Llama 2 7B on the NVIDIA RTX 4090
Best measured: 190 tokens/s generation and 10,830 tokens/s prompt processing (Q4_0, Vulkan, FA on).
Every run
| Quant | Backend | Generate tok/s | Prompt tok/s | Test | Build / by / notes | Source |
|---|---|---|---|---|---|---|
| Q4_0 | Vulkan FA on | 190 | 10,830 | tg128 | 4ae88d0 coopmat2 | llama.cpp Vulkan scoreboard (#10879) |
| Q4_0 | CUDA FA on | 189 | 14,771 | tg128 | 2241453 @lhl | llama.cpp CUDA scoreboard (#15013) |
| Q4_0 | Vulkan FA off | 188 | 9,452 | tg128 | 4ae88d0 coopmat2 | llama.cpp Vulkan scoreboard (#10879) |
| Q4_0 | CUDA FA off | 186 | 11,993 | tg128 | 2241453 @lhl | llama.cpp CUDA scoreboard (#15013) |
tg128 = generating 128 tokens; pp512 = reading a 512-token prompt (llama-bench). LocalScore averages several prompt and output lengths, so its numbers run lower than tg128.
No card yet?
Try it on a rented GPU (RunPod)Affiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.
All results for the NVIDIA RTX 4090 · Llama 2 7B on other hardware