TL;DR: A user-tested patch from Zaggita reportedly removes Nvidia's compute throttle on consumer cards, delivering ~20% faster prefill in vLLM with no decode gain.
Summary: A community performance unlock attributed to Zaggita claims to restore hardware instructions Nvidia disabled on consumer GPUs for AI compute workloads. An independent tester ran it on two RTX 3080 20GB cards (Threadripper 1920x, PCIe 3.0, 240W power limit, CUDA 13.3, vLLM 0.32.0 in tensor-parallel mode) serving a Qwen3-27B AWQ INT4 model with FP8 KV cache and MTP speculative decoding, measuring roughly +20% prefill throughput and 0% change in decode speed for single-stream requests.
Why it matters: Prefill is the dominant cost for long-context and RAG workloads, so a free ~20% prompt-processing gain on repurposed mining or modded cards matters for self-hosted inference economics. Test it on your own setup and validate thermals, driver compatibility, and decode-heavy workloads before relying on it, since token generation appears unaffected.
Source: reddit