[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f18370askaz4w7":3},{"code":4,"message":5,"data":6,"timestamp":31},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":21,"summary_md":22,"topic":23,"github_topics":26,"license":27,"homepage":28,"updated_at":29,"history":30},"JustAResearcher\u002Fgpu-unlock-ada",4,"GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER; firmware generated locally","C",1271,147,904,1,[16],{"id":17,"title":18,"points":19,"comments":14,"posted_at":20},50223787,"GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER",3,"2026-10-24","Lets Linux users unlock full tensor-core throughput on consumer RTX 40-series cards their firmware artificially limits.","## The problem\n\nNVIDIA's consumer Ada cards ship with tensor cores running at half their rated speed-select setting, so the same silicon in an RTX 4090 delivers roughly half the FP16\u002FBF16 tensor throughput a workstation part achieves. The unlock has to happen between firmware bringing the card up and the stock NVIDIA driver binding it, which no userspace tool or driver patch can reach. This project is a Linux kernel module plus locally generated firmware payload that performs that override before driver handoff, then verifies the change survived a PCIe FLR.\n\n## What's different\n\nThe usual routes are a patched NVIDIA driver (fragile per driver version, needs recompiling) or a Windows UEFI utility. This one sits below the driver, so the stock `615.78.08` \u002F `580.178.04` drivers load unmodified and the applied state survives reset. It generates the firmware from pinned NVIDIA inputs fetched locally rather than distributing binaries. And it is explicit about its limits: measurements are register-resident instruction rates, not application speedups — the README itself reports cuBLAS gains of 12–34% for FP16 but FP8 GEMM essentially flat at 0.3%, and INT8\u002FINT4 unchanged.\n\n## Who it's for\n\nOwners of the specific listed cards (RTX 4090 `10de:2684` is the only non-experimental target; 4070 Ti SUPER `10de:2705` and 4060 Ti 16GB `10de:2805` are single-card experiments) running Linux from the initramfs, who benchmark or run FP16\u002FBF16 tensor workloads and can tolerate boot-level risk. Look elsewhere if you are on Windows, run A-series or data-centre parts, need INT8\u002FFP8 gains, or want a supported, turnkey product — the README never claims that.\n\n## Cost of adoption\n\nYou need Linux kernel headers, a DKMS toolchain, and one of the supported distro families (Arch, Debian, RPM-based); `make install` dispatches to makepkg\u002Fdpkg\u002Frpmbuild. Builds fetch and SHA-256-verify pinned NVIDIA `610.43.03` inputs into `~\u002F.cache\u002Fgpu-unlock` by default. The unlock only activates via kernel command line and requires a full power-off for the first attempt; a warm reboot may leave firmware in a state the module refuses. It is GPL-2.0-only and a day-old fork of `Zagitta\u002Fgpu-unlock` integrating a payload from a third-party Hugging Face release — evidence so far is a handful of owner runs, not an extended stability test.\n\n## Verdict\n\nThis is a brand-new repository: created 2026-10-23, tracked for one day, and already at +668 stars\u002Fday onto a new board at #6 — a spike, not a track record. The underlying achievement is real and narrowly verified on one card per target, but the adoption risk profile is boot-time kernel intervention against undocumented GPU firmware, and the performance story is much narrower than the doubling headline suggests. Worth watching and worth reading for anyone doing Ada tensor benchmarking on Linux; not something to deploy on a machine you depend on until reports accumulate across boards and driver versions.",{"slug":24,"name":25},"hardware","Hardware & IoT",[],"GPL-2.0",null,"2026-10-25T01:05:12Z",[],1792915142]