submission 17892
KernelBot user 105 · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 1 lines, June 9 Researcher Reciprocity License v1.0.
fp8t_v3.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-fp8-mm-17892?include=source"interfacepython
Compatibility
measured onAMD Instinct MI300X
declared hardwareAMD Instinct MI300X
architecturesgfx942
dtypesfp32, fp8_e4m3fnuz
Benchmark evidence
18 measurements across 1 GPU, fastest first.
Operation / workload
Hardware
Latency
Rank
Observed
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=512 · seed=6563
AMD Instinct MI300X
198.1µs
#72 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=1536 · seed=8135
AMD Instinct MI300X
199.3µs
#65 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=512 · seed=12341
AMD Instinct MI300X
205.3µs
#46 of 98
2025-05-08
Show all 18 measurements ›Showing all 18 measurements ⌄
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=576 · seed=12346
AMD Instinct MI300X
228.0µs
#76 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [1024, 7168] · n=4608 · seed=7531
AMD Instinct MI300X
231.8µs
#44 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=576 · seed=9863
AMD Instinct MI300X
262.9µs
#56 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=1536 · seed=6543
AMD Instinct MI300X
420.4µs
#51 of 98
2025-05-08
AMD FP8 blockwise GEMMfp32/fp8_e4m3fnuz · [6144, 7168] · n=4608 · seed=65436
AMD Instinct MI300X
1.14ms
#50 of 98
2025-05-08
Reported · How evidence levels are derived →
Source and license
sourceavailable
revision digestsha256:bf6b6c7b7a752491b361a640ecc4d0b7c69c967705a4716d0bc4db3ca1f5b9f6
license declaredunknown
license concludedunknown
imported2026-08-15
Kernel source
fp8t_v3.py1 lines
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Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON