Skip to content

ADR 0003: NVIDIA Multi-Generational CUDA & Driver Segmentation

Status

Accepted

Context

NVIDIA driver releases regularly deprecate older GPU architectures (e.g. CUDA 12.8 drops Pascal sm_60/61). Conversely, newer Blackwell and Hopper GPUs require cutting-edge 615/610 drivers, Open Kernel Modules, and Fabric Manager for NVLink, whereas Turing and Ampere remain most stable on the 565 branch. A single monolithic NVIDIA image causes driver breakage on older hardware and missing features on cutting-edge chips.

Decision

We partition NVIDIA support into dedicated provisioners and flavor tiers:

  1. *-nvidia-legacy: NVIDIA 535 LTS branch (535.309.01, CUDA 12.2) for Pascal and Volta.
  2. *-nvidia-mainstream: NVIDIA 565 branch (565.77, CUDA 12.8) for Turing and Ampere.
  3. *-nvidia-modern: NVIDIA 610 branch (610.57.04, CUDA 13.3.1) for Ada Lovelace and Hopper.
  4. *-nvidia-bleeding: NVIDIA 615 branch (615.71.09, CUDA 13.4) for Blackwell (RTX 5090, B200).
  5. *-nvidia-datacenter: NVIDIA Open Kernel Modules + Fabric Manager for Hopper and Blackwell multi-GPU NVLink meshes.

Consequences

  • Pascal and Volta users retain fully functioning CUDA without driver panics.
  • Turing, Ampere, and Ada users run stable production-grade drivers.
  • Blackwell RTX 5090 / B200 users get instant day-zero acceleration with CUDA 13.4 and native CDI.
  • Datacenter operators get turnkey NVLink mesh networking without manual setup.