X
Innovation

'We are the best-funded AI startup,’ says SambaNova co-founder Olukotun following SoftBank, Intel infusion

SambaNova added $676 million in venture money to build out its Dataflow SaaS service for training deep learning neural networks.
Written by Tiernan Ray, Senior Contributing Writer
kunle-sambanova-april-2021.jpg

"I think most people would say we are the most credible competitor to Nvidia," says Kunle Olukotun, Stanford University computer science professor and co-founder of AI startup SambaNova Systems. SambaNova Tuesday announced a new round of venture capital funding that brings its capital to date to over $1 billion.

In yet another sign of the rising interest in alternative computing technology, AI systems startup SambaNova Systems on Tuesday said it has received $676 million in a Series D financing from a group of investors that includes the SoftBank Vision Fund of Japanese conglomerate SoftBank Group; private equity firm BlackRock; and the Intel Capital arm of chip giant Intel.

The new funding round brings the company's total investment to date to over $1 billion. The company is now valued at more than $5 billion.

"With this $676 million, we are the best-funded AI startup," said Kunle Olukotun, a professor of computer science at Stanford University, and a co-founder of SambaNova, in an interview with ZDNet via Zoom. SambaNova competes with other heavily funded startups, including Cerebras Systems and Graphcore.

"We're using this round to build our software, hardware, and systems collateral, to challenge the incumbent," said Olukotun. 

The incumbent, in this case, is Nvidia, whom SambaNova claims to be able to best on benchmark deep learning tasks, with a much-smaller footprint in an equipment rack, and much less power.

sambanova-datascale-2020.png

SambaNova's DataScale computers are built with novel chips and software and memory routing technology, with a focus on solving the training tasks of large deep learning forms of neural network AI.

SambaNova has pioneered a novel approach to moving neural network programs through circuits to train those programs, and a novel chip design, which it calls a data-flow architecture. 

The company has built computer systems to run the chips and software, called DataScale. The DataScale system is comparable to sixty-four of Nvidia's DGX-2 rack-mounted systems running the A100 GPU, but in only one quarter of a standard telco rack, says SambaNova. 

Also: SambaNova claims AI performance rivaling Nvidia, unveils as-a-service offering

And the company claims it can get a 2000-times performance-per-watt improvement in energy use on the same benchmark deep learning tasks as the DGX systems.

"I think most people would say we are the most credible competitor to Nvidia," observed Olukotun.

In December, SambaNova it began offering a subscription service, called Dataflow as a Service, where companies effectively lease the equipment to run on-premise, with support added in to ease the task of building deep learning programs such as very large natural language processing models like Google's BERT.

It is the service offering in particular that is going to be boosted by the capital infusion, the company said.

Also: 'It's not just AI, this is a change in the entire computing industry,' says SambaNova CEO

"There's been a lot of customer pull and demand in that space," said Marshall Choy, Vice President of product at SambaNova, referring to the services offering, in the same video interview. "And so we're continuing to amplify the R&D investment there on the tooling and the tool chain and the usability to provide people the very seamless API interface, and take away the struggle of dealing with the infrastructure.

Choy declined to offer stats on the customer uptake of Dataflow as a service.

SambaNova is now over 350 people, and is actively hiring, said Olukotun. The company intends to get to over 500 people by the end of this year, he said. That will be across functions, he said, including R&D engineers, field service engineers, go-to-market staff, and others.

Editorial standards