Supermicro plans AI inference racks pairing Nvidia and SambaNova processors
SambaNova says the collaboration initially plans for 10,000 SN50 chips, with AI-focused cloud providers among the first customers targeted.
Supermicro will manufacture and market racks that use both SambaNova SN50 chips and Nvidia GPUs to run AI models, the companies said. According to SambaNova, the collaboration initially plans for 10,000 SN50 chips.
SambaNova said the processors will perform different tasks: Nvidia GPUs will read incoming prompts, and SN50 chips will produce the answers. Supermicro will add the racks to its Data Center Building Block Solutions range of ready-made systems.
Assembly, testing, delivery and customer support will be Supermicro’s responsibility, with on-site service included. Engineers from SambaNova will adapt and tune AI models for SN50 and help customers deploy them.
The initial sales effort will focus on neoclouds—providers built around AI—and sovereign clouds that retain data and computing within national borders. The companies plan to approach businesses running AI in their own data centers afterward.
SambaNova identifies SN50 as the fifth generation of its Reconfigurable Dataflow Unit, a processor intended for inference. It puts average power consumption for a SambaRack SN50 at 20 kW and says operators can increase capacity in existing facilities rather than await new construction delayed by power and cooling constraints.
Oracle also plans to put SN50 chips into its cloud data centers and make the capacity available to AI model builders. It announced those plans during its AI World conference on 27 October.
