No, the Ancient Egyptians built the pyramids, but it is worth asking who is building some of the chips capturing the media’s attention recently.
Last week, Alibaba announced it’s first server CPU, the YiTian 710. This is a big achievement, but for some reason there is not a lot of faith in this product. First, it emerged that they had gotten help from Taiwanese companies. This included GUC, a solid ASIC service provider. GUC is in the business of helping other companies design chips and bringing them to production. But beyond that there is a lot of speculation that they had help from some other, presumably US, chip company.
To be fair, this could be entirely baseless. We have literally spent years disparaging claims against Chinese companies supposed inability to innovate. Everyone wrote off Chinese handset companies (and Mediatek) until they were the dominant players in the market. And everyone said Huawei was just a copycat until they delivered 5G products a full year ahead of Nokia and Ericsson.
So our first instinct was to assign these rumors to that same category of baseless rumor-mongering. In this light, that help from Taiwanese companies is trivial. These are service companies that assist with taking chips to production not actual designers. GUC could probably design a CPU, but our sense is that would be a bit of stretch for them. Just as likely that Ali designed the chip themselves.
But then we were reminded of this news. Alibaba unveiled their Hanguang 800 AI accelerator back in 2019, over two years ago. And since then…silence. A search of Ali Cloud’s otherwise ample service offerings does not return any results for Hanguang or AI accelerators. And we could find no other actual news about the product on Google or Baidu. In the press release for the new YiTian, Ali mentions the Hanguang but implies it is still not in deployment yet.
So we have to conclude that the YiTian may not be a fully baked product. Or at least it is reasonable to want some more information about it. Maybe it is not ready. Maybe it was never meant to be a fully commercial product, just a demonstration of Ali’s abilities or a patriotic endeavor.
Designing CPUs is not easy. Advanced processors for general purpose compute are incredibly advanced products, with immense complexity in how they handle instructions, route data, and manage memory. And then on top of that there is the software question. It takes a lot of work to build a software ecosystem that will support a new chip. For an example closer to home, note this Tweet stream about Dropbox’s support for the new-ish Apple M1 Silicon. A year+ after the launch of that chip and Dropbox’s CEO had to personally tweet that they were adapting their software to work on the chip after a few people complained about poor Dropbox performance on their fancy new MacBooks. (We count ourselves among Dropbox Mac M1 users, and there is a notable power drain here.) And that is just one piece of software. If you are running a CPU for a public cloud, there are dozens, maybe hundreds of software packages that need to be ported and optimized for that CPU. Just building the chip on Arm is not enough to ensure optimal performance, and when we are dealing with cloud-scale deployments small inefficiencies like this are enough to sink the economics of the whole product. Go back and look at the Ali press release – there are no software companies mentioned, let alone quoted in support.
Our intent here is not to pick on Ali. Instead, we are simply pointing out that building chips, especially general purpose processors, is still hard. As much as talk about the growing number of companies designing their own chips, this is still not a project most companies can achieve. To make that clearer, we think there is another internally designed chip that may rely heavily on someone else’s support. Google announced their Tensor chip for smartphones last month, it now looks a lot like that chip relies heavily on Samsung’s Exynos applications processor. If Google, who has already brought half a dozen chips to production, who has their own chip design tools, if even they have to rely on a third party to do a lot of the heavy lifting, it just goes to show how hard this is.