Which GPU unlock results actually match your card and system? Find the original evidence, including the limits behind the numbers.
A report is not a compatibility verdict. These are third-party self-reports, not our hardware tests. Same model names can hide different boards. This page does not inspect, install or modify anything.
One stock boot and two unlocked boots are reported. The same report found no clear improvement in its quantized Ollama workloads.
Controls, scope & missing evidence
170 W cap. Unlocked GEMM clocks and power were not recorded. FlashAttention compared a desktop boot with a text-console boot; do not treat that result as fully controlled.
Omarchy · kernel 7.2.5-3 · NVIDIA 610.57.04 open gpu-unlock e9fb60a
Same 220 W cap, container and frozen benchmark script; five medians of 30 GEMMs at 8192³. FP32 IEEE stayed near 15.6 TFLOPS.
Controls, scope & missing evidence
Uses a custom early-load hook. One baseline boot and one unlocked boot. Parity check covers a 16×16 output sample; monitor output after reboot was not verified.
350 W cap in both conditions. One IQ3_S MoE prefill test rose from 2,759 to 3,156 tokens/s; decode was reported unchanged.
Controls, scope & missing evidence
GPU clocks differed: BF16 1710 → 1425 MHz. An LLM server remained resident during both GEMM runs. Per-clock gains are not absolute speed gains; do not generalize the prefill result to all models.
The author reports the unlock completing but no double-rate BF16 behavior on this specific OEM board.
Controls, scope & missing evidence
No clean locked/unlocked table is supplied. These two numbers compare precisions after unlock, not before and after. They do not establish exactly 0% uplift or prove a fuse limitation.
CachyOS · kernel 7.3.0-4 · NVIDIA 615.78.08 open Original issue and updates
Both power-limit measurements already have unlock enabled. A controlled locked baseline is absent.
Controls, scope & missing evidence
PCI vendor/device ID and original module commit were not provided. The informal LLM observation has no controlled A/B. Do not compute an unlock gain from the two power caps.
One documented system; 4096-square CUDA/cuBLAS 12.6 benchmark, three-run median. The current headless boot lacks a contemporary controlled locked baseline.
Controls, scope & missing evidence
Experimental source-only v0.3.1. Another 3090 or BIOS is not thereby verified. No gaming, ComfyUI, repeated-cold-boot or longevity guarantee. Linux results do not transfer to this Windows port.
Windows 11 Home 26200 · NVIDIA 617.14 · Ryzen 7 7800X3D Windows GPU UNLOCK README 9352abd
At 1.8× throughput, the same amount of fully accelerated work takes 1 ÷ 1.8 of the time: 44.4% less time. If only part of your task benefits, the whole-job saving is smaller.
GEMM is a dense matrix microbenchmark. Loading models, CPU work, memory transfers, quantized decode and other operations may behave differently. A BF16 result does not establish a gaming or ComfyUI speedup.
Explore a hypothetical workload below. The accelerated fraction is an assumption you choose, not a measurement of your GPU.
What-if time calculator
Time fraction = (1 − 0.5) + 0.5 ÷ 1.8 = 0.778
Fixed-work scenario. No prediction about a specific card or application.
Three checks before trusting a result
Match the identity. Check VRAM, board, PCI vendor/device ID and OS. A Linux report does not establish Windows behavior.
Find the locked baseline. Two unlocked power limits do not measure the benefit of unlocking. Check clocks, power caps and benchmark settings too.
Match the actual workload. A matrix test, LLM prefill and token-by-token decode answer different questions. A missing test is a missing test.