SambaNova pairs Intel GPUs with its AI chips in an air-cooled inference demo
The SuperComputing 2026 demonstration separates prompt processing from answer generation. SambaNova’s pitch is more AI capacity without a liquid-cooling retrofit, but it supplies no comparative performance results for the pairing.
The demo splits inference by stage: Intel D-Series GPUs process prompts and build the KV cache, which NIXL transfers to SambaNova’s SN50 RDUs for token-by-token decoding. SambaNova says the stack’s Ethernet connection and serving software let providers add capacity in existing air-cooled data centers and scale GPU and RDU pools independently. Shown for SuperComputing 2026 with GLM-5.3, the system demonstrates hardware integration, not a measured performance or cost advantage; SambaNova provided no end-to-end speed, cost-per-token, or system-energy comparison.
01
Each SambaRack contains 16 RDUs and typically draws 20 kW for inference; 16 racks can work together as a unit of up to 256 RDUs.
02
Intel’s D-Series is a 350-watt PCIe card; Intel-branded cards have 160 GB of memory, while partner designs can reach 480 GB.
03
NVIDIA Dynamo routes requests, vLLM serves both stages, and NIXL transfers the KV cache over UCX/RDMA on RoCEv2 Ethernet.
SambaNova says it has demonstrated an AI serving system that divides work between Intel GPUs and its own processors while running in an air-cooled facility. Presented for SuperComputing 2026, the setup pairs four Intel GPU D-Series cards with SambaRack SN50, targeting providers that want to add capacity inside data centers they already operate.
In its November 15 post, SambaNova describes a recorded demonstration running the GLM-5.3 model across both kinds of hardware. The news is the combined system, rather than a new model or chip launch: Intel handles the incoming prompt, and SambaNova’s chips generate the answer.
Two jobs, different hardware
The split follows two stages of inference—the work of running a model to produce a response. Prefill processes the prompt and builds a KV cache, stored information the model uses as it generates its answer. Decode then produces output one token, or piece of text, at a time. SambaNova assigns compute-heavy prefill to Intel and memory-bound decode to SN50.
Intel’s D-Series is described as a 350-watt, air-cooled card that connects through PCIe. Intel-branded cards carry 160 GB of memory, while partner designs can reach 480 GB. That memory holds model weights and the growing prompt cache; the stated memory bandwidth is 1.5 TB per second.
SN50 is SambaNova’s fifth-generation Reconfigurable Dataflow Unit, or RDU, its specialized inference processor. The company says it links model operations into a continuous flow across chips, reducing trips back to memory during answer generation. Each SambaRack holds 16 RDUs and typically requires 20 kW for inference, according to SambaNova.
A serving-framework diagram separates prefill and decode paths and labels their software layers, network transfer, and hardware components.Source: sambanova.ai.
Software makes the split work
The serving stack uses NVIDIA Dynamo to route requests and vLLM to serve both processing stages. NIXL moves the cache using UCX/RDMA over RoCEv2 Ethernet, the network path connecting the two pools. SambaNova’s deployment argument is that this handoff uses Ethernet rather than a proprietary interconnect, alongside serving tools providers may already use.
Reusing earlier work is another part of the design. SN50 retains cached portions of previous prompts in attached DDR memory. When an AI agent returns to a conversation or shared context, Dynamo is designed to route it to the SN50 holding that information. The GPU then processes only the new context, and less cached data needs to cross the network.
Capacity without a cooling rebuild
Separating the hardware also lets providers grow each pool independently. More long prompts would call for additional D-Series GPUs; longer generated answers would call for more SN50 capacity. For larger models, SambaNova says 16 SambaRacks can work together as a unit of up to 256 RDUs, spreading a workload across multiple lower-power racks rather than one dense, liquid-cooled rack.
The intended audience is cloud and inference providers with existing air-cooled space. SambaNova argues that installing systems there avoids the time and expense of retrofitting facilities for direct liquid cooling. This demonstration does not quantify the tradeoff: the post provides no comparative end-to-end speed, cost-per-token or system-wide energy results for the Intel pairing.
For conference attendees, the next step is concrete: SambaNova invited visitors to see the systems at booth 2246 and directed them to Intel at booth 2819.
Reader comments
Newest comments first. Replies stay oldest first.