2 entries · Checked 2026-10-10
Leaderboard
| Rig | GPUs | Model | Engine | pp512 t/s | tg128 t/s | Wall W | Date |
|---|---|---|---|---|---|---|---|
| An RTX 3090 on a 32-core Threadripper | NVIDIA RTX 3090 24 GB | Qwen3-30B-A3B Qwen3-30B-A3B-Q4_K_M.gguf | llama.cpp b10002 (a7312ae) | 4232 | 207.6 | – | 2026-10-10 |
| A Radeon AI PRO R9700 next to a 3060 | AMD Radeon AI PRO R9700 32 GB | Qwen3-4B-Instruct-2507 Qwen3-4B-Instruct-2507-Q4_K_M.gguf | llama.cpp b11235 (6c7a87f) | 6478 | 139.6 | – | 2026-10-10 |
How runs are measured
llama-bench from llama.cpp produces every result:
llama-bench -m model.gguf -ngl 99 -fa 1 -p 512 -n 128
- pp512: prompt processing speed on 512 tokens, in tokens/s
- tg128: generation speed over 128 tokens, in tokens/s. This is the number people feel.
Each run lists the exact GGUF file, the llama.cpp build, the power limit and, when measured, the wall power in W.
Send your run
Email the full llama-bench output and the command line to contact@needforvram.com. You can also add it to your Garage page.
See LocalScore and LocalMaxxing for large community datasets.
Questions
- How are the speed numbers measured?
- With
llama-benchfrom llama.cpp, on the same command line for every run, so cards, quants and builds can be compared. Each entry lists the exact GGUF file, the build and the power limit. - What is the difference between pp512 and tg128?
pp512is prompt processing on 512 tokens, in tokens per second;tg128is generation over 128 tokens. Generation is the number people feel at the keyboard.- How do I submit a run?
- Paste the full llama-bench output and the command line to contact@needforvram.com, or put it on your own Garage page. Runs without the command line are not comparable and are not listed.