d8dc8c6ba3
infra/cloud-devnet: hcloud (doctl variant) create, builder-VM provision from a git-archive source tarball, systemd units for igneumd --devnet-suffix with a sparse --addpeer mesh and a CPU trickle miner per node, stdlib wRPC client, experiments (latency, partition, hop, collect, observer hookup), README with the command sequence and the Hetzner API prices of 3 Oct 2026. infra/gpu-bench: RunPod image recipes (CUDA 12.8, ROCm), bundle, run.sh (vectors gate, 10-min raw, sweep, inline shortcut ratio, nvcc/NVRTC/OpenCL recompile timings, results row, intake upload), bench-log template. infra/seed-nodes: create-seed (persistent IPv4, firewall), provision on the VM, health check, addPeer from the Mac over grpcurl, seeds.txt; igneum-seed-1 created at 188.245.5.161 (Hetzner cx23, fsn1). docs/plans/cloud-devnet.md and docs/plans/seed-nodes.md. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2.5 KiB
October 2026, rented GPUs: RTX 3060, 3090, 4090, 5090 on RunPod (infra/gpu-bench/run.sh; AMD RX 7900 XTX pending a host)
Image: nvidia/cuda:12.8.1-devel-ubuntu22.04 (RunPod, SSH), bundle of proto-cuda/host.cu plus the three packs and proto-opencl/host.c at commit , nvcc -arch=native, 1 warp per block. Pack igneum-genesis-mh (memory-hard, 104 loads per hash), 1 GiB dataset, batches of 2^24; the raw column is one run sized to about 10 minutes. Sweep sizes 64 to 1024 MiB. Inline = the shortcut kernel of make-inline.sh (every dataset load recomputed from the 256 MiB cache through mh_word), vectors PASS, dataset self-test FAIL by construction. Recompile = nvcc -cubin of kernel.cu out of process (the worker's prepare path), NVRTC in process (nvrtc-time.cu), clBuildProgram of kernel.cl where NVIDIA's OpenCL ICD was present. Cost: RunPod community cloud, USD for the four pods (pricing page: 3090 0.22/h, 4090 0.34/h, 5090 0.69/h; 3060 ).
| card | driver / toolkit | date | Mhash/s at 1 GiB (10 min) | s | igneum-hourly Mhash/s | sweep MiB:Mhash/s 64 / 128 / 256 / 512 / 1024 | 64 MiB over 1 GiB | inline Mhash/s | inline / honest | nvcc cubin ms | NVRTC ms | OpenCL build ms | vectors | build ms |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 3060 | ||||||||||||||
| RTX 3090 | ||||||||||||||
| RTX 4090 | ||||||||||||||
| RTX 5090 | ||||||||||||||
| RTX 5090 (the project lead's PC, 3 Oct 2026, for reference) | 13.4 / 12.8 | 2026-10-03 | 229 (memory-hard) | 185.3 (closed form) | 1340 / n/a / 270 / 242 / 229 (closed form) | 5.9 | n/a | n/a | n/a | n/a | n/a | 96/96 PASS | ||
| Apple M5 Max (Metal, reference) | 2026-10-03 | 45.2 | 36.6 | 569 (4 MiB) / 183 / 94 / 69 / 44 | 9.49 | 0.21 | 20 to 52 (Metal compile) | PASS |
Reading per column: the 1 GiB rate is the mining rate of the card on today's hash; the 64 MiB over 1 GiB ratio is the L2 cliff (ledger M1: the on-chip-cache advantage a chip would have to buy in DRAM); inline over honest is the recompute attacker's rate relative to an honest miner (ledger M16; the M5 Max gives 0.21; a value above 1 on any card means the shortcut beats the honest kernel there); the three compile columns are the hourly program change on real NVIDIA drivers (ledger M11, M17; the 90 s gap of the Windows launcher was a rebuild, not a compile). The rows are raw bench numbers from a rented host with whatever neighbours it had; rerun before quoting a figure outside this log.