SignalTech/AI
GITHUB - ZAGITTA / GPU - UNLOCK : UNLOCK FP / BF16 INSTRUCTION ISSUE RATE ON NVIDIA 3000 CARDS TO THEIR PROFESSIONAL EQUIVALENTS
Unlock FP/BF16 instruction issue rate on NVIDIA 3000 cards to professional equivalents.
Context
What happened: A Linux kernel module from the gpu-unlock project (Zagitta) unlocks FP/BF16 instruction issue rates on Nvidia GA102, GA103, GA104 and GA106 cards to match their professional equivalents, running before the unmodified NVIDIA driver binds the card.
Key numbers: FP16 87.46 to 118.39 TFLOPS; BF16 87.50 to 157.41 TFLOPS; TF32 42.58 to 82.61 TFLOPS. Speed selectors SS0 0x88888888 and SS1 0x8.
Why it matters: The module avoids manual patching for each new NVIDIA driver release, and it re-applies the unlock after suspend to RAM, which resets the speed selectors.
What changed
A Linux kernel module from the gpu-unlock project (Zagitta) unlocks FP/BF16 instruction issue rates on Nvidia GA102, GA103, GA104 and GA106 cards to match their professional equivalents, running before the unmodified NVIDIA driver binds the card.
Why it matters
The module avoids manual patching for each new NVIDIA driver release, and it re-applies the unlock after suspend to RAM, which resets the speed selectors.