We have had some important news over the past two weeks about chips in China.
First, on November 3, gaming and messaging giant Tencent unveiled three chips they have designed in-house (this link is a bit flaky, so here is the English version of the press release which has less information). This news is not surprising in that Tencent is one of the seven largest data center builders in the world. And just as the other six have all dived in to building their own chips, it makes sense that Tencent would follow suit. On the other hand, this news is surprising in that Tencent has been quietest on the subject, befitting their very different corporate culture. In our last post on the state of roll-your-own silicon efforts, we listed Tencent in the very speculative camp, as in they are probably working on something but we have not heard much about it.
We were deeply interested in the three chips they chose to build – an AI accelerator, a smart networking chip, and a video encoding processor.
The Zixiao AI accelerator is the least surprising. At this point, everyone (with the possible exception of Microsoft) has built one of these. AI accelerators are also fairly straightforward to build. They are the ultimate special purpose chip – all they have to do is perform a lot of matrix multiplication. Building these in-house is also important because every company does their own version of matrix math, and small changes in those operations can have outsize impacts on the performance of the chips. For companies operating at Tencent’s scale a custom version of this chip is the obvious path.
The Smart Networking chip, the Xuan Ling, looks more prosaic to us. Tencent obviously has to contend with a lot of networking traffic and having a special purpose chip for this makes sense. The big data center owners have struggled with this problem over the years, trying to reduce their reliance on network box makers, balanced against the fact that those box makers do a pretty good job. As far as we know, Microsoft is the only one of the big data center providers to actually build their own network chip (which was really just an FPGA reliant product), but we strongly suspect that Google has one and probably Amazon as well. Put simply, this is not a huge market and this chip probably does not put a big dent into the share of Broadcom or other merchant providers.
By far, our favorite of Tencent’s three chip is their Video Processor, the Canghai. Google essentially invented this chip six months ago with their VCU. That chip was notable for many reasons, including the fact that it was largely designed by software engineers. The VCU, and now Tencent’s Canghai, are the ideal example of why companies want to build their own silicon. Google did not just invent the chip, they invented a whole category of chips, there really are limited merchant options available, yet the need is so glaring. These processors can greatly reduce the bandwidth costs of serving videos to users, and that bandwidth can be expense. So spending a $100 million (probably a lot less) to offset what is likely billions of dollars in bandwidth costs is an easy decision to make.
All in all this is a solid line-up of smart products, and really promotes Tencent to the big leagues of in-house silicon designers.
The other big news came from Innosilicon. They launched what is arguably China’s first high performance GPU. Admittedly, we know very little about Innosilicon. They are on our (very long) list of emerging Chinese silicon designers. And prior to the launch of their Fenghua GPU, they were largely in the business of providing IP (i.e. designs) of various chip components, so this is a big jump for them. Crucially, Innosilicon partnered with Imagination, who makes IP for graphics processing. This is noteworthy because Imagination is one of the few foreign chip companies to actually get acquired by a Chinese buyer (PE fund Canyon Bridge). So there is a geopolitical angle to this story. As China seeks to build up its semis capabilities, using home-owned GPU IP is an important milestone.
When assessing all this, it is important to look past the product and view their unveiling as part of China’s national semis initiatives. There was not much flag waving in either announcement, but the message was not lost on anyone, and both Tencent and Innosilicon have greatly boosted their domestic political capital.
That being said, we do not know much about the performance of any of these chips, or really any of the technical specs. We found very little information on any of them, other than generic statements about superior performance. Given the very spotted track record on these topics in China, we have to caveat all these claims. We do not know much about the IP used (other than press reports of Arm cores in the Tencent chip), which foundries were used (probably TSMC), or what, if any, foreign companies supported these designs. This should not detract from some pretty impressive news, but we will keep an asterisk next to them until we learn more about their timing and performance.