{
 "schema": "computelab.verified-kernel/1",
 "issued": "2026-10-08",
 "kernel": {
  "name": "chunked matmul overlapped with a copy engine transfer on one GPU",
  "language": "Python and PyTorch, cuda streams backend",
  "source_sha256": "6abd905977056bb53c585168acfe5d878d29928630d3fbf5b1a93b464343da70"
 },
 "release": {
  "version": "1.0.0"
 },
 "tested_on": {
  "chip": "Tesla T4 x1",
  "arch": "sm_75",
  "sdk": "PyTorch 2.11.0+cu130, cuda streams",
  "host": "Linux x86_64"
 },
 "shapes": [
  "a 4096 by 4096, b 4096 by 4096, float32, rows split into chunks",
  "timing a 4096 by 4096, b 4096 by 4096, float32, 64 MiB device to host copy"
 ],
 "edge_cases": [
  "one GPU, compute stream and a side copy stream",
  "4096 rows split into row chunks, each written into the shared output",
  "copied payload compared with a snapshot taken before the copy"
 ],
 "correctness": {
  "result": "pass",
  "rounds": 9,
  "reference": "unchunked a @ b compared with the row chunked product computed under a side stream copy, and 64 sampled rows per run against a float64 CPU product, largest difference 1.9e-05 against the reference, pass bound 1e-03, and 7.3e-06 against float64, pass bound 1e-03",
  "fresh_inputs": true
 },
 "timing": {
  "metric": "exposed copy time per step, CUDA streams pass",
  "unit": "ms",
  "before": 2.518,
  "after": 0.51,
  "before_label": "matmul then copy, compute 35.353 ms, copy 5.141 ms",
  "after_label": "matmul with the copy on a side stream through the copy engine, CUDA streams pass",
  "method": "median of warm steps on one GPU, step time minus pure compute time"
 },
 "scope_note": "CUDA streams pass on the host listed, one GPU. Covers this source and release. A multi GPU NCCL run earns its own certificate.",
 "certificate_id": "CLV-6656C-166A2-002EF-BC989",
 "signature": {
  "alg": "Ed25519",
  "value": "uJrZuJhHOIw22LsThDQFKRLtHsR5088Bdb3My/b1r8YeN5awo6dwPbvY97fIXrrv+Jw3q7MABEial9wQ98KDDQ==",
  "public_key": "yxxzKcfrXLqA9OhWtgA2RO3nwuuY07Ok+jFtRTMsvCI=",
  "key_id": "de06dc9d88379642"
 }
}