» Current | 2022 | 2021 | 2020

Linley Newsletter

SambaNova Takes On Nvidia’s DGX

February 9, 2021

Author: Bob Wheeler

Seldom does a university research project yield a unicorn startup, but that’s the case with SambaNova Systems. Although it isn’t a spinoff, its architecture was born at Stanford University, and two professors cofounded the company in 2017. Despite raising a massive $456 million in funding and shipping its first production systems in 2Q20, SambaNova remained stealthy until recently. Now that it’s ready to serve a broader customer base, it has announced its DataScale systems for the complete neural-network-model life cycle from training to inference and retraining as well as for data analytics. These systems feature a custom AI-accelerator chip called the Cardinal SN10 Reconfigurable Dataflow Unit (RDU).

The RDU employs a coarse-grain reconfigurable architecture (CGRA), using statically configured switches to create an interconnect fabric. The fabric links alternating memory units (MUs) and compute units (CUs), which number 640 each in the startup’s first-generation Cardinal SN10 RDU. Each CU contains 96 SIMD function units that handle various integer and floating-point data types. The 7nm SN10 has six DRAM channels for external memory, a PCI Express host interface, and four ports for a multichip interconnect. The company rates its DataScale SN10-8 system, with eight of its chips in a quarter-rack chassis, at 2.6 petaflops per second for FP16 data.

Several startups developing AI-training chips have achieved working silicon only to struggle in delivering a comprehensive software stack. By contrast, SambaNova had TensorFlow-compatible software before silicon, allowing use of FPGAs for prototype hardware. It touts broad TensorFlow and Pytorch compatibility, letting customers employ pretrained models without modification. The startup also claims to handle large models enabled by 1.5TB of external memory per chip, or 19x the capacity of Nvidia’s largest A100. Most AI-chip vendors focus on inference, but SambaNova joins Cerebras, Graphcore, and Intel (Habana) in challenging Nvidia’s AI-training dominance.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products


Linley Spring Processor Conference 2022
Registration now open: April 20-21, 2022
Hyatt Regency Hotel, Santa Clara, CA
Linley Fall Processor Conference 2021
Held October 20-21, 2021
Proceedings available
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »