Skip to content

ADR-0738: Bump local CUDA toolkit pin to 13.3 + R610 minimum driver (partial — CI deferred)

  • Status: Accepted
  • Date: 2026-05-28
  • Deciders: lusoris
  • Tags: cuda, build, container, ci, deps

Context

CUDA 13.3 was released as GA on or before 2026-05-28. Confirmed facts:

  • Minimum host driver: R610.43.02 (Linux x86_64).
  • New SM targets: Blackwell sm_100, sm_103, sm_120 added. sm_103 is new in 13.3 (GB203 die — GeForce RTX 5070/5060 Ti class).
  • nvcc C++23: CUDA 13.3 nvcc adds first-class C++23 support.
  • apt availability: cuda-toolkit-13-3_13.3.0-1_amd64.deb confirmed present in the NVIDIA ubuntu2404/x86_64 apt repository (verified 2026-05-28 via developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/ listing).

Two distribution channels are not yet updated as of 2026-05-28:

  • Jimver/cuda-toolkit GitHub Action (src/links/linux-links.ts): highest available version is 13.2.0; 13.3.0 not yet present (verified).
  • NVCR nvcr.io/nvidia/cuda: Docker Hub query returns count: 0 for any tag containing 13.3; no 13.3.x-devel-ubuntu24.04 tag published (verified).

The partial bump approach unblocks local development (dev container + host builds with the R610 toolchain) while keeping CI stable on 13.2 until the GitHub Action is updated.

Decision

Bump immediately (apt is available):

  1. dev/Containerfile: replace cuda-toolkit-13-2 with cuda-toolkit-13-3. Update the comment block to reference R610.43.02 minimum driver and CUDA 13.3 changelog items (sm_103, nvcc C++23). Supersedes ADR-0573.
  2. docs/backends/cuda/overview.md: add sm_103 row to the gencode table (gated nvcc ≥ 13.3); add minimum driver section with R610.43.02 note; add CUDA 13.3 toolkit notes section; reference this ADR.
  3. docs/development/build-flags.md: update enable_cuda row to note R610.43.02 minimum with CUDA 13.3 toolkit.
  4. core/AGENTS.md: add rebase-sensitive invariant for the CUDA 13.3 / R610 contract so future rebases preserve the three-file update rule.

Defer (not yet available):

  • CI Jimver/cuda-toolkit pin: stays on cuda: '13.2.0' in .github/workflows/libvmaf-build-matrix.yml until 13.3.0 appears in Jimver/cuda-toolkit src/links/linux-links.ts.
  • NVCR nvcr.io/nvidia/cuda base image: stays on the 13.2 tag in any Dockerfile stages that use the NVCR image directly, until NVIDIA publishes a 13.3.x-devel-ubuntu24.04 tag.

Alternatives considered

Option Verdict
Bump everything blind (apt + Jimver + NVCR) Rejected — Jimver and NVCR do not yet carry 13.3. Changing the CI workflow pin to an unsupported version would break all CUDA CI legs immediately.
Wait for Jimver and NVCR before bumping anything Rejected — the apt package is already available and local users (R610 driver, CUDA 13.3 toolkit installed) are blocked on the stale container. There is no reason to delay the apt / doc bump.
Partial bump (chosen) Unblocks local development and documents the CI gap explicitly. CI follows when the action is updated; a future follow-up PR closes the gap.
Bump only docs, leave Containerfile Under-delivers — the dev container is the primary build environment (CLAUDE §12 r15); leaving it on 13.2 while docs say 13.3 creates user confusion.

Consequences

  • Positive:
  • Dev container and local host builds use the GA CUDA 13.3 toolkit.
  • sm_103 gencode coverage for Blackwell RTX 5070/5060 Ti class GPUs.
  • Documentation accurately reflects the R610.43.02 minimum driver for 13.3.

  • No CI breakage — CI continues using the stable 13.2 Jimver pin.

  • Negative:

  • Local container and CI are briefly on different CUDA toolkit versions (13.3 vs 13.2). This is the same intentional split as the 13.1→13.2 transition (ADR-0573). The containers use the same upstream apt packages; no compatibility gap has been observed.

  • Neutral / follow-ups:

  • When 13.3.0 appears in Jimver/cuda-toolkit src/links/linux-links.ts, open a follow-up PR to bump .github/workflows/libvmaf-build-matrix.yml and update this ADR's status note.
  • When NVCR publishes 13.3.x-devel-ubuntu24.04, update any Dockerfile stages that reference the NVCR CUDA base image.
  • Consider enabling --std c++23 in core/src/meson.build for CUDA TUs as a separate opt-in PR once the 13.3 toolkit is confirmed stable in CI.

References

  • req: "Bump what's actually bumpable NOW; defer what isn't. CUDA 13.3 GA released; apt repo confirmed; Jimver and NVCR not yet updated."
  • NVIDIA apt repo listing (verified 2026-05-28): developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/ — cuda-toolkit-13-3_13.3.0-1_amd64.deb present
  • Jimver src/links/linux-links.ts (verified 2026-05-28): highest version 13.2.0; 13.3.0 absent
  • Docker Hub nvidia/cuda v2 API (verified 2026-05-28): count: 0 for 13.3 name filter
  • ADR-0573 — superseded for the CUDA toolkit pin; ubuntu:26.04 base and other SDK decisions remain valid
  • ADR-0603 — belt-and-suspenders -D__MATH_NO_INLINES nvcc flag retained; ADR-0603 §Consequences follow-up "When NVIDIA publishes cuda-toolkit-13-3: update the pin" is closed by this ADR