Unofficial guide
GPU Unlock benchmarks
Numbers reported by the upstream projects. Outputs are bit-identical; only speed changes.
BF16 dense matmul, stock power limits
StockUnlockedTFLOPS
* RTX 4090 is the peak MMA rate; at stock power it gains much less in large GEMMs.
| Card | Stock | Unlocked | BF16 gain |
|---|---|---|---|
| RTX 3090 Ti | 88 | 157 | +78% |
| RTX 3090 | 75 | 125 | +67% |
| RTX 3080 | 65 | 113 | +74% |
| RTX 3070 | 40 | 72 | +80% |
| RTX 3060 Ti | 37 | 68 | +84% |
| RTX 3060 | 27 | 51 | +89% |
| RTX 4090 | 165 | 329 | +99% |
| RTX 4080 | 104 | 193 | +86% |
| RTX 4070 Ti | 81 | 150 | +85% |
Source: Zagitta/gpu-unlock
What gets faster
Faster
- LLM prefill (vLLM, PyTorch, FP16/BF16, AWQ/GPTQ)+25–55%
- Time to first token−25–30%
- LoRA and full fine-tuning in BF16+25–50%
- Image generation (SDXL, SD 1.5, DiT in ComfyUI)+30–55%
- Video generation (Wan 2.2, BF16)+15–20%
- FlashAttention / SDPA kernels+50–60%
Unchanged
- llama.cpp / Ollama with quantized GGUF0–15%
- Token generation (memory bound)0%
- INT8, FP8 and INT4 models0%
- Gaming and FP32 CUDA0%
Ada matched runs
| TFLOP/s | Stock | Unlocked | Source |
|---|---|---|---|
| RTX 4090 · FP16→FP32 instruction rate | 160.4 | 320.4 | JustAResearcher/gpu-unlock-ada |
| RTX 4090 · cuBLAS FP16 GEMM | — | +12–34% | JustAResearcher/gpu-unlock-ada |
| RTX 4070 Ti SUPER · FP16→FP32 (240 W) | 95.0 | 189.9 | JustAResearcher/gpu-unlock-ada |
| RTX 4070 Ti SUPER · TF32→FP32 | 47.49 | 94.51 | JustAResearcher/gpu-unlock-ada |
| RTX 2080 Ti · cuBLAS FP16 8192³ (250 W) | 52.8 | 77.9 | nlzy/turing-unlock |
How to read these
01
Instruction rates show the cap is gone. Application speed depends on power headroom.
02
Memory-bound work, such as token generation, does not change.
03
Most rows are one card and one boot. Post your own in the upstream issues.