Zagita kernel module lifts RTX 3090 Ti BF16 dense GEMM
At 450 W, the dense GEMM run is the condition behind the rise.
| RTX 3090 Ti | RTX 3080 10 GB | |
|---|---|---|
| BF16 before | 87.50 TFLOPS | |
| BF16 after | 157.41 TFLOPS | |
| FP16 before | 87.46 TFLOPS | 65 TFLOPS |
| FP16 after | 118.39 TFLOPS | 113 TFLOPS |
| power | 450 W | 320 W |
The Zagita module writes speed selectors before the NVIDIA driver binds. The table puts measured BF16 dense GEMM on an RTX 3090 Ti at 87.50 TFLOPS before and 157.41 TFLOPS after.
FP16 moves from 87.46 to 118.39 TFLOPS in the same run.
The power-limit table shows 300 W before, 380 W after a cold boot, and 350 W after a warm reboot.
The 157.41 TFLOPS figure sits outside those limits, so the reported gain depends on the 450 W dense GEMM condition.
The document does not show sustained-load behavior, thermals, or LLM latency from the dense GEMM medians. The page leaves open the cost of the extra power, behavior after driver updates, and actual LLM throughput.
Treat the 450 W condition as the controlling fact, not the module name.
After @KBlueleaf. We did not report this. The pictures, if any, are theirs.
