Recently I have written about China’s problems with EDA tools and chip fabrication. It isn’t all doom and gloom for China, though, and I’d like to talk about some areas where China is doing better. Central processing units (CPUs) are a bright spot.
Much has been written about the Sino-US trade war, especially about Huawei, with a lot of this discussion revolving around access to x86 CPUs from Intel and AMD, and also access to Arm core IP. Currently, Huawei and others still seem to have some access to Intel and Arm technology, and probably will unless sanctions become more robust. If Huawei and other Chinese companies lose access completely, while it would be a heavy blow, it wouldn’t necessarily be a death nail. Chinese companies such as Loongson, Phytium, Huawei, Zhaoxin, and Alibaba have developed CPUs and core IP which can fill in some of the gaps that would appear.
What are CPUs?
Just so we are all on the same page, let’s quickly describe a CPU. The CPU runs the OS and various applications on a device, processing data and giving an output. CPUs used to contain one processor but now usually contain more than one processor (cores). You will have heard of dual-core, quad-core, etc. For example, the latest Qualcomm chips are octa-core based on Arm Cortex processor cores, just like Huawei’s Kirin chip. A computer may even have more than one CPU, each with multiple cores. Multiples CPUs are more common in the server space.
To make a CPU, one of the first things a design company will decide is the core their CPU will use and thus also choosing the instruction set architecture (ISA), an instruction set provides commands to the processor, it is the link between software and hardware. Software designed for one ISA may not work on another without emulation.
ISAs come in two basic types: faster CISC (complex instruction-set computer) architectures, and more power-efficient RISC (reduced instruction-set computer) architectures. A CISC ISA can do multiple things in a single instruction whereas a RISC ISA may need multiple different instructions to complete the same task. Arm, MIPS, and RISC-V are RISC architectures, while x86 is CISC.
Traditionally, if you’re building a server, you need performance and you’ll choose CISC for its speed. If you’re building a phone, you’ll choose RISC.
Once you’ve chosen CISC or RISC, you have to pick a specific processor core to base your design on, and thus the ISA. Choosing a core commits your CPU to its ecosystem of compatible tools, apps, middleware, etc., so it can be hard to switch if your team has already become accustomed to a certain ecosystem. I’ll look at the field by dividing it into processor cores or ISAs.
The smartphone industry grew with Arm and its ecosystem and today basically all handsets use an application processor based on the Arm architecture.
Although Qualcomm left the Arm server market, along with its China JV Huaxintong, and there is endless debate as to whether Arm can ever replace x86 in the server space, there are still a number of western companies such as Marvell, Amazon, and Ampere developing Arm server chips, and for China, it has become even more important to develop them.
In the Arm camp we have Huawei, and the lesser known Phytium. With the sanctions ongoing rumors have arisen that Huawei’s HPC Compute and HPC Storage business lines may be closed sometime this year as they have limited access to Intel CPUs and Intel support. This has increased Huawei’s efforts to replace Intel with its own Arm-based server grade CPUs. The fruit of these efforts to date was the impressive Kunpeng 920. One of the most powerful Arm CPUs in the market Kunpeng boasts 64 cores running at 2.6Ghz, at 7nm. The chip will mainly target cloud services and big data applications in Huawei’s Taishan server range. And while I believe it and future generations of the chip can be a success in this space it does not address Huawei’s concerns at the bleeding edge of HPC or even supercomputing.
The other Arm-based CPU player is China Electronics Corporation subsidiary Phytium. The company has a good relationship with the Kylin OS team, their offices are next to one another, along with the Beidou team. Indeed, much of Phytium’s team is also part of the National University of Defense Technology (NUDT). Although its HQ is in Tianjin, much of the R&D is in Changsha, near NUDT. Phytium has a range of desktop PC and server CPUs but has more of an HPC focus. Its latest Arm chip is also 64 core but runs at around 2.2Ghz and uses a 16nm process. Phytium’s claim to fame is of course its use in supercomputers like the Tianhe-2, but actually the majority of the processing here was done by Intel Xeon and Intel Phi CPUs, Phytium’s was mainly processing front-end tasks. This changed though due to an Obama ban resulting in the removal of Intel Phi, NUDT replacing it with its self-developed 128 core Matrix-2000 processor. This wasn’t as powerful as Intel’s latest offering at the time, but still more powerful than the older Intel Phi processors they were replacing. NUDT and Phytium while cannot instantly replace HPC workloads are moving strongly in that direction.
MIPS, like Arm, is based on a RISC ISA, but it’s less successful. It never got traction in the handset world, but it’s got a long history in China.
The first commercial Chinese CPU I know of (please message me if you know of earlier commercial CPUs), was the Loongson (aka Godson) which began life in the Institute of Computing Technology (ICT) back in 2001 and was commercialized by Loongson and Lemote. You can Google image some nice pictures of the Lemote laptop and desktop PC. These days, its latest CPUs are based on MIPS architecture which is now owned by Waves Computing out of the US and are fabricated by ST Microelectronics from Europe. The latest Longsoon CPU runs at 2Ghz and has four cores on a 28nm FD-SOI process, fine for government civil servant desktop use. For next year it does have some plans to move to a 16 core 2.5Ghz 16nm CPU, not world beating, but a big improvement. Whilst it isn’t going to become a household name it can be used in certain government or military use cases.
While Loongson may have been the first, it’s the only general purpose CPU right now based on MIPS, another company, Ingenic is also using MIPS but for low-power application specific processors. Other Chinese CPU players base their designs off Arm, x86, or RISC-V.
Intel’s x86 is the only major option for CISC, and as such dominates laptop and PC CPUs, and has a leading position in servers.
Some of you may have read my article on AMD’s JVs in China, so will have seen the name of Zhaoxin Microelectronics before. Zhaoxin is a JV between Taiwan’s VIA Technologies and the Shanghai government. Along with Intel and AMD, Zhaoxin has access to the x86 ISA. While I do not see the company competing with its peers at the high-end, Zhaoxin is showing progress in the low-end server, desktop PC, and notebook space. Its latest KX-6000 16nm, 8-core, 3Ghz CPU performs on par with a 2017 Intel i5, enough for China to be independent at the low-end. Zhaoxin is planning a 7nm CPU for 2020 as well. Let’s keep an eye out for performance figures.
Zhaoxin, along with AMD’s JV (THATIC), discussed in my previous article, are China’s attempt to remain independent when it comes to x86 CPUs. Right now, THATIC is limited to AMD’s older Zen architecture, rather than the newer Zen 2, and in all honesty its Dhyana CPU is said to be more or less a rebranded EPYC CPU. Rather than being world leading, the aim has just been to have access to decent x86 CPUs domestically.
Zhaoxin has more freedom to develop than THATIC so I expect to see more from it in the coming years, where as THATIC seems to be in more of a frozen state given restrictions placed upon AMD.
For now, RISC-V, is the open source budget option, mainly seen in low margin IoT applications where price beats ecosystem. But I foresee that changing as China pushes to reduce exposure to US export regulations, its ecosystem grows, and is proven in more and more applications.
I have previously written about the importance of RISC-V to China. In my opinion it is China’s best bet at becoming relatively self-reliant in CPUs. Details are in my previous article, but in brief, despite originating in the US, RISC-V is sanction proof and can be used as a base for CPUs ranging from low-power IoT to high-performance computing.
Whilst there are many companies working on RISC-V based SoCs, MCUs, and CPUs, Alibaba gained most of the headlines in 2019 when its new subsidiary T-Head announced the RISC-V-based c910 processor and RISC-V-based Hanguang 800 AI processor. The C910 CPU and Hanguang NPU help Alibaba improve AI features in its cloud services and allow it become less reliant on foreign chip suppliers like Nvidia, as well as giving it new revenue streams from by potentially selling its CPUs or licensing core IP to third parties.
China is on the road to CPU independence in the low and mid end, but it will be a few years yet before it can be independent in the high-performance computing space, but I am confident it can get there. The current geopolitical situation is only going to speed up the process and even if true, loophole-free sanctions are put into place China will only be slowed down temporarily, life finds a way. China will struggle to break into the top tier with most architectures, which must be licensed and do not allow licensees access to source code. RISC-V gives it the opportunity to do so. Ironically, it is a US open-source processor movement that has provided this opportunity to China.