FP16 → FP16 tensor rate
FP16 multiply with an FP16 accumulator, the path that already runs at full tensor rate.
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.
What the unlock changes on this path
8 rows over 8 cards. 0 improved by more than 5%, 8 did not; the largest gain is GeForce RTX 4090 at 1.02×.
| Mode | Stock | Unlocked | Gain | Card |
|---|---|---|---|---|
f16/f16 | 325.7 | 331.6 TF/s | 1.02× | GeForce RTX 4090 |
f16/f16 | 169.4 | 170.0 TF/s | 1.00× | GeForce RTX 4070 Ti |
f16/f16 | 210.1 | 210.4 TF/s | 1.00× | GeForce RTX 4080 SUPER |
f16/f16 | 153.4 | 153.6 TF/s | 1.00× | GeForce RTX 4070 SUPER |
f16/f16 | 133.9 | 133.9 TF/s | 1.00× | GeForce RTX 4070 (Gigabyte) |
f16/f16 | 36.6 | 36.6 TF/s | 1.00× | GeForce RTX 3050 6GB |
f16/f16 | 214.2 | 213.9 TF/s | 1.00× | GeForce RTX 4080 |
f16/f16 | 132.6 | 132.2 TF/s | 1.00× | GeForce RTX 4070 (AD103) |
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.