Updated:09:24 AM CEST Oct,23
(new)
(c) 1998-2026 Gameguru Mania
Privacy Policy statement
|
Linux Kernel Module Unlocks Full BF16 Speed on RTX 3090s - tech|
| (hx) 08:50 AM CEST - Oct,23 2026 | NVIDIA has shipped RTX 3090 graphics cards with BF16 tensor math running at half speed for six years. The driver reports this as FMLA32 = 1/2, and NVIDIA’s own specifications list BF16 with FP32 accumulation at roughly half the rate of FP16.
A Linux kernel module called gpu-unlock, created by developer Zagitta, flips the setting to full speed before the NVIDIA driver loads. After a cold boot, cards report FMLA32 = full. One user measured the BF16 tensor pipe jumping from 82 to 155 TFLOPS on the same hardware, driver, and power limit.
Real-world gains are smaller but still useful. On a Qwen model in the Strata inference engine, prefill speed improved by about 10–12 percent. Outputs stayed identical to the stock run. Independent tests and Windows UEFI ports of the same method have shown similar roughly 1.8–2× gains on BF16 and FP16 tensor math with FP32 accumulation.
The module does not change clocks, power limits, or permanently alter the card. Community reports say it mainly helps workloads heavy on BF16 matrix multiplies, such as image generation, some speech and music models, and certain training steps. LLM decode speed sees little benefit because it is usually limited by memory bandwidth.
NVIDIA has not commented on the module. The half-speed setting matches official published rates for consumer Ampere cards.
|
|
last 10 comments: All comments
|
|
Add your comment (free registration required)