Third-party testing by Artificial Analysis showed SambaNova's heterogeneous compute platform, combining four H200 GPUs and 16 SN50 RDUs, achieved a decode speed of 763 tokens a second in MiniMax M2.7 at short context lengths of 10,000 input tokens, beating GPU-only inference platforms. For longer context lengths, the platform sustains more than 450 tokens a second. The computationally intensive prefill phase was handled by four Nvidia H200 GPUs, while memory-bandwidth-bound decode operations were done on a single SambaNova rack containing 16 SN50 accelerators. SambaNova hopes to demonstrate how customers can extend the life of aging GPU fleets by using its air-cooled systems as decode accelerators, deployable in existing datacenters. The startup plans to show off configs with 128 and eventually 256 accelerators.