Rodrigo Liang, SambaNova Systems | AI Infrastructure Silicon Valley - Executive Series
summary
SambaNova Systems claims a native Dataflow chip architecture with TSMC manufacturing that compiles Hugging Face models without an ISA, and ships full-stack racks that run LAMA 400B–class models. Measured wins are specific: 1,000 tokens/sec versus ~100 on typical NVIDIA stacks, and racks at 19 kilowatts vs 120 kW; the system supports multi-tenant deployment, hosting hundreds of fine-tuned models per rack, enabling private on-prem inference and lower power and footprint at production scale.
AI-written summary · about
