The Chinese CPU design firm Loongson wants to raise $544 million in an IPO as part of a bid to raise funds and further the development of its next-generation products. Loongson is a company we’ve discussed several times over the last few years. Its CPUs use an instruction set based on MIPS64 and it has steadily released more powerful CPUs.
Earlier this year, Loongson announced a new instruction set architecture, LoongISA 2.0. According to a post on zhihu.com (it’s described as a question and answer website and it’s formatted a bit like Medium), Loongson has spent the past decade steadily building to this moment. Beginning with the Loongson 2E/2F, the company began to introduce its own custom instructions for MMX as well as SIMD instructions more generally. I apologize for any errors on my part, as Google Translate makes some clear mistakes when translating the original post.
Loongson has continued to add new instructions to its CPU design. It formalized Loongson 1.0 with the release of the Loongson 3A2000 in 2015. At that time, a Loongson official reported that 1,014 LoongSIMD instructions had been added to the CPU, in addition to 148 general-purpose extension instructions and 213 instructions intended to improve the execution of ARM and x86-native binaries.
LoongISA 2.0 was announced earlier this year, with reportedly nearly 2,000 instructions implemented in total. While this sounds like a lot, I wanted to see how x86 would compare.
Here’s a mildly interesting fact I learned as a result: Because x86 supports instructions of variable length up to 15 bytes, the number of possible x86 instructions ranges between 981 and ? depending on what you count as an instruction.
Loongson has not been terribly forthcoming with details about its processors and has not released instruction manuals, but the company has claimed that it was able to improve x86 and ARM binary translation performance a great deal by adding an additional 213 instructions, while paying less than a 5 percent die penalty.
Reading this article, one gets the distinct impression that Loongson is sabotaging itself by refusing to release any kind of data on its processors. It is apparently quite difficult to optimize for the chips because virtually nothing is known about their internal structures or best practices.
Here’s another post, this one discussing Loongson’s relatively weak market uptake. Again, I apologize if there are errors in Google Translate:
For a long time, the R&D and production of general-purpose CPUs have been in the hands of foreign IT giants Intel and AMD…. Loongson has been carrying the mission of researching and developing domestically-made CPUs since its establishment. Chinese people have always placed high hopes on them. They hope that in the near future, they can use the CPUs independently produced and developed by the country, so that the core technology will no longer be controlled by others… [A]fter so many years, Godson still exists only in newspapers and media reports, and has not yet entered the homes of ordinary people… netizens are everywhere questioning the voice of Godson.
The author attributes Loongson’s low market penetration to a lack of domestic software support as opposed to any fundamental weakness in the processors. According to him, many popular (Chinese) domestic software manufacturers continue to prioritize Windows instead of Linux.
It’s interesting that Loongson has continued to add instructions as a means of improving performance. The upcoming 3A5000 is expected to be the company’s last MIPS64 core before it pivots to building RISC-V CPUs. Right now, Intel and AMD account for something like 99.5 percent of the Chinese CPU market, but the Chinese government has supported Loongson’s work for years as part of its effort to make China less reliant on foreign chip manufacturing. Asia Nikkei writes: “Part of computers used by Chinese customs that handle sensitive tasks are already equipped with domestic CPUs and operating systems, a customs official in South China’s Guangdong Province told Caixin. Their performance isn’t great, the official said.”
This is probably unsurprising considering that Loongson’s current efforts are built on 28nm and clocked around 2GHz. Still, the company’s designs have evolved rapidly in the past six years. It’s an opportune time for a pivot to RISC-V, given the interest that ISA has drawn. A $500M IPO may not be much compared with Intel or AMD, but it’s not exactly nothing, either. We’re curious to see where the company goes from here, especially since it’s not currently subject to US trade restrictions like some other Chinese technology firms are. Loongson’s CPUs may not pose a near-term performance risk to Intel or AMD’s business, but it’s always interesting to see details of what other CPU design firms are bringing to the table. Being forced to work with limited resources can sometimes drive research breakthroughs.
Feature image by Windows 1089, CC BY-SA 3.0