Computelab.co
← Chip & kernel registry
Google / Grouped GEMM and MoE

Segment matmul from jax-md, Pallas Triton to Pallas TPU

Google TPU v6e

Pass · Real chip

Recorded result

Pass, 65 of 65 shapes. Port median 155.4 us against 176.7 us for the JAX reference on the production shape

Recorded 2026-10-01

Execution conditions

  • TPU v6e-1, single chip, JAX 0.11.2, libtpu 0.0.48
  • Every shape 5 of 5 inputs inside the FP32 tolerance
  • Reduced precision on the same shape is caught and fails
  • Signed card, certificate CLV-E85D6-A4580-59D83-082AD

Historical result; a new workload requires compatibility checking and fresh Verify.

Use this as your starting point

Save the reference, describe your workload and set the metric that matters. Your project keeps the selection and its recorded conditions together.

Evaluate for my workload →

A historical result is not a compatibility guarantee. Your configuration needs a fresh Verify run.

Evidence summary imported from the local registry on October 2, 2026. A summary is not a newly issued proof pack.