$2100 'Troll Rig' runs full DeepSeek R1 671B Q2_K at 7.5 t/s

gpu4.pics
via 1119745302
NVIDIA RTX 3080
1× Modded RTX 3080 20GB · 20 GB total VRAM · AMD EPYC 7763 (QS)
10 GB VRAM NVIDIA 320 W TDP
running DeepSeek R1 671B· Q2_K
7.5
tokens / sec
Generation
7.5tok/s
higher is better
Prompt / prefill
tok/s
higher is better
Time to first token
lower is better
Composite score
Model size
671B
Runtime
AMD EPYC 7763 (QS) · 320 GB RAM
gpu4.pics/r/curated-1j1p9an

All benchmarks on this rig

ModelQuantGenPromptTTFTPowerPerf/W$/MtokScore
DeepSeek R1 671BQ2_K7.5

How it stacks up — DeepSeek R1 671B

NVIDIA RTX 3080
7.5
Apple M3 Ultra
18