tensor_modes_issue.cu measures register-resident FP16/FP32, TF32/FP32, and
unsigned INT4/INT32 mma.sync instruction issue rates. Each warp keeps eight
independent accumulator chains in flight. tensor_modes_driver_runner.py
times 20,000 iterations three times, then checks all 270,336 output cells
per mode against the expected all-ones result. It requires the exact GPU UUID,
power cap, and cubin SHA-256 so a different card or binary cannot silently
enter a matched comparison.
Build on Linux with the CUDA toolkit for an SM89 card:
nvcc -ptx -arch=sm_89 -O3 -std=c++17 \
-o tensor_modes_issue_sm89.ptx diagnostics/bench/tensor_modes_issue.cu
ptxas -arch=sm_89 -O3 -o tensor_modes_issue_sm89.cubin \
tensor_modes_issue_sm89.ptx
sha256sum tensor_modes_issue_sm89.cubin
python3 diagnostics/bench/tensor_modes_driver_runner.py \
tensor_modes_issue_sm89.cubin \
--sha256 PASTE-THE-CUBIN-SHA256 \
--expect-uuid GPU-PASTE-YOUR-EXACT-UUID \
--expect-power-limit 240Set --expect-power-limit to the cap actually used on your card. Run only
when the GPU is free of customer and owner workloads. For a stock/unlocked
comparison, use the same card, cubin, power cap, driver, launch settings,
and clock policy in both states. Record live clocks and restore the original
workload afterward.
This is an instruction-throughput probe, not a GEMM or inference benchmark. All-one inputs check the exercised arithmetic path but do not validate general numeric accuracy or predict application speedup. The RTX 4070 Ti SUPER report records one matched 240 W comparison and its limits.
validate_gemm.cu is a separate correctness check using cuBLAS and a CPU
reference. It uses three seeded FP16/BF16 input sets with square, rectangular
and non-tile-aligned shapes. Inputs vary across both coordinates and use exact
dyadic values. Run the same built binary on the same GPU in stock and unlocked
states; these small checks do not establish general application accuracy.
nvcc -O3 -arch=sm_89 -std=c++17 -o validate_gemm \
diagnostics/bench/validate_gemm.cu -lcublasmma_bench.cu accepts positive iteration and repetition counts. Repetition
results are stored dynamically, so longer runs do not use a fixed text buffer.
The testing guide gives a stock/unlocked collection workflow, including the separate large cuBLAS GEMM probe. The matched RTX 4090 evidence records both improved and unchanged cases, numerical checks and telemetry.