DeepSeek accelerated to record speed on just 16 SambaNova chips – 20 times more Nvidia chips would have been needed

AI startup SambaNova Systems, founded in 2017 by people from Sun, Oracle, and Stanford, has announced the world’s fastest DeepSeek-R1 671B deployment. That’s not possible with Nvidia hardware.

Image source: sambanova.ai

SambaNova achieved a DeepSeek-R1 speed of 198 tokens per second on just 16 custom-built accelerators, a feat that would require 40 racks of 320 Nvidia GPUs. “SambaNova’s SN40L RDUs are the fastest platform to run DeepSeek, a 5x increase over the speed of the latest GPU on a single rack, and we’ll be offering 100x the capacity for DeepSeek-R1 by the end of the year,” promised Rodrigo Liang, co-founder and CEO of SambaNova.

While computationally intensive AI workloads have traditionally been powered by Nvidia GPUs, SambaNova claims its configurable dataflow architecture is a more efficient solution. Its hardware runs three times faster and uses five times less power than today’s most powerful GPUs, while still delivering the full computing power of the DeepSeek-R1. The achievement was confirmed by experts at Artificial Analysis, an independent AI assessment firm.

The open source DeepSeek-R1 671B is available on the SambaNova cloud via API. The company is actively increasing its capacity and hopes to reach a total throughput of 20,000 tokens per second in the near future.