Alphawave: Hyperscaler Customer Chip Development

This is a snippet of the transcript, sign up to read more.

You referenced Broadcom's acquisition strategy which is similar to Synopsys, who also sell IP in the space. Alphawave claim all their chips for every interconnect are built with the same architecture which leverages their specialty in DSP. How easy or difficult is it to work with a company who has different architectures bought through different acquisitions, versus one with one architecture consistent across interconnects?

The ASIC business Broadcom runs is separate from their other businesses. A good friend of mine works on the SerDes transceivers Broadcom make themselves. Alphawave has an advantage when they work with one of those hyperscalers because their main business is this IP plus the SOC part to make a chip out from it. For Broadcom, if it's not profitable and doesn't extract much cash for Hock Tan, he will simply get rid of the division. If I were Google, I would want to partner with Alphawave rather than Broadcom. Synopsys' problem is their main business is EDA tools, however there are many good open-source EDA tools coming online. Google is a big proponent of that and I am now building chips without paying a dollar to Synopsys by using open-source EDA tools.

This is a snippet of the transcript, sign up to read more.

Do others like them exist? Obviously, Broadcom have their own internal teams but how is that different from OpenFive?

The difference is that OpenFive did not have any IP whereas Broadcom has high speed SerDes IP, which Alphawave also has after acquiring OpenFive. Other players do exist and I have used them when I have done low budget chips in the past. Broadcom are the most expensive design house to take RTL to GDSII. Uniquify is a small shop who Marvell and Broadcom outsource to when they run out of time. Several small houses are doing good work and are viable businesses. TSMC has an arm called GUC, and I made a chip at Google using them. They might have the TSMC name behind them but they don't speak English well. Broadcom has an advantage as you can speak to their guys who are in the Bay Area. GUCs documentation is also not the best quality.

This is a snippet of the transcript, sign up to read more.

Do the other hyperscalers also act that way or do they move slightly faster?

Everybody is faster than Google, even IBM. Facebook interviewed me on several occasions and they definitely hired many people to do something similar to Google's TPU, but they haven't announced they have something, so maybe Facebook is slower. Amazon is the quickest, then comes Microsoft, but I have no experience working with either. When Google makes a TPU, it is not a big volume business. They are willing to pay $10 million for it but could get it for under a million from Uniquify. Automotive will be more interesting because every car needs to have more compute. Tesla has a self-driving chip which is a much more interesting business for Alphawave.

This is a snippet of the transcript, sign up to read more.