SambaNova and Intel demonstrate air-cooled AI inference system

SambaNova and Intel are demonstrating an air-cooled inference system that pairs Intel GPU D-Series processors for prefill with SambaNova SN50 RDUs for decode. The companies say the design can help providers deploy disaggregated AI workloads in existing air-cooled data centers.
- Intel D-Series GPUs perform prefill for long-context workloads.
- SambaRack SN50 RDUs perform the decode phase.
- The demonstration runs GLM-5.3 on four D-Series GPUs and an SN50 rack.
- The serving stack uses NVIDIA Dynamo, vLLM and NIXL.
- The companies target existing air-cooled data centers.
A split architecture for inference
The demonstration at SuperComputing 2026 runs GLM-5.3 across both hardware stacks. It uses four Intel D-Series GPUs for prefill and a SambaRack SN50 for decode, showing how the two phases of inference can run on hardware designed for different workloads.
Prefill processes the incoming prompt and builds the key-value, or KV, cache. Decode then generates output tokens one at a time. According to SambaNova, separating these phases allows providers to scale the prefill and decode tiers independently, depending on whether their traffic is dominated by long prompts or long generated responses.
Intel GPUs handle prefill
Intel's D-Series, codenamed Crescent Island, is a 350 W air-cooled PCIe GPU for data center inference. The company describes it as a dedicated prefill engine for long-context workloads, with 32 Xe3P cores, 256 XMX engines, 160 GB of LPDDR5X and 1.5 TB/s of memory bandwidth.
Intel's reference design shown at OCP includes 80 D-Series GPUs, 208 petaflops of peak FP4 compute and between 12 TB and 38 TB of memory per rack, depending on configuration. The rack can provide 120 TB/s of aggregate memory bandwidth. SambaNova says other configurations can adjust GPU count, memory and power for a provider's facility and workload mix.
SambaRack SN50 handles token generation
SambaRack SN50 contains 16 SambaNova fifth-generation Reconfigurable Dataflow Units, or RDUs, and typically requires 20 kW for inference. SambaNova says the RDUs chain model operations into a continuous dataflow across chips, reducing memory traffic and kernel relaunch overhead during token generation.
For larger models, up to 16 SambaRacks can work together as a scalable unit with as many as 256 RDUs. The company says this approach spreads one workload across lower-power racks rather than a single dense, liquid-cooled rack.
The system uses NVIDIA Dynamo for request routing, vLLM for serving on both inference paths and NIXL to move KV cache between the pools over UCX/RDMA on standard RoCEv2 Ethernet. SambaNova says cache-aware routing can send requests to an SN50 that already holds a relevant prefix, reducing repeated prefill work and KV movement.
Why existing data centers matter
SambaNova argues that air-cooled deployment could help inference providers expand without waiting for new facilities or undertaking liquid-cooling retrofits. The company cites Sightline Climate, as reported by Axios in February 2026, saying as many as half of data center projects worldwide slated to come online in 2026 could face delays because of power constraints, grid equipment shortages and community opposition.
For users, the proposed architecture is intended to preserve familiar serving software while allowing providers to add prefill or decode capacity separately. SambaNova and Intel are showing the system at SuperComputing 2026, with SambaNova at booth 2246 and Intel at booth 2819.
SambaNova: The Air-Cooled Token Factory: Intel GPU D-Series + SambaRack SN50

