Llama 2 7B on the NVIDIA A30
Best measured: 132 tokens/s generation and 3,069 tokens/s prompt processing (Q4_0, CUDA, FA on).
Every run
| Quant | Backend | Generate tok/s | Prompt tok/s | Test | Build / by / notes | Source |
|---|---|---|---|---|---|---|
| Q4_0 | CUDA FA on | 132 | 3,069 | tg128 | 583cb83 @Hedede | llama.cpp CUDA scoreboard (#15013) |
| Q4_0 | CUDA FA off | 125 | 2,767 | tg128 | 583cb83 @Hedede | llama.cpp CUDA scoreboard (#15013) |
| Q4_0 | Vulkan FA on | 117 | 3,158 | tg128 | 583cb83 | llama.cpp Vulkan scoreboard (#10879) |
| Q4_0 | Vulkan FA off | 111 | 3,291 | tg128 | 583cb83 | llama.cpp Vulkan scoreboard (#10879) |
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.
Rent one by the hour
Rent a GPU on RunPodAffiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.
All results for the NVIDIA A30 · Llama 2 7B on other hardware