SambaNova and Intel Demonstrate Air-Cooled Heterogeneous Inference: D-Series GPU for Prefill, SN50 for Decode
At SC26, the companies showcased a disaggregated inference system combining Intel's D-Series GPUs and SambaNova's SN50 RDUs, designed to leverage existing air-cooled data centers for AI scaling.
At SuperComputing 2026, AI chip company SambaNova and Intel jointly demonstrated a new heterogeneous inference architecture. The solution pairs Intel's GPU D-Series (codenamed Crescent Island) for the compute-intensive "prefill" phase with SambaNova's fifth-generation Reconfigurable Dataflow Unit (RDU), the SN50, for the memory-bandwidth-intensive "decode" phase. This combination aims to run each inference stage on the hardware best suited to its characteristics, creating a fully air-cooled "Token Factory."
The Intel D-Series is a 350-watt air-cooled PCIe GPU designed for data center inference. It features 32 Xe3P cores and 256 XMX engines, equipped with 160 GB of LPDDR5X memory (up to 480 GB in partner designs) and 1.5 TB/s of memory bandwidth, making it suitable for processing long-context prompts and building key-value (KV) caches. In contrast, the SambaNova SN50 uses a continuous dataflow architecture that avoids the kernel launch overhead and memory round-trips common in GPUs, optimizing it for token-by-token generation during decoding. A single SN50 rack holds 16 RDUs with a typical power draw of 20 kW, allowing it to fit into many existing air-cooled data centers.
The system's software stack relies on open standards, including NVIDIA Dynamo for request routing, vLLM as the serving engine, and NIXL for KV cache transfer over RoCEv2 Ethernet. This design allows service providers to implement efficient cache-aware routing using existing network infrastructure without learning proprietary stacks. When an agent returns to a conversation or revisits shared context, Dynamo routes the request to the SN50 node already holding that prefix, reducing repeated prefill work and lowering time-to-first-token latency.
The demonstration addresses a critical industry constraint: global data center construction faces delays due to power constraints and community opposition. According to Sightline Climate, as many as half of the data center projects slated to come online in 2026 could face delays. Since much of the existing space is air-cooled while newer AI systems often require direct liquid cooling retrofits, this heterogeneous approach offers a path to scale AI capacity without costly facility upgrades. The system was shown running Zhipu AI's GLM-5.3 model at the event.