SambaNova and Intel Demo Disaggregated Air-Cooled Inference on GLM-5.3
Summary
SambaNova and Intel demonstrate a disaggregated inference system at SuperComputing 2026 that runs Intel's 350 W air-cooled D-Series GPUs for prefill and SambaNova's SN50 RDUs for decode on GLM-5.3. The demo uses four D-Series GPUs; a SambaRack SN50 holds 16 RDUs at about 20 kW. Both companies say it fits existing air-cooled data centers.
Key Points
- Intel's D-Series card carries 160 GB of LPDDR5X and 1.5 TB/s of memory bandwidth for the prefill role, SambaNova says.
- The demo routes requests with NVIDIA Dynamo and moves KV cache between the two pools over standard RoCEv2 Ethernet via NIXL.
- Sightline Climate reports up to half of data center projects due online in 2026 could face delays from power constraints.