Who will go into volume production first for AI ASIC?

AI and deep learning has generated lot of excitement over the past few years. Many semiconductor start-ups have emerged since to build chipset optimized for AI. They tackle compute, communication and memory related problems specific to AI algorithm accelerations and build highly optimized architecture that promises low power and high performance. Nervana was perhaps the first companies to build a chipset specifically for AI who got started around 2014. Nervana wanted to sell cloud services based on their chipsets and bypass the ASICs altogether. When Intel bought them for $350+ million, it got everyone’s attention and suggested an exciting times ahead for AI chipset industry. At the time, Intel announced that the silicon will be available in H1 2017.

Since 2016, a large number of startups have appeared who have raised quite a bit of funding till date. They are generating headlines and the race to get to market first is on.

Graphcore, a UK based company has raised over $110 million and got the valuation of $625 million at its last round.  Earlier they announced that the chip will be released in 2017 but recent announcements suggest that it will be released sometimes in 2018. They have been publishing their benchmarks of late suggesting the early silicon samples are back. Cerebras, a somewhat late entrant to the game in 2016, based in San Jose, California, has raised over $112 million giving a valuation of $700M. That is a great number for a company without any product to show. However, as of 2018, they have neither declared details of their architecture nor announced a date for shipment.

Groq, another high-profile semiconductor start-up was founded in 2017 by Google engineers.   They claim to be able to run best inference and released some impressive benchmarks on their web site. They’ve announced that the chip has been taped out and expect first silicon samples by the end of 2018.

That leaves Campbell based start-up Wave Computing. The company is led by semi industry veteran Derek Meyer and has been making steady progress over the last two years. In 2017 they announced that their chipset has taped out and started accepting applications for early engagements last year.  During our recent briefing, they announced that they’ve started shipping to early access customer and demonstrated their AI workstation in the company conference room. 

Who will go into volume production first for AI ASIC?

AI and deep learning has generated lot of excitement over the past few years. Many semiconductor start-ups have emerged since to build chipset optimized for AI. They tackle compute, communication and memory related problems specific to AI algorithm accelerations and build highly optimized architecture that promises low power and high performance. Nervana was perhaps the first companies to build a chipset specifically for AI who got started around 2014. Nervana wanted to sell cloud services based on their chipsets and bypass the ASICs altogether. When Intel bought them for $350+ million, it got everyone’s attention and suggested an exciting times ahead for AI chipset industry. At the time, Intel announced that the silicon will be available in H1 2017.

Since 2016, a large number of startups have appeared who have raised quite a bit of funding till date. They are generating headlines and the race to get to market first is on.

Graphcore, a UK based company has raised over $110 million and got the valuation of $625 million at its last round.  Earlier they announced that the chip will be released in 2017 but recent announcements suggest that it will be released sometimes in 2018. They have been publishing their benchmarks of late suggesting the early silicon samples are back. Cerebras, a somewhat late entrant to the game in 2016, based in San Jose, California, has raised over $112 million giving a valuation of $700M. That is a great number for a company without any product to show. However, as of 2018, they have neither declared details of their architecture nor announced a date for shipment.

Groq, another high-profile semiconductor start-up was founded in 2017 by Google engineers.   They claim to be able to run best inference and released some impressive benchmarks on their web site. They’ve announced that the chip has been taped out and expect first silicon samples by the end of 2018.

That leaves Campbell based start-up Wave Computing. The company is led by semi industry veteran Derek Meyer and has been making steady progress over the last two years. In 2017 they announced that their chipset has taped out and started accepting applications for early engagements last year.  During our recent briefing, they announced that they’ve started shipping to early access customer and demonstrated their AI workstation in the company conference room. 

Chip development has always been a 12 to 18 month cycle and has gotten even longer with shrinking manufacturing geometry. These companies have been in the development phase for some time now. Nvidia has openly challenged start-ups by making rapid progress in its GPU compute capacity and open sourcing its DLA (Deep Learning Accelerator) engine. The challenges with chip design have remained the same over the years. Just because someone has taped out, doesn’t mean that they’ll sample on time. Similarly, just because someone is sampling, doesn’t mean that they will be able to go into volume production. In semiconductor industry, Murphy’s law – ‘everything that can go wrong will go wrong’ seems to come true more often than not and the race for new AI ASIC is now reaching the final stages. We’ll be monitoring the space closely to see who crosses the line first.