Artificial intelligence (AI) and deep learning have generated lot of excitement over the past few years. Many semiconductor startups have emerged to build chipsets optimized for AI. They are tackling compute, communication, and memory-related problems specific to AI algorithm accelerations and building highly optimized architectures that promise low power and high performance. Nervana was perhaps the first company to build a chipset specifically for AI, which got started in 2014. Nervana wanted to sell cloud services based on its chipsets and bypass the application-specific integrated circuits (ASICs) altogether. When Intel bought the company for $350+ million, the move got everyone’s attention and suggested that exciting times were ahead for the AI chipset industry. At the time, Intel announced that the silicon would be available in 2H 2017.
Numerous Well-Funded Startups Entering the Race
Since 2016, a large number of startups have appeared and they have raised quite a bit of funding to date. They, too, are generating headlines and the race to get to market first has begun.
Graphcore, a U.K. based company, has raised more than $110 million and received a valuation of $625 million during its last round of funding. Graphcore announced that its chip would be released in 2017, but recent announcements have revised that timeline to sometime in 2018. Graphcore has been publishing its recent benchmarks, suggesting that the early silicon samples are back.
Cerebras, a somewhat late entrant to the game in 2016, based in San Jose, California, has raised more than $112 million giving it a valuation of $700 million. That is an impressive number for a company without any product to show. As of 2018, it has neither declared details of its architecture nor announced a date for shipment.
Groq, another high-profile semiconductor startup, was founded in 2017 by Google engineers. It claims to be able to run the best inference and released some impressive benchmarks on its website. It has announced that the chip has been taped out and it expects the first silicon samples by the end of 2018.
Campbell, California-based startup Wave Computing is led by industry veteran Derek Meyer and has been making steady progress over the last 2 years. In 2017, it announced that its chipset has been taped out and it started accepting applications for early engagements last year. During a recent briefing, Wave Computing announced that it had started shipping to early access customers and demonstrated its AI workstation in the company’s conference room.
Long Development Cycles on the Way to the Finish Line
Chip development has always been a 12 to 18-month cycle at a minimum, and has become even longer with shrinking manufacturing geometry. These companies have been in the development phase for some time now. NVIDIA has openly challenged startups by making rapid progress in its graphics processing unit (GPU) compute capacity and open sourcing its Deep Learning Accelerator (DLA) engine. The challenges with chip design have remained the same over the years. Just because someone has taped one out, does not mean that it will sample on time. Similarly, just because someone is sampling, does not mean that it will go into volume production. In the semiconductor industry, Murphy’s law of everything that can go wrong will go wrong seems to come true more often than not and the race for new AI ASICs is now reaching the final stages. Tractica will be monitoring the space closely to see who crosses the line first.