SILICON | Server chips: a growing opportunity for Chinese fabless players

In recent years, handsets have been key to semiconductor industry growth. So when analysts predicted a grim 2020 for the handset markets, things didn’t look great for semiconductor companies either. In December 2019, analysts expected handset sales to drop 2% to 270 million units in 2020.

Predictions for the rest of the year are weaker, but still miles ahead of the handset market. Some analysts expect a 5-10% drop in global chip sales for 2020 as wireless, automotive, industrial, and general consumer electronics sales all fall. Others predict 3.3% growth for the whole year. It might not be huge growth, but it’s growth nonetheless.

If consumer electronics sales are in freefall, what’s keeping the semiconductor industry from dropping further, even giving it hope of growth—and is this the opportunity Chinese chip makers have been waiting for to make their mark on the industry?

TSMC saw sales of every chip it manufactures drop in volume in H1 2020 growth. Except for one, which grew by 12%: high-end server computer chips.

Less phones, more laptops

While the Covid-19 pandemic accentuated the trend of falling handset sales, it put fuel on the fire of cloud computing.

Cloud services were already moving data storage and processing from edge devices, like phones and laptops, to data centers. But then the pandemic and lockdowns made work from home the norm around the world.

This has not only led to an increase in PC and laptop sales, which grew 11.2% year-on-year globally in Q2 2020, but also an increase in the use of video conferencing, distance learning, and video streaming services.

In China, up to 300 million people were working from home in the first quarter of 2020—and tech companies jumped at the opportunity. Virtually every major internet company brought out new apps to deal with this demand.

Baidu brought out its collaboration tool Baidu Hi. Alibaba released DingTalk 5.0. Bytedance created Feishu. Tencent pushed Tencent Meeting (which it had luckily just released in December 2019). Even Sohu and Pinduoduo got in on the action with Little E and Knock. 

This surge in remote working services has led to a surge in internet traffic, which demands more processing power from cloud providers. More processing power needs more servers, and servers are made of chips: general-purpose CPUs and accelerators like graphics processing units (GPUs) and field-programmable gate arrays (FPGAs).

READ MORE:  SILICON | China’s progress on homegrown CPUs

I don’t believe we will see internet usage drop back to pre-pandemic levels. The amount of data collected by individuals and companies has been exploding for a while now, and it will only grow.

The sudden change in human behavior brought by Covid-19, along with increasing workloads and 5G connectivity, represents an opportunity for Chinese companies to break into a market dominated by Intel, AMD, and Nvidia.

Do one thing, do it well

More people working from home doesn’t just mean more servers; it means a greater mix of servers to cater to the varying needs of different applications. Some servers need to be flexible; some need to be low-cost, and some need to have specific accelerators designed for specific applications.

Not only will the world need more chips, but it will need a greater variety of them. This gives Chinese companies the chance to pick a market and develop a product.

Different types of chips have made their way into the server space in recent years and ever increasingly so. General purpose Central Processing Units (CPUs) aren’t suitable for some applications, so they need help from various different kinds of accelerators.

Accelerators are processors to which the main general-purpose CPU offloads some workload. Graphics processing units (GPUs), used for image and video processing, and more flexible field-programmable gate arrays (FPGAs), and application specific integrated circuits (ASICs) are the main types of accelerators in use.

In 2012, Nvidia found that its GPUs, normally used for processing images and video, were great for AI applications. It has ridden the AI wave to now be worth more than Intel. Some applications have required more flexibility, so FPGAs from Xilinx and Intel have also made their way into data centers.

Chinese Jingjia Micro, and recently Zhaoxin, are working on GPUs, but at this stage they are low-end laptop/PC offerings that don’t meet the demands of servers. The same can be said for Gowin, Anlogic, Pango, and others doing FPGAs; Chinese players are still far behind the likes of Intel, Xilinx, and Achronix.

READ MORE: China’s first homegrown x86 PCs are here, but don’t get too excited

Sometimes it makes sense to create a chip for a specific purpose and to do that one thing really well. Enter Application-Specific Integrated Circuits (ASICs).

China’s chip sector has proven to hold its own in at least one type of server ASIC: cryptocurrency mining rigs. Bitmain and Canaan are the world’s top producers of crypto mining equipment. This suggests that it is possible for China to lead innovation in at least one kind of server chip.

But many companies have popped up in China looking to ride the AI server ASIC wave in recent years, and none have found great success yet. Like most industries in China, lots of people jump on the bandwagon and make large profits difficult for one another. Many will die, but a few will survive and prosper.

The old ISA conundrum?

While ASICs are probably the best opening, Chinese companies in the server space now are focused on general-purpose CPUs. Companies working on both types of chips have chosen a variety of Instruction Set Architectures (ISAs).

Instruction set architectures (ISAs) are a set of instructions that control communication between software and hardware in processors. They are owned and licensed by western companies, which means Chinese chipmakers rely on deals with IP licensing firms like the UK’s Arm.

Can Chinese companies even begin to make inroads into a market that is 98% x86 architecture, of which almost 90% is Intel and 10% AMD?

It’s difficult, for all the same reasons why Chinese companies can’t wrangle US superiority in semiconductors.

Whether because of luck, economic planning, or market forces, a couple of companies have emerged around each ISA, spreading China’s bets. Huawei and Phytium are using Arm; Zhaoxin and Montage are using x86 (I consider Hygon defunct); Loongson is using MIPS; and Sunway, something else altogether, possibly developed in-house.

Huawei’s Hisilicon has been by far the most successful in the server CPU space. Some Chinese analysts say it may sell 1.5 to 2 million of its Kunpeng server chips this year. Its Taishan server chip might see its market share grow to 3% share globally by the end of 2021. We all know Huawei’s current troubles, so such predictions aren’t exactly watertight.

One way out of the ISA conundrum, as I’ve written before, is using the RISC-V open-source architecture. Huawei and others are jumping on the bandwagon, trying to develop high-performing chips using the free-to-use architecture, and should continue to. It won’t be a fast transition.

READ MORE: China’s chipmakers could use RISC-V to reduce impact of US sanctions

But when it happens, it will remove one key weapon from the US arsenal. The US won’t be able to block Huawei and other Chinese companies from getting their hands on the fundamental architecture.

However, even if one of them created a CPU, based on any of these ISAs, that was superior in power, performance, and area, there are other barriers to entry.

Snatching some of the global market share is not just about having a great performing chip. The software, applications, standards bodies, etc. create an entire ecosystem that can help customers integrate, optimize, and get to market faster.

Huawei has tried to create such an ecosystem by opening up OS source code, compilers, tools, etc. it has done better than any other Chinese company, but still relies on the Arm ecosystem.

Too many cooks

Whether we like it or not China is looking to design and manufacture homegrown chips to replace US imports, and server chips are key to this. The stability and growth in this market means it’s ripe for investment, even if barriers to entry are high.

Making server chips is a long and painful process. But the Chinese companies listed above have identified the cloud as an opportunity and have been investing in it.

One extreme example of trends in China’s server chips industry came from Tencent earlier this year. The Shenzhen company announced it would buy 1 million servers over the next five years, spending around $70 billion.

Domestic demand for server chips will only grow in the coming years. With government preference for domestic chips in this industry and a need for custom accelerators, it could be one of the better semiconductor verticals for Chinese companies to build a customer base in.

However, like many industries in China, there is increasing risk of over-fragmentation, which will make it difficult for everyone to make solid profits. Inspur and Sugon, two leading Chinese server companies, recently set up their own chip divisions, adding to market fragmentation. While I doubt they will start with CPU design as their first foray, it might be coming in later years.

It remains to be seen if Chinese chip makers can compete internationally. But increasing revenues from China will give them better footing to go about global business development, especially in China-friendly countries.

Chinese companies need to pick their fights. Competing in the general-purpose CPU space is an uphill battle. RISC-V could provide a long-term self-reliant option, but dominant players Intel and AMD are strong competition.

Custom ASICs for specific tasks is where China already has plenty of talent and companies that are up to the task. While it would be nice to see extra competition in the GPU and FPGA space, Chinese companies here face the same barriers as those creating general-purpose CPUs.

Most importantly, ecosystems need to be built. It is a difficult and time-consuming endeavor. Even a recent SOE I met with preferred to use Intel, simply because he understood it; it works, it’s mature.

With this mindset, even grabbing the domestic market is a far cry from where we are now. The government needs to step in and use “Made in China” incentives.

China has talented engineers in the ASIC and FPGA design space, but there are simply too many companies for any one of them to have the economies of scale and R&D spend to truly compete. Consolidation and collaboration are needed if Chinese design companies are going to seize the server opportunity.