Community
Reddit user reports 1.9x BF16 tensor speedup on RTX 3090 after gpu-unlock
A community test using the gpu-unlock module raised 3090 BF16 tensor throughput from 82 to 155 TFLOPS, but real Qwen prefill gained only about 10 to 12 percent and outputs were unchanged.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead