GeForce RTX 4060
Status and notes
24 SMs, 8 GB. FP16/FP32 gained 55% on the cold boot and 41% after a warm reboot; TF32 gained the most (+63%). FP8 and the int8 path barely moved, and FP32 without tensor cores went slightly backwards — the unlock only widens the tensor rate.
Measured throughput
| Mode | Stock | Unlocked | Gain |
|---|---|---|---|
cuBLAS f16/f32 4096³ | 29.97 TF/s | 46.49 TF/s | 1.55× |
cuBLAS f16/f32 8192³ | 31.38 TF/s | 44.23 TF/s | 1.41× |
cuBLAS bf16/f32 4096³ | 31.46 TF/s | 49.57 TF/s | 1.58× |
cuBLAS bf16/f32 8192³ | 31.59 TF/s | 45.37 TF/s | 1.44× |
cuBLAS tf32 4096³ | 15.86 TF/s | 25.86 TF/s | 1.63× |
cuBLAS tf32 8192³ | 15.89 TF/s | 25.30 TF/s | 1.59× |
cuBLAS fp8 4096³ | 63.97 TF/s | 96.19 TF/s | 1.50× |
cuBLAS fp8 8192³ | 63.49 TF/s | 100.43 TF/s | 1.58× |
cuBLAS f16/f16 4096³ | 46.75 TF/s | 52.48 TF/s | 1.12× |
cuBLAS f16/f16 8192³ | 49.47 TF/s | 49.19 TF/s | 0.99× |
cuBLAS fp32 4096³ | 7.85 TF/s | 7.60 TF/s | 0.97× |
cuBLAS int8 4096³ | 28.10 TF/s | 28.27 TF/s | 1.01× |