Skip to content

The device model knows two GPUs (CC 8.6, 7.5); Hopper and Blackwell are model errors #52

Description

@vyncint

TodayDEVICES holds two entries: CC 8.6 (A10G) and CC 7.5 (T4) (model/lib.rs:36–58). There is no 8.0 (A100), 8.9 (L4/L40), 9.0 (Hopper: 2048 threads/SM, 64 warps, 32 blocks, 228 KB) or 10.0 (Blackwell). --cc on anything else is a model error, and docs/LIMITATIONS.md correctly says the model is a capacity question — but the capacity table stops at Ampere.

Separately, reconverge's cc.rs carries a 7.0 → 12.0 shared-memory table for the same devices. Two tables of the same facts in two repositories will drift; they already disagree in shape (one has threads/warps/blocks, the other smem per block).

Fix. Complete DEVICES for 8.0, 8.9, 9.0 and 10.0 from the CUDA Programming Guide's compute-capability table, each entry citing the row it came from; then decide with reconverge whether one small shared simt-device-table crate should own both tables. Keep sm_count per named product, since it is a product fact and not a CC fact.

Done when

  • device("9.0") and device("10.0") return parameters with a cited source.
  • A test asserts every CC in the table is ascending and every field is within the CUDA-documented range.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions