Chinese Chips Are Being Artificially Slowed To Dodge US Export Regulations

Once upon a time, countries protected their domestic industries with tariffs on imports. This gave the home side a price advantage over companies operating overseas, but the practice has somewhat fallen out of fashion in the past few decades.

These days, governments are altogether more creative, using fancy export controls to protect their interests. To that end, the United States enacted an export restriction on high-powered computing devices. In response, Chinese designers are attempting to artificially slow their hardware to dodge these rules.

I Got New Rules, I Count ‘Em

Companies like NVIDIA and AMD have had to rework certain products to comply with US regulations against Chinese exports. The A100 datacenter GPU was banned from export, so NVIDIA developed the lower-specced A800 instead. Credit: NVIDIA Press Site

The new export rules come as the US government grapples with the ascendance of China’s military, in both size and technological sophistication. The regulations restrict the export of advanced integrated circuits, but the regulations don’t stop there. The tooling, software, and other manufacturing equipment required to fabricate such hardware is also subject to the rules. The often-stated aim is to slow or halt the development of advanced military devices that could be used by the Chinese government or sold on to other countries. Alternatively, it could be painted as an attempt to safeguard the advantage of existing players in the semiconductor market.

Chips capable of an “aggregate bidirectional transfer rate over all inputs and outputs” exceeding 600 GB/s, not counting to volatile memory, may not be exported or re-exported to China, under the new rules. Advanced manufacturing tools used for electroplating, chemical vapor deposition, and other chip-production processes are similarly banned from export. Just to cover all bases, software packages for the design, manufacturing, or use of these chips or associated hardware is also subject to the sanctions. Companies can apply for a license to export such material to China, however, as with most such restrictions, there is a presumption that such licenses will be denied. Other restrictions apply to chips exceeding certain machine learning performance limits and powerful supercomputers.

Additionally, regarding exports of items not subject to the above restrictions, “US persons” must have a license if the items will be used in the “development” or “production” of ICs in China meeting certain criteria. This includes chips that use a non-planar architecture, or are made at a technology node of 14 nm or less, as well as NAND memory with 128 or more layers, and DRAM made at a node of 18 nm or less. The category of “US persons” is broad, too, including US citizens, permanent residents, and companies and legal entities established in the US, even when operating abroad.

Dodging the Issue

Chinese tech giant Alibaba and smaller startup Biren Technology have since found themselves struggling with the restrictions. Along with a variety of lesser-known Chinese chip firms, they’ve invested heavily in designing new chips for high-powered computing applications. These include new chips to rival GPUs from companies like Nvidia and AMD, along with processors for machine learning applications.

But Alibaba and Biren are fabless, outsourcing the actual production step. Many of these firms have their designs produced by Taiwan Semiconductor Manufacturing (TSMC), considered a world-leading silicon foundry.

Some of the latest designs from these companies are in contravention of the new export rules, in terms of data rates or other factors. While they’re slated for production in Taiwan, the US export regulations nonetheless have an effect. That’s due to the fact that the vast majority of semiconductor fabs around the world rely on US-made equipment and software. If foreign fabs started shipping such designs to China, they would quickly be cut off from US equipment and software necessary to the facility’s work. China is spinning up its own semiconductor production facilities, but they’re presently years or decades behind the cutting-edge and thus can’t produce such advanced designs.

Biren Technology has been touting its new datacenter GPUs as outperforming NVIDIA’s A100 offering. Given the latter is no longer legal to export to China, having a homegrown replacement is key. Credit: Biren Tech News Site

Biren Technology is at risk of running over the limit with its BR100 GPU, which is intended for machine learning applications. Early statements quoted a figure of 640 GB/s, in excess of the stated limit. Since then, the company’s website has listed the card’s bandwith at various figures from 512 GB/s to 448 GB/s. According to some researchers, the company may be disabling parts of the BR100 chip to slide past the limit, while potentially allowing it to be re-enabled later.

Alibaba’s own efforts are facing similar troubles. The company has been working on advanced machine-learning chips for AI work at TSMC’s 5 nm technology node. Reportedly, the team is exploring reworking the designs to avoid issues with the regulations, but this is a costly exercise that would take many months and millions of dollars.

Engineers have complained that the rules aren’t clear cut, as there are various ways to calculate the bidirectional transfer rate. Regardless, many are already working to reduce processor speeds to skirt by the rules. The key is remaining low-key, according to one source speaking to Ars Technica. Some companies have had press materials out in public for chips with transfer rates in excess of the regulations, alerting authorities to monitor shipments of such parts. In cases where a chip’s capabilities aren’t yet widely known, though, engineers have more potential to work with the fab to find a redesign that could bypass the regulations.

Part of a Trend

It’s not the first time that US export regulations have tried to clip the wings of Chinese tech firms. Huawei’s semiconductor arm, HiSilicon, fell afoul of a previous set of export rules in 2019. Initial sanctions were placed on the company due to backdoors allgedly found in Huawei’s communications equipment. These rules cut off Huawei’s access to software and hardware from companies like Intel, Google, and Qualcomm. However, Huawei persevered with its own chips and apps, with Chinese buyers propping up the company’s sales as international business faltered. HiSilicon quickly became the number one supplier of smartphone chipsets in China.

From there, the US government went up the chain, making it illegal to supply HiSilicon with equipment for its semiconductor fabs. That was the death knell for the company’s flagship smartphone chipsets, and it was quickly overtaken by other companies in the market.

Access to the world’s best silicon fabs is key to building high-performance chips. Presently, the US holds the keys to those, thanks to a monopoly on the supply of cutting-edge manufacturing equipment and design software. China will persevere on spinning up its own capacity, in much the same way as it has pursued the production of its own jet engines and other technologies. However, in much the same way, that’s a long, slow road to walk, and a costly one to boot.

Banner image: “Silicon Wafer” by Enrique Jiménez

68 thoughts on “Chinese Chips Are Being Artificially Slowed To Dodge US Export Regulations

  1. Restricting AI-capable technology is good, not because of trying to keep ahead of Chinese tech, but because China is using it to violate human rights of their population, and to oppress minority people groups.

      1. Bad motivations can still lead to good results.

        AFAIK those countries you mention aren’t doing Chinese style monitoring of citizens with AI, so I’m not sure why you’d want or expect a ban on AI tech to them?

        There certainly are controls on exports to those countries.

        1. Chinese are heavy handed. Why put cameras everywhere when everybody already carries a GPS/Video/Audio/NFC bug.

          One Israeli company in particular is known to produce software that turns an android or iphone into an always on data vacuum. They only sell to cops and spooks. Not that Google is any slouch in that area. They also only sell their complete data to cops and spooks.
          The Israelis are the ones you’d want to goto if you wanted incriminating data put onto someone’s phone. But if you were in that position, you’d likely have a staff guy to do it.

      2. Are you suggesting the US government and/or Its civilian corperations are somehow not doing the same? Seems human rights take a side seat when AI is involved… I mean it kinda makes sense that the Human element be the last thing considered when creating a being that is by definition absent of humanity.

    1. While the USA is doing such a great job with its AI? Seems like it’s beibg using to oppress minority groups everywhere. Almost as if the governmental system using AI is at fault, or the AI training data, and not the concept of AI.

      “Hey, we have this cool crime prediction system, it can tell us who will be involved in a gun crime! But it is just as likely to pick the criminal as it picks the victim . . . and putting lots of police to work surveling an innocent person might make the criminals living near-by think they are an informant . . . but who cares, someone will get shot and we can solve the crime!”

      1. What about in the US, the minorities can voice their opinions, recruit others to magnify their voices, and organize to effect change. Is the same possible in China or Saudi Arabia?

        1. Do not mistake the paper-thin aesthetic veneer of political activism we have in this country for a functional democracy. We run the biggest prison system on Earth, the biggest military industrial complex on Earth, have the most expensive healthcare and education systems on Earth, and even still we are constantly strangling public services to shower our sprawling military and police state with new weapons and surveillance systems.

      2. Thank you for sharing summer perspective. Australia is following American style politics, policing and militarization as fast as it can. It’s disturbing. So many get distracted by the media over hyped atrocities abroad, and forget to take a closer look at what’s happening at home. Racist and bigoted news will distract voters and does so very effectively.

      3. The verge story wasn’t really done well. It’s kinda of a half assed follow up to their previous story. It mixes all kind of shit together mostly to puff out the story. There is an episode of Nova that covers this better.

        No “AI” was involved. Just stupid algorithms. Used inappropriately.
        No GPU’s were used.
        No cameras were involved.
        If I recall correctly the data being used were police reports.
        It wasn’t being used for the purpose of oppressing a minority group, suppressing a political movement, religious group etc. Just dealing with homicides. A specific subset of which the potential victims were just as likely to be potential offenders.

        There were a shit ton of problems with the idea and it’s implementation.
        The Mayor and State’s Attorney were voted out. The Police Chief was replaced.
        We voted out the governor and replaced the Attorney General.
        We’ve been busy reforming city ordinances, and passing new state laws to deal with this kind of thing.

        The Mayor, State’s Attorney, Police Chief, County President, Attorney General are all minority’s now. And the lieutenant governor.

        But I understand your point. We should all just pretend that China isn’t doing what it’s doing. It’s not genocide if China does it. If somebody else does something bad some place else, that means it’s ok for china to do worse.

  2. Back in the late 20th Century, the USA restricted sales of weapons to South Africa, as part of a pressure campaign against Apartheid policies.

    The result? South Africa built up their own weapons industry, and ended up being a net exporter of weapons.

    Politics really is the realm of unintended consequences. It’s not hard to see where these restrictions are leading.

  3. It’s a double edged sword here. Because, for now, China struggle to get the technology, so it puts effort to this, at chinese investment speed. Once they’ll have it, they will simply apply the same regulation to you, americans. And from there, who will be the most impacted by the restriction ?

      1. China doesn’t care about patents, just accessibility and cost. Can’t implement a core when you don’t have the design files. So everything was 8051 because the designs were readily available and free.

    1. Appeasement of authoritarian regimes either economically or militarily is never a good idea. China already pumps tens of billions if not hundreds of billions into its high tech industries, regardless of US sanctions. I certainly hope China is severely impacted by our technology restrictions.

      Maybe Hackaday can do an article on alternatives to tech products from China, ie ESP and aliexpress :)

    2. If the PRC elects to go the import substitution route, that’s fine. No point in giving them the cow and milk (in essence) for free. At such point PRC fabs start cranking out GPUs and CPUs that are superior to foreign products, and don’t export them, who cares? The OECD’s military-industrial infrastructures won’t use their stuff in any case for this very reason. PRC industry will have grossly increased their unit cost. It’s unlikely there would be more than a momentary industrial capability cost for OECD members, given that it’s unlikely they would have as much of a technology gap to cross as PRC industry has now.

    1. Ditto! We need collaboration so that everyone can produce the same high-specs chips, like it is done in science, rather than gate-keeping them.

      It is computation. Magnetic and electric fields twiddling around here and there on a piece of metal somewhere in a building. It is just dream at some point. A dream the army is fiercely defending.

    2. *He says from his armchair on a computer network that covers the world*

      Globalization is not a mistake. Corruption has been allowed to run unchecked for decades. DECADES! How about we put some guard rails back that were taken out by the corrupt politicians that are still making large cash-grabs across international conflicts.

  4. Why would limiting the performance of a chip require a redesign? Correct me if I’m wrong, but the specs are just a “promise” that the chip will work reliably within some parameters. Overclocking is a prime example of how these values are not really hard limits.

    So the manufacturer can take a chip that they know could handle a higher throughput, and simply state that they believe it to be reliable up to a clock speed that produces a figure of 599GB/s. Anything higher is overclocking.

    Obviously with chips already on the market, or those whose specs have already been published, it would be difficult to backtrack. Otherwise, the only thing that needs to be redesigned are the datasheets.

    1. That’s exactly how the Cocom restrictions were circumvented decades ago. Most famous example: Intel’s P5, available with 60MHz and 66MHz.
      Export restrictions based on performance values only work on paper but not in real life. If China wants it, they’ll get it, just like the Eastern Bloc did in the cold-war era.

      1. As someone from the Easter block in the 80s and 90s, I can tell you that yes we did get some stuff, but it was gigantic PITA and it did severely restrain the R&D in the block.

        For example local university CS department had exactly one 386 PC. I was lucky enough to gain access for a few hours a day, but you can imagine that did not scale well.

        1. Back at that time I was at the other end of this supply chain. The exporting industry bought all the nice stuff, the actual destination remains unknown – probably government or military.

    2. The restrictions have nothing to do with clock rate. They are all about manufacturing technology.
      The export restrictions are on chip fabrication tech that can produce transistor gates smaller than 20 nanometers.

      For example, go grab an 80’s 6502 processor built at 8um (8000 nanometer) and overclock it into the GHz range.

      Electromagnetic signals can’t travel faster than C, so the only way to get them from point A to B faster is to shorten the trip with smaller components. The regulations are intended to make them spend the time developing their own, instead of writing a check and only spending time waiting on shipping.
      (I make no assertions on the wisdom or potential success of that plan)

      1. Sub 20 nm transistors.
        Is that even a thing yet?

        Last I checked TSMC’s N7 were producing features in the 20-24 nm region. Let alone whole transistors. At least TSMC has stopped naming their nodes “nano meters” and instead just goes with a meaningless “N”. And their N5 and N3 nodes aren’t all that much smaller.

        But the node names and actual features getting manufactured started diverging well over a decade ago.

        Mainly since semiconductor device makers started measuring other aspects than feature size to determine when a “new” generation is established. So when a sufficient improvement is made one slaps on the next step down on the list, even if one’s features aren’t close to that.

        Things that started determining the next generation is by measuring performance (transition times isn’t really governed by feature size), power efficiency (at least somewhat tied to feature size), transistor density (there used to be a lot of emptiness in chips and still is, transistors gets too warm to pack more tightly at some point), among other aspects.

        Improve a few of these and it wouldn’t be “unreasonable” to say that one’s “45nm” node has matured into a “28nm” one. However, existing manufacturers at the “28nm” node will have matured in this time so it wouldn’t be too fair of an assessment in practice.

        But in the end.
        There is currently no sub 20 nm transistors on the market.
        Features might be dropping bellow now, but whole transistors are much larger.

        And as a quick aside.
        The propagation speed of electrical signals in a chip usually aren’t a major limit to clock speeds. Haven’t been a limit for a long time now at the very least. Transition times of output stages tends to be a lot slower, not to mention thermal limits limiting the practical peak operating clock speed.

        And these thermal limits then becomes the clock speed goal. Making logic stages that transition “as fast as possible” just burns extra power due to increasing transient currents necessary to drive the output capacitance, and this extra power just lowers the peak operational clock speed.

        Power dissipation in our wiring is at the square of the current. Halving the transition time doubles the current, that in turn quadruples the power dissipation. So if anything, chip designers should intentionally increase transition times to improve peak operational clock speeds.

        As long as the chip is thermally limited, then the assumption of “Slower logic = Faster chip” is true. Eventually one isn’t thermally limited and the chip will start becoming slower again. And a chip vendor might consider improvements in thermal management to unlock additional future performance, or just use clock boosting to handle serial bottlenecks faster, since Amdahl’s law is a thing. So in the end how much extra speed beyond thermal limits one needs is highly debatable.

        So propagation speed of signals through the conductors becomes rather fast all things considered. Often to the point where taking long detours isn’t problematic if it avoids other downsides.

  5. The Raspberry Pi 4B (BCM2711​) CPU is made with a 28 nm process. A 28 nm process was chosen by the Raspberry Pi foundation(/Broadcom) because it has the lowest cost per transistor (older processes with larger transistors would physical use more space on the silicon wafer which would cost more money per chip, and smaller transistors could only be created with newer machines which cost more to use). The 28 nm process is the “value node” and will probably remain there for at least the next 5 years, probably longer if there is a major global recession. TSMC (Taiwan Semiconductor Manufacturing Company Limited – based in what is officially called the “Republic of China” – ROC) have had a 28 nm process since 2011 and SMIC (Semiconductor Manufacturing International Corporation – based in what is officially called the “People’s Republic of China” – PRC) have had a 28 nm process since 2013. It is not bleeding edge being nearly a decade old, but due to price point most silicon that requires a low price high performance will be build with using 28nm process.

    Silicon chips created using the latest bleeding edge processes will use less power for the same performance (or greater performance for the same power). But there is nothing preventing more silicon chips produced with older processes being used, other than higher carbon emissions.

        1. I have wondered if the processor could be produced with a GaN (Which will decompose into metallic Gallium and nitrogen gas above 650 °C/1202°F at atmospheric pressure) process, how much higher frequency could be pushed.

          1. I don’t really get what for? Fundamentally Moore’s law was driven by CPU power hungry markets with dollars to spend. I haven’t upgraded my PC in years now (don’t ask about the GPU).

            We’re already using stupidly bad languages/libraries (cough, Javascript, cough) that waste 99+% of CPU resources. Suggest teaching assembler to compsci students again, use as weed out.

            If we had terraHz processors, ‘they’ would just grow their libraries into even larger steaming piles. We’d have Javascript renderers used in VR shooters.

            Also a phone with a 300 degree C chip inside? Nope.
            Larger form factors already have active cooling, which can get a lot more active if needed. Raised floors with cooling water were common not that long ago.

          2. @HaHa
            > I don’t really get what for?
            Basically for certain maths intensive tasks to happen faster, sometimes the only way is a CPU that is clocked faster. With GaN, perhaps an order of magnitude higher (or more) clock rate might be possible.

            “Amdahl’s law” from wikipedia “Amdahl’s law is often used in parallel computing to predict the theoretical speedup when using multiple processors. For example, if a program needs 20 hours to complete using a single thread, but a one-hour portion of the program cannot be parallelized, therefore only the remaining 19 hours’ (p = 0.95) execution time can be parallelized, then regardless of how many threads are devoted to a parallelized execution of this program, the minimum execution time cannot be less than one hour. Hence, the theoretical speedup is limited to at most 20 times the single thread performance.”

  6. 600 gigabytes per second? That’s hard to imagine.
    A CPU, clocked at 5GHz, putting 12 bytes in (or out) per clock cycle, is 60 GB/sec. Ten of these would be needed, to break 600GB/sec.

    Maybe a 10 gigabit router would qualify? Not sure how many ports you need, to risk breakiung that limit.

  7. I can’t help but feel that somebody’s missed a point here: all the Chinese are doing is restating the speed at which they believe they will get an acceptable yield, and it’s most unlikely that anybody in the USA will be able to demonstrate that in practice the design is capable of far higher performance since they don’t know how many chips are scrapped as duds.

    Whether the original Chinese spec was attainable, or was merely an optimistic claim to make the company look good to its paymasters, is barely relevant.

  8. They’re missing a trick here, build a well behind cutting edge 15nm plant, use it to lithograph silicon fumed onto a shrinky dink, put it in the oven, voila, instant 5nm tech … ;-)

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.