⬢github C · 3.5K ★ +2.6K since we first saw it · pushed 7 h ago · GPL-2.0
JustAResearcher/gpu-unlock-ada
GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER; firmware generated locally
A Linux kernel module that applies a register/firmware override to unlock FP16 and BF16 tensor throughput on RTX 40-series GPUs (demonstrated on RTX 4090 and RTX 4070 Ti SUPER), roughly doubling instruction-issue tensor rates while stock NVIDIA drivers and CUDA keep working. Firmware is generated locally; no binaries are distributed, and real-world GEMM gains vary.
Why now: Trending as a newly-starred repo with independently reproduced ~2× tensor-rate measurements on consumer Ada GPUs, plus an experimental Windows port report.
Who it is for: Linux GPU enthusiasts and ML researchers with RTX 4090/4070 Ti SUPER cards who want more tensor throughput for compute workloads.
Stars over our 132 snapshots: 878 to 3.5K, since 1 d ago.
Where people talked about it
- ⬢github new repos, most starred 7 min ago
- ⬢github new repos, most starred 11 h ago
API: https://socialmediatrends-api.osmike.com/v1/repos/JustAResearcher/gpu-unlock-ada