TF32 with an FP32 accumulator

19-bit TF32 inputs, aimed at training throughput. Also fuse-gated on Ada, also roughly doubled by the unlock.

tf32 · TFLOP/s · rate-limited by the fuse

1rows

1changed

1.99×peak gain

How this mode was measured

Register-resident MMA issue rate at a 240 W cap, 20,000 iterations, three timed reps, all 270,336 output cells checked per state

Every card measured on this mode

Mode	Stock	Unlocked	Gain	Card
`tf32`	47.49 TFLOP/s	94.51 TFLOP/s	1.99×	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

FP16 with an FP16 accumulator · 1
FP16 with an FP32 accumulator · 2
BF16 with an FP32 accumulator · 1
INT8 with an INT32 accumulator · 1
INT4 with an INT32 accumulator · 1
FP8 E4M3 through CUTLASS · 1
The dp4a integer dot product · 1
Scalar FP32 fused multiply-add · 1
FP16/FP32 as a cuBLAS GEMM · 4