GPUUnlock

FP16 → FP32 tensor rate

FP16 multiply with an FP32 accumulator — the shape most inference and attention kernels use.

How this path is measured

Register-resident mma.sync issue rate: all-one inputs, eight independent accumulator chains per warp, 20,000 iterations and three timed repetitions, best of three reported.

Governed by the SM speed-select fuse: yes

What the unlock changes on this path

11 rows over 11 cards. 11 improved by more than 5%, 0 did not; the largest gain is GeForce RTX 4070 SUPER at 2.00×.

ModeStockUnlockedGainCard
f16/f3276.7153.6 TF/s2.00×GeForce RTX 4070 SUPER · AD104 · 56 SMs
f16/f32107.0213.7 TF/s2.00×GeForce RTX 4080 · AD103 · 76 SMs
f16/f32104.8209.7 TF/s2.00×GeForce RTX 4080 SUPER · AD103 · 80 SMs
f16/f32 · 26.48 TF/s (209.02 ms) -> 52.97 TF/s (104.51 ms)26.553.0 TF/s2.00×GeForce RTX 4050 Laptop GPU · AD107 · 20 SMs
f16/f3218.336.6 TF/s2.00×GeForce RTX 3050 6GB · GA107 · 18 SMs
f16/f32165.6330.8 TF/s2.00×GeForce RTX 4090 · AD102 · 128 SMs
f16/f3284.8169.4 TF/s2.00×GeForce RTX 4070 Ti · AD104 · 60 SMs
f16/f3267.0133.8 TF/s2.00×GeForce RTX 4070 (Gigabyte) · AD104 · 46 SMs
f16/f3266.3132.2 TF/s1.99×GeForce RTX 4070 (AD103) · AD103 · 46 SMs
f16/f3295.0189.0 TF/s1.99×GeForce RTX 4070 Ti SUPER · AD103 · 66 SMs
f16/f3246.492.3 TF/s1.99×GeForce RTX 4060 Ti 16GB · AD106 · 34 SMs

The smallest ratio in this set is GeForce RTX 4060 Ti 16GB at 1.99×.

Reading the numbers

A 1.00× row is not a failed unlock. It says that particular path was never rate-limited, so the override had nothing to restore. If you benchmark your own card and see 1.00× on the mode you care about, the unlock may well have applied correctly and simply does not help that workload.

Other measurement paths