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

bench-log: the real RTX 4070 alone (7,553 MiB compressed in 7.7 s at 2^26; core-only 5,761 MiB at 2^25): a 12 GB card proves alone, measured on the card

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-06 12:01:54 +00:00
parent 16f02e19ce
commit 25d36336a2
1 changed files with 30 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
docs/bench-log.md
View file

@ -1657,6 +1657,36 @@ a pool rule, 20x the bytes a shard on the relay; nothing on chain. Every number
the real 12 GB card (route 1, `tools/prover-floor/pc-12gb-card.ps1`, kit `igneum-prove-wsl2-floor.zip`
e377c7df...) is the measurement that decides the public line.
Route 1, THE REAL 12 GB CARD (6 October 2026, the project lead's word at 11:25Z through main; an ASUS Dual RTX 4070 OC 12 GB,
12,282 MiB, nvidia-smi index 1, PCIe 4.0 x4 through a USB4 enclosure, mining beside the 5090 in PC 1). PC 1 already
had WSL2 Ubuntu-24.04 with nvcc 12.8 and cargo 1.99 (jobs `floor-pc1-wsl2-enable`, elevated, changed nothing, and
`floor-pc1-toolchain`, 11:21 to 11:23Z); the v4 server built cold in 353 s (`floor-pc1-build`, 11:25:33 to 11:31:34Z,
480 crates, `ed7b1e4b862c451b4a2cefa0e544f7b6a9e04279690c3f8c9881666bd7f8f23a`, sm_86/89/120, sm_89 native for the
4070); the kit `igneum-prove-wsl2-floor.zip` (e377c7df...) fetched (`floor-pc1-kit`, 11:32:49Z) and the host built
from it on PC 1 (adcad5ea..., the pinned ids matching). Every proof steered to the 4070 by `IGNEUM_CUDA_DEVICE=1`
(the host change: the SDK starts the server with `CUDA_VISIBLE_DEVICES=1`); `CUDA_DEVICE_ORDER=PCI_BUS_ID`.
Alone (job `card12-alone`, 11:53:01 to 12:00:53Z, the miners stopped by the job, the prover off then on; the 4070's
own idle **0 MiB** read before and after, so every peak is the server's own; `--mode shard` = the core proof, its
CPU verify, then the compressed proof, the sampler split at the core RESULT; every proof VERIFIED):
| Threshold | Fixture | Cycles | Core-only peak MiB | Full compressed peak MiB | Core s | Compressed s | Core proof bytes | CPU verify s |
|---|---|---|---|---|---|---|---|---|
| 2^26 | v1 shard | 4,717,439 | 7,265 | **7,553** | 3.7 | **7.7** | 14,379,043 | 0.432 |
| 2^26 | block-72854 (the "empty block first" fixture is NOT empty: a 5x bigger shard) | 23,594,943 | 7,041 | 7,617 | 17.2 | 35.9 | 69,847,451 | 2.204 |
| 2^27 | v1 shard | 4.7 M | 10,177 | 10,177 | 3.1 | 5.5 | 10,147,581 | 0.308 |
| 2^27 | block-72854 | 23.6 M | 10,017 | 10,241 | 12.9 | 22.8 | 35,910,731 | 1.120 |
| 2^25 | v1 shard | 4.7 M | **5,761** | 7,617 | 5.2 | 12.1 | 25,621,187 | 0.793 |
| 2^25 | block-72854 | 23.6 M | 5,761 | 7,649 | 31.0 | 74.3 | 164,004,701 | 5.248 |
Reading. On the card itself a 12 GB card PROVES ALONE: the compressed v1 shard at 2^26 holds 7.55 GB of 12.28 GB
(4.7 GB spare) in 7.7 s, at 2^27 10.2 GB (2.1 GB spare) in 5.5 s, and a 23.6 M-cycle shard fits at 2^26 in 36 s.
The 5090's allocation pattern overstated the 4070 by about 0.65 GB (8.2 GB "own" against 7.55 measured). The
4070 proves the v1 shard 1.35x slower than the 5090 alone (7.7 against 5.7 s at 2^26) through a PCIe 4.0 x4 link.
Core-only at 2^25 is 5.76 GB on the card. Consequence for the public line: "12 GB proves (prove-only)" is now
measured on a 12 GB card, so the line moves from 24 GB once the patched server ships (the packaging row before
0.3.12); mine-and-prove is the beside-the-miner job's row (below, `card12-miner`).
Consequences (the rule of 5 October 2026), as they stood after sweep 1 (superseded by the reading above for the 12 and 16 GB tiers): the shipped SP1 GPU server refuses every
card under 24 GB before allocating, so 8, 12 and 16 GB NVIDIA cards cannot prove on it whatever the shard; the v1
patch takes the shard's term out (20.5 to 12.7 GB on the v1 shard) at a 1.26x time cost (5.3 s against 4.2 s,