SambaNova Releases Fourth-Gen Chip
Author: Anand Joshi
SambaNova Systems, an AI start-up targeting the data-center market, continues to march forward with its fourth-generation chip. The new SN40L employs TSMC 5 nm technology and offers 688 FP16 Tflop/s. The chip is available to customers on the SambaNova-hosted cloud with on-premises shipments scheduled in 2024.
SambaNova’s chip architecture is based on its reconfigurable dataflow unit (RDU) originally developed at Stanford. The company plans to build training and inference systems based on its chip. Although it continues to target large AI models, the company’s initial customers are using the machine for HPC. When used for AI, it is best suited for training, as it offers no additional acceleration for INT8 and other inference data types.
The Silicon Valley startup raised $1.1 billion in funding from investors such as BlackRock, Google Ventures, Softbank, and Intel Capital. Its latest Series D round valued SambaNova at $5.1 billion, a record valuation for a chip company at the time. The company started shipping its first production systems based on its initial AI chip in Q2 2020 and has two more chips in production.
Three years after shipping its first products, the company has disclosed little about its customer success. It has published no AI benchmarks, and its TFLOPS/W falls short of Nvidia’s. Its announced customers include Lawrence Livermore Laboratory, Argonne National Laboratory, and OTP Bank, Hungary’s largest commercial bank.
