Skip to content

docs(research): CUDA 13.3 impact assessment for VMAFX fork - #63

Merged
lusoris merged 1 commit into
masterfrom
research/cuda-13-3-impact-20260528
May 28, 2026
Merged

lusoris merged 1 commit into
masterfrom
research/cuda-13-3-impact-20260528

Conversation

@lusoris

@lusoris lusoris commented May 28, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Adds docs/research/research-0734-cuda-13.3-impact-assessment.md — a scoping digest for the CUDA 13.3 toolkit bump, covering Tile C++ kernel rewrites, CompileIQ autotuning, green contexts for vmafx-server, DLPack interop in the ONNX Runtime path, and NVCR/Jimver image availability.
  • Per-item recommendation table: wait-for-benchmark (Tile C++), do-now pilot (CompileIQ on filter1d.cu), wait-for-Torch-TRT (green contexts), not-worth-it yet (DLPack), wait-for-NVCR (image availability).
  • Includes LOC + complexity estimates per CUDA kernel candidate measured against the live core/src/feature/cuda/ tree (3 049 LOC across the candidate set; 845-LOC filter1d.cu is the strongest Tile-C++ structural match).
  • All cited claims survived a 3-vote adversarial verification pass; refuted claims (e.g., the "Tile C++ is performance-portable without per-arch tuning" marketing framing, several DLPack-in-ORT historical claims) are deliberately omitted.

Test plan

  • Reviewer: confirm per-kernel LOC numbers against wc -l core/src/feature/cuda/**/*.cu
  • Reviewer: confirm grep -rn 'DLPack\|dlpack' core/src/dnn/ returns no hits (digest's "current state" baseline)
  • No code changes; doc-only PR; no build / test gates apply beyond markdown linting

Deliverables (ADR-0108)

  • research digest: this PR is the digest
  • decision matrix: per-item recommendation table in §"Per-item recommendations"
  • no AGENTS.md invariant note needed: no rebase-sensitive invariants changed
  • reproducer / smoke command: wc -l core/src/feature/cuda/**/*.cu and grep -rn 'DLPack' core/src/dnn/
  • no changelog fragment needed: docs-research-only
  • no rebase impact: docs-research-only

🤖 Generated with Claude Code

@lusoris
lusoris force-pushed the research/cuda-13-3-impact-20260528 branch from 78d2247 to 059c4b9 Compare May 28, 2026 19:38
Per-item scoping for the CUDA 13.3 toolkit bump: Tile C++ rewrite
(wait for benchmark), CompileIQ autotune (pilot on filter1d.cu), green
contexts for vmafx-server (wait for Torch-TRT PR #4232 class of bugs
to clear), DLPack in ORT (not worth it until tiny-AI training sidecar
lands), and image availability (gating on NVCR + Jimver). Includes LOC
and complexity estimates per candidate CUDA kernel and adversarially
verified source citations.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@lusoris
lusoris force-pushed the research/cuda-13-3-impact-20260528 branch from 059c4b9 to c9a106d Compare May 28, 2026 21:55
@lusoris
lusoris merged commit 733b4fc into master May 28, 2026
17 of 20 checks passed
@lusoris
lusoris deleted the research/cuda-13-3-impact-20260528 branch May 28, 2026 22:34
@lusoris lusoris added this to the 1.0.0 — First release milestone Sep 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant