Measured work, with the scope beside every number.

Kernel engineering, AMD Instinct MI300X

86.3898 to 0.1015 ms

Block-local counts replaced global atomics on a skewed histogram of 67,108,864 keys into 4,096 bins, ROCm 7.2.4.

TorchRec fusedTBE, Blackwell MIG slice

19.74 to 15.93 ms

Forward and backward median on an RTX Pro 6000 Blackwell MIG 2g.48gb slice, bit-exact across three SGD steps.

Copy and compute overlap, Tesla T4

2.518 to 0.510 ms

Exposed copy time per step. The copy moved under the matmul on a side CUDA stream, 9 checks within tolerance, signed record.

Triton fused softmax wrapper, Tesla T4

48 of 48 cases

A host wrapper fault found and corrected. Triton softmax about 1.35x faster than PyTorch in float16 GPU graph timing.

Numerical correctness, Kernelwitness

1,091 to 4 wrong bins

A one percent tolerance accepted 1,091 wrong bins. The measured 432 ULP rule accepted four in the same example.

Google TPU, Pallas

TPU Kernel Parity Lab

A Pallas masked-softmax kernel compared with JAX and XLA under stated test conditions.