Companies including Cerebras Systems, KnuEdge Inc., Graphcore Ltd., Cornami, and Wave Computing are all addressing deep learning, as I reported a year ago. Now Cade Metz in The New York Times has an update on the startups placing big bets on AI. Venture capitalists who were once hesitant to fund companies potentially going up against giants like Intel, Qualcomm, and Nvidia are beginning to show some enthusiasm.
“Today, at least 45 startups are working on chips that can power tasks like speech and self-driving cars, and at least five of them have raised more than $100 million from investors,” Metz writes. “Venture capitalists invested more that $1.5 billion in chip startups last year, nearly doubling the investments made two years ago, according to the research firm CB Insights.”
Metz quotes Bill Coughran, formerly with Google and now a partner at Sequoia, as saying, “Machine learning and AI has reopened questions around how to build computers.” Metz adds, “Sequoia has invested in Graphcore, a British startup that recently joined the $100 million club.”
Metz notes that the startups are looking to find a profitable niche or get acquired. The latter approach worked for Nervana, a 50-employee Silicon Valley startup that Intel acquired for $400 million, according to recode. Mets says Cerebras subsequently scooped up five Nervana engineers and, according to Forbes, has raised more than $100 million in funding.
In addition to Graphcore and Cerebras, Metz reports, Silicon Valley-based Wave Computing and two Beijing companies—Horizon Robotics and Cambricon—have all raised more than $100 million. Another Beijing startup, DeePhi, has raised $40 million.
And a previous report had put investment in KnuEdge, not necessarily a startup at more than a decade old, at the $100 million mark.
Metz quotes Mike Henry, chief executive of AI chip startup Mythic, as saying money in 2015 and early 2016 was a nightmare, but “…with the big, acquisition-hungry tech companies all barreling toward semiconductors,” that has changed.
The companies are reserved in explaining exactly how their chips will work. KnuEdge, for example, says simply, “We’re building the world’s most scalable computing fabric for machine intelligence based on neurobiological principles.” A recent blog post notes, “In high school math class we learned about the two most common modes of logical argumentation: deduction and induction…. For machine learning, we need to augment deduction and induction with two additional modes of reasoning—abduction and transduction. Abduction provides the justification of using statistical methods (‘mostly true’) to look for patterns in data. Transduction implies a specific-to-specific mapping by way of a general class.” Another post is titled “Biomimicry: Emulating Neurons with Hardware.”
Cornami gets a little more specific, writing that its “…multicore TruStream architecture creates a new semiconductor category that allows for any C++ programmer to maximize the concurrency and pipelining of any algorithm directly in hardware. To support this capability, Cornami has developed and patented technology that maps a directed graph (cyclic or acyclic) onto a scalable and potentially unlimited sized collection of processors and efficiently and concurrently executes it.”
Commenting on potential products from Graphcore, Cerebras, Wave Computing, Horizon Robotics, Cambricon, Mythic, and DeePhi, Metz commented, “It is still unclear how well any of these new chips will work. And the chip startups will face competition from Nvidia, Intel, Google, and other industry giants.”
