SambaNova Pairs Intel GPUs with SN50 for Air-Cooled AI Inference

What happened
On November 15, 2026, SambaNova Systems and Intel demonstrated a new air-cooled AI inference system at SuperComputing 2026. The system uses Intel D-Series GPUs to process prompts and SambaNova's SN50 hardware to generate output tokens. The design splits AI inference into two phases, with each phase handled by different hardware optimized for its specific workload. The demonstration was intended to show how this setup can be deployed in existing data centers without requiring liquid cooling.
Key details
- Demonstration took place at SuperComputing 2026 on November 15, 2026.
- Intel D-Series GPUs (D270) handle the prefill phase, each with 32 Xe3P cores and 350 W air-cooled design.
- SambaRack SN50 contains 16 Reconfigurable Dataflow Units (RDUs) and typically uses 20 kW per rack.
- The system uses open software including NVIDIA Dynamo, vLLM, and NIXL for context caching and routing over standard Ethernet.
- The design allows independent scaling of prompt processing (prefill) and token generation (decode) hardware.
Why it matters
Data center operators can add AI inference capacity using air-cooled hardware, avoiding the need for expensive liquid-cooling retrofits. This setup allows organizations to scale prompt processing and output generation separately, which can help manage power and cooling constraints in existing facilities. Providers can deploy this system using standard server racks and Ethernet networking, potentially reducing infrastructure costs and complexity.
