{"full_name":"JustAResearcher/gpu-unlock-ada","owner":"JustAResearcher","name":"gpu-unlock-ada","url":"https://github.com/JustAResearcher/gpu-unlock-ada","description":"GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER; firmware generated locally","homepage":null,"topics":[],"language":"C","license":"GPL-2.0","avatar":"https://avatars.githubusercontent.com/u/62031663?v=4","created_at":"2026-10-23T21:00:53Z","pushed_at":"2026-10-26T04:24:46Z","metrics":{"stars":4860,"forks":631,"watchers":30,"issues":89,"prs":54,"releases":2,"issues_open":null,"issues_closed":null,"prs_open":null,"prs_closed":null,"prs_merged":null,"discussions":0},"growth":{"day":{"stars":8,"forks":4,"base_day":"2026-11-09"},"week":{"stars":159,"forks":31,"base_day":"2026-11-03"}},"history":{"dates":["2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-09","2026-11-10"],"stars":[2546,3730,4221,4399,4499,4573,4701,4736,4852,4860],"forks":[317,492,551,565,575,587,600,604,627,631],"watchers":[14,19,23,24,26,26,26,27,30,30],"issues_open":[null,null,null,null,null,null,null,null,null,null],"prs_open":[null,null,null,null,null,null,null,null,null,null]},"boards":{"monthly":198,"weekly_rate":125,"monthly_rate":38,"new":142},"same_owner":[],"issues":[{"number":109,"title":"Idea: Use the unused 0x009A0148 slot in generate_v67.py to open FUSE_DEBUGCTRL_PLM (0x00820054) for a read-only RIR fuse dump pre-FLR","comments":8,"created_at":"2026-10-28T16:13:46Z"},{"number":117,"title":"Windows Hybrid success: MSI GE78HX RTX 4090 Laptop (2717/13cd), FP16 +85%, Intel display and brightness preserved","comments":4,"created_at":"2026-10-29T18:10:10Z"},{"number":105,"title":"Windows UEFI report: MSI RTX 4060 Ti 16GB (2803/5174), transaction + direct Windows handoff; benchmark pending","comments":4,"created_at":"2026-10-28T05:24:07Z"},{"number":95,"title":"[Windows UEFI Support] MSI RTX 4090 Suprim Liquid X (AD102, 1462:5104)","comments":3,"created_at":"2026-10-27T17:28:13Z"},{"number":92,"title":"[2705] Windows UEFI unlock working on RTX 4070 Ti SUPER — different VBIOS + ComfyUI A/B +5.3%","comments":2,"created_at":"2026-10-27T12:24:51Z"},{"number":71,"title":"[Silicon Report] Verified unlock on AD107 / RTX 4060 (10de:2882) — Proxmox VE host, LXC consumer","comments":2,"created_at":"2026-10-26T17:13:47Z"},{"number":70,"title":"[Success Report] ASUS RTX 4080 SUPER (AD103, 10DE:2702) — UEFI PLM/SS Unlock Verified, FP16 +97%, BF16 +83%","comments":2,"created_at":"2026-10-26T15:39:53Z"},{"number":140,"title":"Rtx 4060 ti 16gb Windows 11 unlocked\n","comments":1,"created_at":"2026-11-05T23:27:17Z"}],"discussions":[],"latest_release":{"name":"GPU Unlock 1.1.1 — TF32 and INT4 evidence","tag":"v1.1.1","published_at":"2026-10-24T00:28:11Z","body":"GPU Unlock 1.1.1 publishes a reproducible SM89 tensor issue-rate probe and matched stock/unlocked TF32 and unsigned INT4 results on one RTX 4070 Ti SUPER (`10de:2705`, NVIDIA `580.178.04`, 240 W cap). The Ada unlock transaction is unchanged from 1.1.0; the module and DKMS version metadata advance to 1.1.1.\r\n\r\n| Register-resident mode | Stock | Unlocked |\r\n| --- | ---: | ---: |\r\n| FP16 inputs, FP32 accumulate (control) | 94.96 TF/s | 188.97 TF/s |\r\n| TF32 inputs, FP32 accumulate | 47.49 TF/s | 94.51 TF/s |\r\n| Unsigned INT4 inputs, INT32 accumulate | 754.81 TOP/s | 754.68 TOP/s |\r\n\r\nThe same cubin, card, driver, 240 W cap, and launch settings were used in both boots. Each mode checked 270,336 output cells with zero mismatches per state. **TF32 instruction issue rate gained 1.99×; unsigned INT4 did not improve.** These all-one, register-resident loops do not establish an application GEMM or inference speedup. INT8 and one previously tested FP8 GEMM were also unchanged.\r\n\r\nRead the [canary report](https://github.com/JustAResearcher/gpu-unlock-ada/blob/v1.1.1/docs/ada-2705-report.md), [sanitized numeric output and hashes](https://github.com/JustAResearcher/gpu-unlock-ada/tree/v1.1.1/docs/evidence/ada2705), and [GPL benchmark source and reproduction steps](https://github.com/JustAResearcher/gpu-unlock-ada/blob/v1.1.1/diagnostics/bench/README.md). Installation and boot rollback are in the [Quick Start](https://github.com/JustAResearcher/gpu-unlock-ada/blob/v1.1.1/QUICKSTART.md).\r\n\r\nThe exact source-only archive ran 22 Linux host tests with 8 optional-input skips, and its kernel module compiled against the tested host's `5.15.0-177-generic` headers as version 1.1.1. The owner miner stayed running through this CPU-only validation. The downloadable archive contains no NVIDIA firmware binary; pinned NVIDIA inputs are generated locally under NVIDIA's separate license.\r\n\r\n## Additional FP8 evidence — October 23, 2026\r\n\r\nThe earlier unchanged FP8 result was the selected staged-accumulation kernel. A later custom harness using NVIDIA CUTLASS 4.8.1's fast-accumulation GEMM measured **158.54 -> 309.08 dense TFLOP/s (1.949527x)** for 8192x8192x8192 on the same RTX 4070 Ti SUPER, same CUBIN and same 240 W cap. The requested core ceiling was 2400 MHz; unlocked core boundary samples were 2265/2370 MHz versus 2400/2400 MHz stock, with normal 10251 MHz memory in both states. This is a same-power-cap measurement, not clock-normalized throughput.\r\n\r\nAll 19 matched kernel/shape cases passed full-output cuBLAS pedantic FP32 checks before and after timing. The predeclared inputs were independent deterministic +/-1 matrices, K <= 8192. The original staged profiler still gained only about 0.7%; the faster result uses the identical fast kernel in both hardware states. This does not establish arbitrary-input fast-accumulation accuracy, a 2x full-model gain, or Pearl mining performance.\r\n\r\nSeparately, both boots reran the original Ada FP8 numerical-profile suite: **35 suites, 3,227,648 FP32 output cells, zero bitwise differences** between stock, unlocked and the archived Ada results. These rounding-sensitive tests preserve the observed Ada 16+16 model on the tested inputs; Ada still requires its own Pearl verifier/circuit profile.\r\n\r\nEvidence assets:\r\n\r\n- [cutlass-fp8-stock-vs-unlocked.zip](https://github.com/JustAResearcher/gpu-unlock-ada/releases/download/v1.1.1/cutlass-fp8-stock-vs-unlocked.zip) — SHA-256 `b9468068a25cc208749eb0004fc86e15ac98c25c06517fc7c76d3a4f8ac6e861`\r\n- [ada-fp8-profile-stock-unlocked-evidence.zip](https://github.com/JustAResearcher/gpu-unlock-ada/releases/download/v1.1.1/ada-fp8-profile-stock-unlocked-evidence.zip) — SHA-256 `5e9fa282298b996828e9554a47bc32c2da9ac65037fe50368e80f5bb95bac05c`\r\n- [FP8 evidence checksums](https://github.com/JustAResearcher/gpu-unlock-ada/releases/download/v1.1.1/FP8-EVIDENCE-SHA256SUMS.txt)\r\n\r\nThe CUTLASS package includes source, binaries, disassembly, all timings, telemetry, correctness results and the d"},"readme":true,"readme_zh":false,"news":[{"title":"GPU tensor unlock source for RTX 4090 and experimental RTX 4070 Ti SUPER","url":"https://github.com/JustAResearcher/gpu-unlock-ada","hn":"https://news.ycombinator.com/item?id=50223787","source":"Hacker News","lang":"en","date":"2026-10-24","points":3,"comments":1},{"title":"开源评测：gpu-unlock-ada｜RTX4090张量吞吐解锁，三次独立复现实测翻倍","url":"https://blog.csdn.net/TunerT_TQ/article/details/169451263","source":"blog.csdn","lang":"zh","date":"2026-10-29"},{"title":"開源工具幫 RTX 20-40 解鎖 AI 算力 研究者實測最多可以增倍","url":"https://www.msn.com/zh-hk/%E6%8A%80%E8%A1%93/%E7%A1%AC%E9%AB%94%E8%88%87%E8%A3%9D%E7%BD%AE/%E9%96%8B%E6%BA%90%E5%B7%A5%E5%85%B7%E5%B9%AB-rtx-20-40-%E8%A7%A3%E9%8E%96-ai-%E7%AE%97%E5%8A%9B-%E7%A0%94%E7%A9%B6%E8%80%85%E5%AF%A6%E6%B8%AC%E6%9C%80%E5%A4%9A%E5%8F%AF%E4%BB%A5%E5%A2%9E%E5%80%8D/ar-AA2g4Kdd","source":"香港 unwire.hk on MSN","lang":"zh","date":"2026-10-26"},{"title":"GitHub 今日推荐｜gpu-unlock-ada：Ada 显卡张量算力翻倍解锁","url":"https://blog.csdn.net/weixin_40013817/article/details/169049722","source":"blog.csdn","lang":"zh","date":"2026-10-24"},{"title":"UEFI GPU Unlocker Lifts RTX Tensor Limit On Windows","url":"https://news.google.com/rss/articles/CBMiakFVX3lxTE4wYTI4bFYxWnllY2dSRzQ4djV2VXVaMU53b0ZtejRQX1dVWmxTT3FnamVkM1IxOHIxZU9SSVlLUzdrSUVCempsenAyOWJNMkJyLXZud1hyQnB0bUZveEw4NUxKMFp1N0NadGc?oc=5","source":"Modders Inc","lang":"en","date":"2026-11-02"},{"title":"Open-source USB tool unlocks Nvidia's half-speed GPU AI limiter on Windows — lifts GPU's FP16 math performance up to 65%, UEFI unlocker covers RTX 20, 30, and 40-series desktops and laptops","url":"https://news.google.com/rss/articles/CBMi2AJBVV95cUxPVUtXVThBWVBJbnBvZk1QVUZvQ0l3aEVQd1FETFlLdGZocmtJaGp5VmI5N0JnLTlUSWV3cTdoR1FYLXpzcjN1TjBzR0JEU0RubUpUbUlPd21YaHpldFI0dTJxTXM2ak9SRnV0ZVI5UWxFSVhCQUFzcnd5TTYzNlB1amRFdERkYmpFQmQxMlBndFVoV0JIZmRkbnhCUHJPb2l0djNyQ3BOejR6d3ZEYWtURW1ua0c3YzN0N3BTUEJNRzQwTndnVFNjdTRQcWQySXc4SVc1UEFWY0ZxdzJjTV9sTXdKeWluRlBFc0Vydnpldzk4OUh3dVpzTFRnQmdjTE01MWtYRmZLNFc5aVVuQkNudWZjdFNUclBJS3BULWNUVGZMQ2dfYVhKNHpSYUhUN3BJWThlQWIyZmFfUGhtUzRtMHkxbjVvbFRib1paN25JdTZ1amtvdDlYLQ?oc=5","source":"Tom's Hardware","lang":"en","date":"2026-11-01"},{"title":"Modders unlock hidden AI power on Nvidia RTX 20, 30, and 40 series GPUs, nearly doubling some benchmarks","url":"https://news.google.com/rss/articles/CBMilwFBVV95cUxQaDlieVFTNW5uUEtSQ3Fsd3dVOF96OEh4YUZfb19UZjd6YU5WMGoxSC1tSkxMQUJ6QWQzMXdaTmFTOUo2bWRDU09RaUNKbnRzOEtjN3FyOWZnTzM1ZEI2SDZTMlpEX1JLT3l4MGJwdGlCdDhNMkkyLVVNbFhicmNEM29IYlo3YmRncWxMdDlmc1F0dzdkVnlV?oc=5","source":"TechSpot","lang":"en","date":"2026-10-27"},{"title":"GPU Unlock Mods Reach RTX 4090 and RTX 2080 Ti, Doubling FP16/BF16 Tensor Rates","url":"https://news.google.com/rss/articles/CBMia0FVX3lxTE50c2tvQ3dwQXNsaHJCdU0xbTlhMEwwZ3daeGF1TjZvYWowZ2xsMTFXbzlHX0pHRjBXWHFyRmVlTThUbTJBT1NrTzJnc2FkYkI3UzlBUXNFUTlDc3ZENEI3UFJBNW5rdkRPWnlZ?oc=5","source":"Modders Inc","lang":"en","date":"2026-10-27"},{"title":"NVIDIA RTX 30 GPUs Get Huge AI Performance Boost With New Unlock Mod","url":"https://news.google.com/rss/articles/CBMifEFVX3lxTE1EYks3cXV5S1RQcEFXelhWaEhFWUZjVjVjaDhUU1Fwa1REbzRubzU1SHdRUGZmdDBTSnN0VnExczF6Rm1JUDIzUzJoNmhEN2FZSjZQdld0dGFMVVRQbFlrcGJlLU02YmZRSlcwM3VxT2VFRl9oYWFJTVdPRkM?oc=5","source":"ThePCEnthusiast","lang":"en","date":"2026-10-25"},{"title":"英伟达GPU隐藏算力被开源解锁，不改官方驱动，RTX 4090吞吐翻倍","url":"https://news.google.com/rss/articles/CBMiTkFVX3lxTE92b242TzdMcE9vMk9SRWlwVlhEd0pSbU15SmFjcGROYm5PSnNWb3AxWm9Mc0RvaTFFb1RXVHdfQnhVMzJWRzJDSUpZb1Y4UQ?oc=5","source":"36Kr","lang":"zh","date":"2026-10-25"}],"stale":false,"snapshot_at":"2026-11-10T09:35:25Z","history_start":"2026-09-06"}