Llama-3.3-70B-Instruct

on 2x RTX 3090

19.1

tok/sdecode

2x RTX 3090 (24GB) + AMD EPYC 7H12 64-Core Processor (128c) + 1008GB

rank in tier

1/12x RTX 3090 runs

best workload

chat-short

where the rig flew

slowest workload

chat-long18.7 tok/s

where the rig struggled

backend

llama.cppllama.cpp is the universal backend — broad model support, modest speed ceiling.

faster than

llama-3.3-70b-Instruct-4bit on M3 Ultra16.8 tok/s
qwen2.5-72b-Instruct-4bit on M3 Ultra16.3 tok/s
phi-4-4bit on M3 Pro15.6 tok/s

This is a shareable view of one signed llm-speed suite-suite-v1 run. Numbers are not edited. The raw record has every workload, the public key, and the fingerprint hash.