中文版

gpu-unlock（GitHub：https://github.com/Zagitta/gpu-unlock ，2026-10-18 创建，已 11,363★）是一个 Linux 内核模块：它在启动时、原厂 NVIDIA 驱动加载之前，解除 GeForce 显卡上被限制为一半速度的张量核心（FP16 / BF16 / TF32），不用打补丁的驱动、不用刷 VBIOS，驱动升级后也不用重做。

按项目给出的实测：张量算力最高可翻倍，LLM 预填充（prefill）快 25–55%，BF16 微调快 25–50%，图片生成快 30–55%。支持 RTX 20 / 30 / 40 系列，RTX 50 系列不支持；仅限 Linux，且需要手动在内核命令行里选择开启（gpu_unlock.targets=all），属于启动层面的改动。

为什么值得了解

• 本地跑模型的人最缺的往往不是新显卡，而是把手里这张卡的算力用满——这个项目瞄准的正是这个缺口
• 10 天 1.1 万 star，说明”消费级显卡被人为限速”戳中了很多人的痛点
• 注意边界：它只加速张量计算，游戏、解码、INT8/FP8 模型基本不受益；动手前先确认自己的卡在支持列表里

English Version

gpu-unlock (GitHub: https://github.com/Zagitta/gpu-unlock , created 2026-10-18, now 11,363★) is a Linux kernel module: at boot, before the stock NVIDIA driver loads, it removes the half-speed cap on GeForce tensor cores (FP16 / BF16 / TF32) — no patched driver, no VBIOS flash, and nothing to redo when NVIDIA ships a new driver.

Per the project’s own benchmarks: tensor throughput can double, LLM prefill runs 25–55% faster, BF16 fine-tuning 25–50% faster, and image generation 30–55% faster. It supports RTX 20 / 30 / 40 series cards (not RTX 50), is Linux-only, and must be opted into via a kernel command-line flag (gpu_unlock.targets=all) — a boot-level change.

Why it’s worth knowing

• People running models locally often don’t need a new card — they need the card they already own to run at full speed, and that’s exactly the gap this targets
• 11k stars in 10 days shows that “consumer GPUs being artificially speed-limited” hits a nerve
• Know the boundaries: it only accelerates tensor math — gaming, decoding, and INT8/FP8 models barely benefit; check that your card is on the supported list first