The chip cold war between China and the United States continues to heat up like a processor with a heat sink that it is a little too small. And the scuttlebutt is that this week Alibaba, the Chinese online retailer and cloud service provider that is roughly analogous to Amazon and its Amazon Web Services division, will launch its own Arm server processor.
The announcement, according to various reports, will happen sometime this week at Alibaba’s annual Aspara conference, which runs from October 19 through 22. And it will represent yet another step in China’s desire to have semiconductor independence, which we discussed a month ago in detail.
A report in the South China Morning Post, which cited a report in the media aggregator site Caixin that has been subsequently removed, says that Alibaba taped out whatever this server chip is called in the early summer and may be “releasing it soon.”
This report also indicated that this rumored server chip is based on a 5 nanometer manufacturing process, which strongly suggests that this processor will be etched by Taiwan Semiconductor Manufacturing Co (TSMC). How long this will last remains to be seen for reasons we will discuss in a moment. Suffice it to say, Semiconductor Manufacturing International Corp (SMIC), the largest indigenous foundry in China, has peaked thus far with a 14 nanometer FinFET processor and can’t do it, and neither can United Microelectronics Corp (UMC), which is also based in Taiwan and which is actually an older foundry than TSMC, since it has peaked at 14 nanometer processes. Both SMIC and UMC offer chip etching in larger geometries spanning up to 90 nanometers. Korean memory maker and server chip wannabe Samsung has 7 nanometer processes in the field and is manufacturing IBM’s “Cirrus” Power10 processor, launched a month ago in the Power E1080 server, but does not yet have production 5 nanometer processes in production. Intel Foundry Services, the chip manufacturing operation that Intel created earlier this year to open up the chip maker’s foundries to the designs of other chip makers, would undoubtedly love to get business from Alibaba. Unfortunately Intel can barely get 10 nanometer processes out the door and has delayed its 7 nanometer processes. While it may get on track with 5 nanometer, that ain’t happening right now.
So, if this Alibaba server chip is etched in 5 nanometer processes, and if it is happening any time soon, it pretty much has to be coming from TSMC.
Ironically but also obviously, Alibaba is a strong partner with Intel when it comes to chip buying, of course, as all of the big clouds are, and is in fact a sponsor of the Aspara conference this week. But Alibaba also has cloud instances based on Ampere Computing’s 80-core Altra and presumably is in line to get the 128-core Altra Max processors that will ship any day now.
It may seem odd that Alibaba would develop its own Arm server chips when it has a partnership with Ampere Computing, but as we discussed this summer with Ampere Computing chief executive officer Renee James, there is every reason to believe that many of the big clouds will have a dual source approach for Arm server chips, developing their own as well as relying on Ampere Computing, thus hedging their bets against the risk of a failure either internally or externally to deliver a chip on time. When new chips don’t get rolled into datacenters on time, it affects cloud product launches, price/performance and competitive stance, and profits at the big clouds. Moreover, if any organization in China does anything with Altra and Altra Max chips, that makes the US government touchy. Ampere Computing’s customers could be put on the US Department of Commerce’s Bureau of Industry and Security entity list, which blocks sales of chippery and puts pressure on TSMC and other suppliers not to do business with China. Or there could be Coordinating Committee for Multilateral Export Controls (CoCom) export restrictions placed on Ampere Computing’s chips. Or anyone else’s, for that matter. (We are not picking on Ampere Computing, but just pointing out that if its processors are used for military purposes, there are mechanisms for restricting export and import to any country.
This, in fact, just happened to Phytium Technology, the upstart Chinese Arm server chip designer that decloaked six years ago with the 64-core “Mars” FT-2000/64 server chip based on the Arm instruction set and which provided some details about this chip a year later. The original FT-2000/64 Mars chip was etched in 28 nanometer processes, but there was a kicker FT-2000/64+ chip based on TSMC’s 16 nanometer processes. These are not advanced processes at all, so the chips are big and hot by comparison to those using 7 nanometer or 5 nanometer processes.
Here is why we bring up Phytium, which had an aggressive roadmap to adopt 7 nanometer and 5 nanometer processes at TSMC to drive its roadmap, jacking up the core counts to 128 and beyond. Apparently, Phytium chips were used in the supercomputers that did the simulations for the DF-17 hypersonic medium-range missiles revealed by China at its National Day parade two years ago this month. After finding this out, the US Department of Commerce put Phytium on the Entity List in early April this year, and two weeks later TSMC stopped supplying chips to Phytium. And last week, lo and behold, both Alibaba and Baidu took stakes in Phytium. Another hedging of bets by two of the four Chinese hyperscalers and cloud builders, it would seem.
Alibaba has been working towards creating its own chips since seeing the success that AWS has had with its “Nitro” DPUs, which came from its acquisition of Annapurna Labs in 2015 and which has been expanded into the Graviton and Graviton2 general-purpose CPUs, the Trainium machine learning training accelerators, and the Inferentia machine learning inference accelerators. And, quite possibly, homegrown AWS switch and router ASICs. If AWS can and has vertically integrated its stack, then Alibaba has to as well — particularly if there is a cold war between the United States and China where tech is wielded as a weapon.
Alibaba had been doing chip research through its Alibaba Damo Academy research and development arm for some time before then. This chip development arm is variously known as T-Head and Pingtouge Semiconductor. Then Alibaba got serious and acquired a startup called C-Sky Microsystems in April 2018 and announced its plan to create its own AI chip aimed at video surveillance in September 2018, with the hope to get its Hanguang 800 AI inference chip out of the fabs by the end of 2019 and into production through 2020. This chip has 17 billion transistors and was made using TSMC’s 12 nanometer process — the same one used by Nvidia to make its “Volta” V100 accelerators. In July 2019, Alibaba said it was working on the Xuantie 910 RISC-V chip, which it unveiled at Hot Chips in the summer of 2020 with a 16-core variant.
Alibaba, like other Chinese hyperscalers, has to walk a fine line, and particularly since it runs a cloud and is not just running its own applications (which Baidu and Tencent do a lot more of). With most of the world’s application code running on Windows Server or Linux running on X86 processors, and no indigenous suppliers of X86 processors (excepting the Tianjin Haiguang Advanced Technology Investment Co partnership AMD has with the government of Tianjin, which is also the backer of Phytium, by the way), Alibaba has to be careful to not annoy the US government so it can get processors from Intel, AMD, and Nvidia. But it also has to work towards the Chinese government’s goal to have chip design and manufacturing independence.