GPUUnlock

FP16 with an FP32 accumulator

The shape most inference kernels use. This is the path the SM speed-select fuse rate-limits on Ada, which is why an unlock doubles it.

f16/f32 · TF/s · rate-limited by the fuse

2rows
2changed
2.00×peak gain

How this mode was measured

Register-resident mma.m16n16k16 issue rate, 8 independent accumulator chains per warp, best of three timed runs

Distance to the specification peak

  • GeForce RTX 4090: 330.8 TFLOP/s measured against a 330.3 TFLOP/s specification peak — 100.2% of it.
  • GeForce RTX 4070 Ti SUPER: 188.97 TFLOP/s measured against a 190.6 TFLOP/s specification peak — 99.1% of it.

330.3 TFLOP/s (Ada whitepaper, RTX 4090, 128 SMs, 2.52 GHz, FP16 dense) = 1024 FLOP/cycle/SM

1024 FLOP/cycle/SM (330.3 TFLOP/s ÷ 128 SMs ÷ 2.52 GHz) × 66 SMs × 2.82 GHz = 190.6 TFLOP/s

Every card measured on this mode

ModeStockUnlockedGainCard
f16/f32165.60 TFLOP/s330.80 TFLOP/s2.00×GeForce RTX 4090 · 128 SMs
f16/f3294.96 TFLOP/s188.97 TFLOP/s1.99×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