Primary sourceLxfater
View sourceEvent ·Multi-source
3090 被解锁 FP16 限制,PyTorch 算力从 75 提到 135 TFLOPS
有网友利用 GPU 启动固件的漏洞,在 NVIDIA 驱动加载前拿到寄存器写入权限,解除了 RTX 3090 上 FP16/BF16 输入、FP32 累加运算的速度限制。功耗仍限制在 360W 的情况下,PyTorch 测试成绩从 75 TFLOPS 提升到 135 TFLOPS,提升约 80%。4090 也有分支项目将同类运算的指令吞吐从约 165 TFLOPS 提升到 331 TFLOPS,接近翻倍。这一改动不是超频,而是绕过了驱动层的算力权限保护。
Current read
有人钻了显卡固件的漏洞,3090 不超功耗直接多出 80% 算力,跑 AI 的老卡党快看。
7 sources72° AI heatLast updated 2026/10/24 06:35:08
Evidence
7 sourcesRelated source 1Teortaxes
View sourceRelated source 2kimmonismus
View sourceRelated source 3AI21 Labs
View sourceRelated source 4rohanpaul_ai
View sourceRelated source 5VTV Công nghệ
View sourceRelated source 6MiniMax_AI
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.