GeForce RTX 4070 Ti
状态与说明
Measured inside a KVM guest with the card passed through by VFIO, so the unlock survived the guest reboot that VFIO forces. FP16/FP32 84.8 → 169.4 TF/s and TF32 42.4 → 84.0 TF/s, same card, binary and 285 W cap. cuBLAS GEMM gained +50% at 1024³, rising to +94% at 8192³.
实测吞吐
| 模式 | 解锁前 | 解锁后 | 提升 |
|---|---|---|---|
f16/f16 | 169.40 TF/s | 170.00 TF/s | 1.00× |
f16/f32 | 84.80 TF/s | 169.40 TF/s | 2.00× |
bf16/f32 | 84.70 TF/s | 168.90 TF/s | 1.99× |
s8/s32 | 338.80 TOP/s | 338.00 TOP/s | 1.00× |
dp4a | 84.10 TOP/s | 84.00 TOP/s | 1.00× |
ffma | 40.10 TF/s | 40.20 TF/s | 1.00× |
tf32 | 42.40 TF/s | 84.00 TF/s | 1.98× |
u4/s32 | 673.90 TOP/s | 672.40 TOP/s | 1.00× |
cuBLAS f16/f32 1024³ | 51.90 TF/s | 77.80 TF/s | 1.50× |
cuBLAS f16/f32 4096³ | 79.80 TF/s | 154.00 TF/s | 1.93× |
cuBLAS f16/f32 8192³ | 83.40 TF/s | 162.10 TF/s | 1.94× |