As industries from automotive to healthcare race to take advantage of massive amounts of unstructured data, another startup is moving into the high stakes world of custom chips that can accelerate machine learning software. And the new company is getting off the ground with lots of financial fuel.
SambaNova Systems closed a $56 million funding round led by the venture capital arm of Google parent Alphabet and Walden International, a firm chaired by Cadence’s chief executive Lip-Bu Tan, among others. The company is building not only the silicon but also the software for machine learning and data analytics tasks, such as parsing databases.
The Palo Alto, California-based company was founded in part by C.E.O. Rodrigo Liang, former vice president of processor development for Oracle. He came to Oracle via its acquisition of Sun Microsystems, which previously purchased Afara Websystems, where he was director of engineering next to another SambaNova founder, Kunle Olukotun.
Founded by Olukotun, an electrical engineering and computer science professor at Stanford, Afara created server chips based on the SPARC architecture. Olukotun was one of the architects of Sun Microsystem’s T1 processor – codenamed Niagara – which combined multicore and multithreading technology for the first time.
SambaNova’s other founder is Chris Re, a professor of computer science at Stanford. He previously founded a startup called Lattice that used machine learning to sift through databases to answer queries. Along with Michael Cafarella, the co-creator of Hadoop, a popular big data processing technology, he sold the company to Apple last year.
Olukotun said that SambaNova, which was only founded in November, is focused on “software-defined hardware” that balances flexibility and energy efficiency and adjusts to advances in machine learning. The company indicated that its chips would scale from edge devices to enterprise systems. But it is still unclear whether the company is targeting training or inferencing – or both.
A spokesperson did not respond to multiple requests for an interview with SambaNova.
The funding is an adrenaline shot for SambaNova. But the company will have to stretch it further than its biggest competitors, such as Intel and Xilinx, which have thousands of engineers and war chests filled with billions of dollars in spoils from when Moore’s Law ran smoothly. But no one yet has successfully challenged Nvidia, whose graphics chips dominate the machine learning space.
The race to artificial intelligence has reinvigorated the semiconductor industry and unleashed an explosion of investment in chips. “We’re in this era of exponential change, and the speed of innovation is outpacing the silicon design cycle,” said Victor Peng, C.E.O. of programmable chip maker Xilinx, in a recent conference call with reporters.
Following a long hiatus, investors are once again targeting semiconductor startups . The investors in Graphcore, which has raised $110 million since 2016, include Sequoia Capital and researchers like DeepMind’s Demis Hassabis. Wave Computing has drummed up funding from Dado Banatao, while Cerebras Systems, reportedly valued at almost $1 billion, is funded in part by Benchmark Capital.
Several of these companies could enter volume production soon. Graphcore plans to scale up production of chips based on its Colossus architecture by the end of the year. Wave Computing has started an early access program for server systems using its dataflow processing unit, which unlike Nvidia GPUs is not an accelerator.
It is not clear where SambaNova is in the development of its hardware, but the company plans to hire aggressively in the next few months and accelerate the rollout of its technology. “We have exposed our technology to some of the world’s largest companies across different industries,” Liang said in a statement.
SambaNova’s technology and management team have enamored investors, which also include Redline Capital and Atlantic Bridge Ventures. Alphabet is not completely unfamiliar with the sector, with Google on the second generation of the TPU, its custom chipset for machine learning training and inferencing. It is now available through Google’s cloud.
“Other platforms have been designed for A.I. and machine learning techniques that exist today. SambaNova’s software-defined infrastructure anticipates and supports a rapidly-evolving ecosystem,” said Dave Munichiello, general partner at GV, previously known as Google Ventures. “We firmly believe that over time this computing approach will lead the industry.”