New Nvidia unlock works on Chinese 3080 20GB too
retained text
Just ran my own tests of the new GPU performance unlock [by Zaggita](https://github.com/Zagitta/gpu-unlock). If you missed it: Nvidia locked down compute speed on consumer cards for some specific hardware instructions needed for AI; the unlock fixes it. TLDR: got +20% prefill impovement, 0% decode change after unlock.
Details:
Test bench: my two old [2x RTX3080 20GB](https://www.reddit.com/r/LocalLLaMA/comments/1p0bbrl/rtx_3080_20gb_a_comprehensive_review_of_chinese/), this time paired with Threadripper 1920x, both running PCIe 3.0 x16. Both GPUs power limited to 240W.
System: Proxmox 9.2.3, linux kernel 6.14.11-9-pve, proprietary Nvidia drivers 610.57.04, CUDA 13.3.
Inference: vLLM 0.32.0 running inside Debian 13 LXC container, tensor parallel mode.
Model: [cyankiwi/Swift-1.5-Qwen3.8-27B-AWQ-INT4](https://huggingface.co/cyankiwi/Swift-1.5-Qwen3.8-27B-AWQ-INT4), fp8 kv cache, MTP enabled.
Full vLLM launch config:
vllm serve cyankiwi/Swift-1.5-Qwen3.8-27B-AWQ-INT4 --max-num-seqs 4 --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --tensor-parallel-size 2 --enable-prefix-caching --kv-cache-memory-bytes 6G --kv-cache-dtype fp8_e4m3 --enable-prompt-tokens-details --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":3}' --enable-chunked-prefill --attention-backend FLASHINFER --default-chat-template-kwargs '{"reasoning_effort": "medium"}'
Benchmark results for single stream:
|Prompt length tok|TTFT before|TTFT after |TTFT change|ITL mean before|ITL mean after|Prefill tok/s before|Prefill tok/s after|Prefill speed change|Decode steps/s before|Decode steps/s after|
|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|
|8 192|8 471 ms|7 013 ms|**-17.2%**|22.27 ms|29.63 ms|967|1 168|**+20.8%**|16.05|15.98|
|16 384|17 573 ms|14 418 ms|**-18.0%**|23.93 ms|22.65 ms|932|1 136|**+21.9%**|16.26|15.88|
|32 768|36 754 ms|29 970 ms|**-18.5%**|23.89 ms|27.20 ms|892|1 093|**+22.6%**|16.54|16.36|
|65 536|79 432 ms|64 173 ms|**-19.2%**|27.86 ms|24.57 ms|825|1 021|**+23.8%**|15.98|16.50|
|131 072|183 593 ms|146 010 ms|**-20.5%**|22.42 ms|25.48 ms|714|898|**+25.7%**|16.36|16.73|
|261 900|465 890 ms|362 192 ms|**-22.3%**|24.18 ms|23.00 ms|562|723|**+28.6%**|15.42|16.67|
TTFT: Time To First Token. Prompt processing speed is defined as prompt\_length / TTFT. Each prompt in the benchmark was randomly generated, so there was no prefix cache hits during test.
ITL: Inter-Token Latency: amount of time passed between two streaming responses from the engine. Decode steps/s are defines as 1s/ITL.
Why I report ITL and not token generation metric: due to activated MTP, the engine will respond with multiple tokens at once; but the amount of tokens in single responce will vary, due to acceptance rate. To isolate this varuable, I measure decode step rate. If MTP acceptange length is 0, each step generates 1 token; if MTP has perfect acceptance length of 3, then each step generates 4 tokens. So this means theoretical generation performance between 15 and 60 tok/s. Real-life MTP acceptance rate for agentic coding fluctuates around 2 for this model, so real-life perforamnce fluctuates around 45 tok/s.
During the test, both GPUs hit light thermal throttling and lowered their core clocks to 1750Mhz. Cold GPUs reach 1950MHz. therefore, I could have about +10% more performance if I had better cooling. I also re-run the test with full 320W power limit, got both prefill and decode speeds within 2% of 240W, so for those particular cards, power limit does not impact real life speed, but does impact real life power bill.
Congratulations to fellow Ampere owners, it's a bright day for us today!document fetched 2026-10-24 10:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xdn33h/new_nvidia_unlock_works_on_chinese_3080_20gb_too