gpu4.pics
data via LocalScore
NVIDIA RTX 3090
llamafile 0.9.1
24 GB VRAM NVIDIA 350 W TDP
running Qwen2.5 14B Instruct· Q4_K_M· 14.8B
59.9
tokens / sec
573 · Strong
Generation
59.9tok/s
higher is better
Prompt / prefill
2,005tok/s
higher is better
Time to first token
640 ms
lower is better
Efficiency
0.18 tok/s/W
Electricity cost
$0.259/Mtok
Power draw
329 W
AMD EPYC 7352 24-Core Processor (znver2) · 126 GB RAM · Linux · llamafile
gpu4.pics/r/ls-result-2

All benchmarks on this rig

ModelQuantGenPromptTTFTPowerPerf/W$/MtokScore
Qwen2.5 14B Instruct14.8BQ4_K_M59.92,005640 ms329 W0.18 tok/s/W$0.259/Mtok573

How it stacks up — Qwen2.5 14B Instruct

NVIDIA RTX 3090
59.9
NVIDIA H100
72.2
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
70.8
NVIDIA A100 80GB
66.3
NVIDIA RTX 3090 Ti
64.2
NVIDIA RTX A6000
52