-cuda/nvrtc/worker.cpp
View file

@ -18,8 +18,9 @@
// prepare <epoch_seed_hex 64> <day_seed_hex> <pack_dir> compile that pack in the background, build its cache
// and dataset, self-test it; a job on it then switches
// quit
// stdout: ready cuda <device> pack <seed string> dataset-log2 N batch B regs R prepare 1 path nvrtc ...
// found <job_id> <nonce u64> <hash_hex 16>
// stdout: ready cuda <device> pack <seed string> dataset-log2 N batch B regs R prepare 1 path nvrtc ... readback select|full ...
// found <job_id> <nonce u64> <hash_hex 16> <epoch_hex 16> the fifth field (8 October 2026, I05): the epoch of the pair
// the hash was computed under, so stale work is never read as current
// done <job_id> <hashes> <ms>
// error <job_id> <text>
// need <epoch_seed_hex> <day_seed_hex> before the mismatch error: the pair this worker lacks (the miner prepares it)
@ -1416,12 +1417,12 @@ static int runServe(Ctx& c, const Options& o, Pair* cur) {
}
for (uint32_t a = 0; a < nHits; ++a) {
uint64_t nonce = ((uint64_t)hi << 32) | (uint64_t)(uint32_t)(lo + (uint32_t)sel.hHits[2 * a]);
std::printf("found %s %llu %016llx\n", jobId.c_str(), (unsigned long long)nonce, (unsigned long long)sel.hHits[2 * a + 1]);
std::printf("found %s %llu %016llx %.16s\n", jobId.c_str(), (unsigned long long)nonce, (unsigned long long)sel.hHits[2 * a + 1], cur->epochHex.c_str());
}
} else {
for (uint32_t i = 0; i < chunk; ++i) if (hOut[i] <= target) {
uint64_t nonce = ((uint64_t)hi << 32) | (uint64_t)(uint32_t)(lo + i);
std::printf("found %s %llu %016llx\n", jobId.c_str(), (unsigned long long)nonce, (unsigned long long)hOut[i]);
std::printf("found %s %llu %016llx %.16s\n", jobId.c_str(), (unsigned long long)nonce, (unsigned long long)hOut[i], cur->epochHex.c_str());
}
}
scanMs = wallMs() - r0;

11
proto-opencl/host.c
View file

@ -1156,6 +1156,13 @@ typedef struct {
cl_uint stateLeaves; /* class v5: the state leaves the dataset was built from (uploaded for the build, released after); 0 otherwise */
} ServePair;
/* The found line's fifth field (8 October 2026, I05): the epoch the pair's program was built for, 16 hex, so stale work is never read
* as current; the compiled-in pack's epoch comes from its program.h. */
#ifndef IGNEUM_SEED_BYTES_HEX
#define IGNEUM_SEED_BYTES_HEX "0000000000000000"
#endif
static const char* pairEpochTag(const ServePair* p) { return p->epochHex[0] ? p->epochHex : IGNEUM_SEED_BYTES_HEX; }
static int hexEq(const char* a, const char* b) {
size_t i;
if (strlen(a) != strlen(b)) return 0;
@ -1929,12 +1936,12 @@ static int runServe(Device* dv, const DeviceInfo* di, const Options* o) {
for (a = 0; a < nHits; ++a) {
uint32_t idx = (uint32_t)hHits[2 * a];
unsigned long long nonce = ((unsigned long long)hi << 32) | (unsigned long long)(uint32_t)(lo + idx);
printf("found %s %llu %016llx\n", jobId, nonce, (unsigned long long)hHits[2 * a + 1]);
printf("found %s %llu %016llx %.16s\n", jobId, nonce, (unsigned long long)hHits[2 * a + 1], pairEpochTag(cur));
}
} else {
for (i = 0; i < chunk; ++i) if (hOut[i] <= target) {
unsigned long long nonce = ((unsigned long long)hi << 32) | (unsigned long long)(uint32_t)(lo + i);
printf("found %s %llu %016llx\n", jobId, nonce, (unsigned long long)hOut[i]);
printf("found %s %llu %016llx %.16s\n", jobId, nonce, (unsigned long long)hOut[i], pairEpochTag(cur));
}
}
scanMs = wallMs() - r0;

17
tools/ci/batches/f10-worker-20261008.json Normal file
View file

@ -0,0 +1,17 @@
{
"run_id": "f10-worker-20261008",
"manifest_sha": "964cdf6c1",
"evidence_dir": "docs/analysis/review-2026-10-08-b/f10",
"cells": [
{
"cell": "harness:cuda-select-check",
"status": "RUNNING",
"evidence": "docs/analysis/review-2026-10-08-b/f10/emu-test.log"
},
{
"cell": "harness:pack-geometry",
"status": "RUNNING",
"evidence": "docs/analysis/review-2026-10-08-b/f10/packfile-test.log;docs/analysis/review-2026-10-08-b/f10/geometry-check-mini.log"
}
]
}

29
tools/ci/test-map.json
View file

@ -459,6 +459,35 @@
"UX-05": "partial: the vote-key commitment on the wire (verify::tests::a_share_under_another_key_is_refused_before_the_hash: known-pass the member's key, known-fail another key on the share and a job naming another key, refused with code vote_key before the hash), the open pool's sidechain::check_key (the claimed key, the header's key and the reveal are one key), the same-nonce-other-template wrong_hash test, the TLS binding test; the malicious-pool run and the leave-and-retain run on the devnet-4 pair are the UX-05 batch's other evidence",
"UX-04": "partial: PPLNS distribution tests (fee first, never overpays, late joiner), the payout key file round trip, the signed transfer decode; the live payout path on the devnet-4 pair is the UX-04 batch's evidence"
}
},
"harness:cuda-select-check": {
"command": "proto-cuda/nvrtc/emu/test.sh on a box (the CUDA worker under the CPU emulation backend): emu/select-check.sh (--readback select against the full read, known-failed first) and emu/serve-check.sh (the hash sample against igneum-pow hash-bound, both sides of the 32-bit boundary, the swap)",
"box_class": "suite (CPU emulation, any build box)",
"fixtures": [
"F0",
"F1"
],
"cases": [
"ROT-01",
"POW-01"
],
"coverage": {
"ROT-01": "partial: the worker's high-32 nonce rollover inside a job (a job from 2^32 - 32) finds the same set on both read-back paths and the boundary nonces hash as igneum-pow; the hourly program boundary itself is the fast-time harness",
"POW-01": "partial: the CUDA select pass equals the full read line for line (zero hits, partial, the all-hit overflow fallback, a 96-nonce tail with a 32-lane launch, queued jobs, hits then zero, the epoch tag) and 17 sampled hashes equal igneum-pow; the million-vector campaign is harness:p01-vectors"
}
},
"harness:pack-geometry": {
"command": "proto-cuda/nvrtc/emu/packfile-test.sh on a box (packfile.h pf_load over emu/geometry-vectors.txt and the checked-in packs) and igneum-bench --geometry-check proto-cuda/nvrtc/emu/geometry-vectors.txt <pack dirs> on the Mac mini (proto-metal/main.swift packGeometry over the same file)",
"box_class": "suite (a build box for C; the mini for Swift)",
"fixtures": [
"F0"
],
"cases": [
"POW-02"
],
"coverage": {
"POW-02": "partial: the dataset geometry rule (exact word count, items, multiply-shift, 64-bit bytes, the floor-log2 and 2^32 refusals, missing leaves) agrees across the CUDA and OpenCL loader and the Metal host on 19 conformance vectors and the two checked-in packs (master review I06, gate INT-10); the rest of POW-02 is suite:pow"
}
}
},
"not_run": {