GPUUnlock

FP8 E4M3 through CUTLASS

A real GEMM kernel rather than an issue-rate loop. It measured 1.0029×, the clearest single piece of evidence that FP8 was never gated.

e4m3/f32 · GFLOP/s · never rate-limited

1rows
0changed
1.00×peak gain

How this mode was measured

CUTLASS 4.8.1 profiler, one cutlass_tensorop_s16832gemm_e4m3_128x128_64x3_tn_align16 kernel, 4096³, 10 warmups, 20 profiling iterations, device reference required

Distance to the specification peak

  • GeForce RTX 4070 Ti SUPER: 59495 GFLOP/s measured against a 176400 GFLOP/s specification peak — 33.7% of it.

330.3 TFLOP/s (Ada whitepaper, RTX 4090, 128 SMs, 2.52 GHz, FP8 dense) × 66/128 SMs × 2.61/2.52 GHz = ~176.4 TFLOP/s

Every card measured on this mode

ModeStockUnlockedGainCard
e4m3/f3259325.80 GFLOP/s59495.00 GFLOP/s1.00×GeForce RTX 4070 Ti SUPER · 66 SMs

Same card, same driver, same binary. The unlocked figure is the only variable.

Which cards this covers

Only cards with a same-card measurement on this mode appear in the table. A card missing from it has not been measured on this path, which is not the same as failing.

Other measured modes