Last year, we wrote about Google’s VCU chip. At the time we pointed out that one of the most interesting of many things about this chip is that it was designed with Google’s own EDA design tool, called Taffel. From what we can tell, Taffel is designed specifically for aiding software developers in designing their own chips.
At the time, we idly speculated the potential for Google to someday open source Taffel and let software engineers everywhere design their own chips. Since then there have been hints that Google is thinking about this. Notably their sponsorship of “free” chip tape outs on 130nm with Skywater, a US-based fab. Full disclosure, this is a very small step, and feels like Google is only dipping their toe in the water, at most. More intriguing is this video from the Google engineer who seems to be the public face of Google’s efforts here. (Hat tip to Dylan Patel for bringing this to our attention, we’ll pause for a moment while you go subscribe to his newsletter.) This video is interesting both for its explanation of the complex suite of software tools needed to produce a chip as well as for the explanation of Google’s efforts in the space.
But we think there is potential for Google to do a lot more.
In writing about VCU, we noted the significance of the fact that this chip was designed by software engineers, not by semiconductor engineers. The whole point of a chip is to run software more efficiently, and no one knows more about the requirements of that software than the engineers actually writing the code. We have also been writing a lot recently about our broader thesis that the future of compute systems is going to require closer integration between software and semis. Following this logic then leads to the conclusion that the future of compute is going to eventually optimize around allowing software engineers to design their own chips.
This means that Google has the potential to greatly disrupt the semiconductor industry and alter the course of future technology. If you watch that whole video (and we know it is lengthy) you can see that Google has essentially put all the key tools in place for this. There are now a full suite of EDA tools and, PDK resources for getting chips to production. So has the revolution begun?
The reality is that we are not quite there yet. First off, the semis design cycle has a lot of steps to it. Note we mentioned PDKs above, but did not really explain it. These are essentially an interface between the chip design company and the foundry. One of the many barriers to entry TSMC has working in its favor are the very robust set of PDK tools they offer. Likewise, the EDA oligopoly, is in part built on the integration of their tools with the Foundries’ PDKs. There are other design steps and files, like GDS which have been around so long that even seasoned semis engineers we speak with do not remember what it stands for. Most importantly, many of these steps are specific to a single foundry. So today, anyone can go and design a chip and get a ‘free’ tape out at Skywater, but those are produced at 130nm, a process which is about 15 years behind the current leading edge. Getting a new set of tools integrated with TSMC’s 5nm process is much tougher, and not likely to get open-sourced any time soon.
Then there is the actual cost of producing the chip. Skywater is offering ‘free’ tape outs at 130nm, but that only returns a small number of chips. Taping out a chip at TSMC today is likely to cost $20 million upfront, and more for each wafer. This is not something even the best engineer can do as a side project.
That being said, we do think the end result of all of this will be to greatly reduce the cost of designing chips. This will make it possible for a far wider range of companies to start designing their own chips. At this point, anyone that makes hardware has to be at least considering designing their own chips, especially if those chips do not need leading edge manufacturing and run inside powered devices. Cars are a pretty good example of this. Cameras, aircraft, home networking boxes, agricultural equipment, machine tools and industrial systems are a few more.
In all of this, we definitely can see a world in which literally anyone can design their own chip (let’s not call it “democratization of semis”), but the practical realities of it a bit more mundane. Nonetheless, these programs from Google are very important. Every past change in the model of how chips get built has led to massive growth in the number of companies in the industry – this goes back to the 1970’s and the rise of “merchant semis” companies like Intel, to the fabless model in the 1990’s and most recently the rise of RISC V. The pendulum is definitely swinging towards more companies building semis and that will be an exciting place to work and invest.