Digital Drosophila Archive
30 lines
2.5 KiB
Markdown
30 lines
2.5 KiB
Markdown
# V100 conditioning acceleration — engineering replay result
|
||
|
||
Status: **PASS-ELIGIBILITY-UPDATE-REPLAY-EQUIVALENCE**. This is not a learning result or a production-training promotion.
|
||
|
||
Benchmark code commit: `221e1f13f758ebe8d9f188bab976c361f7cacf45`.
|
||
|
||
| Device | Real-fixture replay median (prep + transfers + comparison) | Resident CUDA kernel median |
|
||
|---|---:|---:|
|
||
| Tesla V100-SXM2-16GB (v100a-02) | 3.430 ms | 0.122 ms |
|
||
| Tesla V100-SXM2-16GB (v100b-01) | 3.344 ms | 0.122 ms |
|
||
| NVIDIA GeForce RTX 4090 (rtx4090-01) | 3.327 ms | 0.074 ms |
|
||
|
||
Each GPU passed14sealed real fixtures +16fixed synthetic cases, with sampled trace, final trace, allUS snapshots and delta byte-identical to the original independent CPU oracle. Six invalid inputs were rejected on each GPU.
|
||
|
||
CPU warm reference includes the original API's repeated provenance checks and AST parsing; it is only one warm sample perfixture. GPU authenticates the source and complete anatomical mapping once. Therefore ratios in JSON describe API implementation replay, not pure hardware acceleration.
|
||
|
||
The CUDA kernels preserve every timestep, float32 separate round-to-nearest operations, exact US timing, noUS zeros, compartment/frozen/D0 identity and the depression floor. No gain/learning-rate/tau/threshold/connectome/sign changes. No dense time×neuron matrix.
|
||
|
||
A new candidate-only CuPy environment and explicit CUDA12.6 nvcc compilation are used. Original science and ComfyUI environments are unchanged. All three ComfyUI queues were checked before each fixture; no services were stopped.
|
||
|
||
Integrity: 64034 saved files, 3103901058 bytes, 16 registered source files checked; no mismatch. This is the saved Phase4AR inventory, not a new Phase1–4AQ numerical regression.
|
||
|
||
Phase4AR remains stopped at1033/30000CPU-verified epochs and138/3920probes. No weights were installed and no training was resumed.
|
||
|
||
Next: a separately validated persistent-worker pipeline and end-to-end benchmark. This local optimization does not justify dividing the9-day estimate by the kernel speedup. Independent-seed V100 parallelism is not yet measured.
|
||
|
||
Preserved failures: candidate packaging incompatibility (--report unsupported bypip22); first CuPy compile FTZ flag conflict (before any kernel ran). Both were corrected only in the independent branch.
|
||
|
||
Compiler references: [CuPy precompiled RawModule](https://docs.cupy.dev/en/v13.6.0/reference/generated/cupy.RawModule.html), [CUDA12.6 compiler options](https://docs.nvidia.com/cuda/archive/12.6.3/cuda-compiler-driver-nvcc/index.html).
|