Performance

Real-world benchmarks.

Measured on production hardware. No synthetic scores. No cloud acceleration.

Consumer-grade GPU cluster · Up to 4× RTX 4090

Model Inference — Tokens / Second

Llama 3 8B (Q4)142 tok/s

Llama 3 70B (Q4)38 tok/s

Mistral 7B (Q4)161 tok/s

CodeLlama 13B (Q4)89 tok/s

Phi-3 Mini (Q4)184 tok/s

Measured with Ollama · Q4_K_M quantization · single concurrent request

Power Efficiency

85W

Idle Draw

Full system at rest

320W

Inference Load

Single model active

580W

Peak Load

All GPUs saturated

0.44

Tokens / Watt

Llama 3 8B Q4

Measured at wall with a kill-a-watt meter · ambient 22°C

Thermal Performance

GPU Idle Temp38°C

GPU Load Temp71°C

CPU Temp52°C

Throttle Threshold95°C

Ambient 22°C · sustained 30-min inference load · custom fan curve applied

Benchmarks represent real-world measurements on reference configurations. Actual results may vary by workload, quantization, and system configuration.