Supermicro to build AI inference systems with SambaNova SN50 chips
Supermicro is building and selling inference systems for agentic AI that combine NVIDIA GPUs with SambaNova SN50 RDUs. The initial shipment covers 10,000 SN50 chips, with cloud service providers set to receive the systems first.
- Supermicro is building systems that combine NVIDIA GPUs with SambaNova SN50 RDUs.
- The initial shipment covers 10,000 SN50 chips.
- NeoClouds and sovereign clouds will receive the systems first.
- Supermicro will provide integration, delivery and support.
- SambaNova expects greater inference capacity to be deployed in 2027.
NVIDIA GPUs and SambaNova chips in one system
According to SambaNova's November 3 announcement, Supermicro will integrate, validate, deliver and support disaggregated inference systems that pair NVIDIA GPUs with SambaNova SN50 RDUs. The systems are designed for agentic AI inference in data centers.
The phased rollout will begin with an initial shipment of 10,000 SN50 chips. The products will first be sold to NeoClouds and sovereign cloud providers, before later reaching companies that run AI in their own data centers.
According to SambaNova, the partnership is driven by expectations that agentic AI systems will significantly increase demand for inference. Such agents can call large language models thousands of times, and completing a single call can take several minutes. As a result, even modest gains in speed and efficiency can be valuable.
Different hardware is needed to process requests and generate responses
According to the company, large language model inference can be divided into two stages. The prefill stage processes the prompt and is primarily compute-intensive, making it a strength of NVIDIA GPUs. The decode stage, by contrast, generates the response one token at a time and is therefore primarily memory-intensive.
SambaNova SN50 RDUs were specifically designed for this second task. According to SambaNova, agentic workloads spend most of their time in the decode stage, making SN50 RDUs particularly well suited to accelerating them. The system's logic is to use the right hardware for each phase.
The company claims that using SN50 RDUs and NVIDIA GPUs together enables faster and more efficient inference than either hardware type can provide on its own. SambaNova also said that SN50 RDUs deliver high-speed, high-throughput inference for large models without requiring exceptionally high rack density or power consumption. This means, the company said, that they can also be deployed in existing data centers.
Supermicro will handle integration and support
Rodrigo Liang, co-founder and CEO of SambaNova, said customers can now order SN50 RDUs through a global partner from which they can already buy GPUs. “This is how fast and efficient inference becomes broadly available,” he said, according to SambaNova's statement.
Charles Liang, president and CEO of Supermicro, said the collaboration expands the company's DCBBS portfolio. According to the statement, this includes support for heterogeneous architectures that combine GPU computing with specialized inference acceleration.
Supermicro will be responsible for integration, validation, delivery and support, including on-site service. SambaNova engineers will work with Supermicro teams on customer deployments. They will also ensure that models run properly and are tuned for the best possible performance on SN50 RDUs.
From the customer's perspective, SambaNova expects the partnership to simplify procurement: customers will be able to buy SN50 RDUs from the same channel partner as the rest of their AI infrastructure. SambaNova expects the collaboration to enable significantly more high-performance inference capacity to be deployed in 2027 than would be possible without it.
SambaNova: Supermicro Will Build and Sell SambaNova-Powered Inference Systems
