Updated:10:34 AM CEST Oct,24
(new)
(c) 1998-2026 Gameguru Mania
Privacy Policy statement
|
RTX 3090 Tensor Cores Nearly Double with Free Linux Unlock - tech|
| (hx) 09:24 AM CEST - Oct,24 2026 | 
A new open-source Linux tool can unlock restricted Tensor Core performance on NVIDIA RTX 3090 and other Ampere cards. Users report BF16 throughput jumping from about 75 TFLOPS to 135 TFLOPS in PyTorch under a 360 W power limit.
The project, called gpu-unlock, is a kernel module that removes a firmware speed limit on FP16 and BF16 operations with FP32 accumulation. NVIDIA deliberately cut the issue rate on consumer cards compared with professional equivalents. The module runs before the standard NVIDIA driver loads, so no permanent VBIOS flash is required.
Real-world LLM gains are smaller than pure math benchmarks. Prompt processing can improve 10–30 percent depending on the engine, while token generation often stays nearly the same. The unlock currently works only on Linux and reverts after a full power-off.
|
|
last 10 comments: All comments
|
|
)
Add your comment (free registration required)