Introduction
Background
In recent decades, U.S. semiconductor policy has been primarily market driven & laissez faire
April, 2015
Obama administration refuses Intel a licence to export its Xeon chip for use in Chinese supercomputer centres, such as the National University of Defense Technology (NUDT). Justification = being used for nuclear weapons research.
This narrow export restriction was easily evaded by China, via shell companies. NUDT able to build leading supercomputers using the banned Xeon chips.
Broad examination of Chinese military equipment have found them to be extremely reliant on U.S. chips. For AI, 95% of chip sales in China are Nvidia.
More generally, China’s policy of “military-civil fusion” (meaning there is no distinction between technologies available to companies and to the military) makes it very hard to target export restrictions at exclusively military use.
August 9th, 2022
US signs CHIPS Act into law. Aim = grow US chip industry (37% of world’s chips made in US in 1990 → now 12%). $52B for US chip companies.
September 1st, 2022
Biden administration announced export controls, restricting the sale of high-end chips to 🇨🇳
Criticism:
- May be able to circumvent this, as with the Xeon chip. If not, then…
- In the short term, this policy will significantly harm Chinese AI capabilities.
- However, in longer term it may be good for China because it will strengthen China’s domestic chip design ecosystem. Why?
- Currently, China’s domestic AI chip companies can’t win customers in China because their designs aren’t competitive.
- With fewer customers, these companies can’t benefit from economies of scale.
- New export controls → revenues formerly flowing to U.S. chip companies now go to Chinese ones, offering a viable path to economies of scale.
October 7th
Becomes clear that September announcement was only first stage in a package of wider measures. Further controls are introduced.
These address the points raised above, and aim to stop China developing their a competitive alternative chip industry. This is done by introducing a wider set of restrictions, targeting 4 “chokepoints” the US has over the global semiconductor industry.
The measures are highly aggressive, with the aim of wiping out China’s AI semiconductor industry.
These actions demonstrate an unprecedented degree of U.S. government intervention to not only preserve chokepoint control but also begin a new U.S. policy of actively strangling large segments of the Chinese technology industry—strangling with an intent to kill.
Chokepoint 1: Access to high-end chips
Military-civil fusion
To counter military-civil fusion policy, high-end AI chips can no longer be sold to any entity operating in China.
The Biden administration is essentially saying to China: “If your policy is military-civil fusion, then the only realistic way of implementing our policy of no military end use is to end all sales to China, and we are now willing to take that step.”
What is a “high-end chip”?
Initial reporting suggested that export controls only targeting Nvidia & AMD’s current high-performance AI chips.
Now clear that the rules set a performance threshold - any chip above it requires an export license from the Department of Commerce.
Applications for such a license will face a “presumption of denial” - a de facto ban.
The Department of Commerce document outlines the conditions for a high-end chip on p103.
Some thoughts on the document:
FLOPs
For 8-bit, we have 8 * TFLOP/s = 4800 → 600 TFLOP/s | for 16-bit, 300 TFLOP/s
H100 Tensor Cores provide 1600 - 2000 TFLOP/s of 8-bit compute, 800 - 1000 TFLOP/s of 16-bit
A100 gives 312 TFLOP/s in 16-bit - so presumably this chip is included too
Bow gives 250 TFLOP/s of FP16
Interconnect Bandwidth
(restriction = 600 Gbyte/s or more)
H100 uses new NVLink with 900 GB/sec
Previous gen NVLink used for A100 is 600 GB/sec
… both thresholds are set at or just below the A100 specs
Conditions specifically target machines with high interconnect bandwidth. Purpose = don’t impact other highly parallel use cases such as video games.
Possible consequence - big Chinese push towards low-comms distributed training (e.g. branch-train-merge)
Foreign-direct product rule
The US is also invoking the foreign-direct product rule, meaning the above license requirement doesn’t just apply to US-produced chips or components, but to any product produced by US-regulated technology.
According to one article:
In addition to more traditional export controls, such as country-based licensing requirements and list-based restrictions, the FDPR can cripple its targets’ ability to source critical items produced from US-origin technology.
If foreign companies don’t comply with the FDPR, they risk losing their own access to US technologies.
Consequence of chokepoint 1
Nvidia’s share price, boosted in August by passage of CHIPS act, tanks as investors fear:
- Hit of reduced sales to China
- Heavy investment by China into its own GPU & CUDA alternatives
(are we seeing an uptick in response to the October announcement?)
Chokepoint 2: Access to US-made chip design software and global manufacturers
Electronic design automation (EDA) is a key part of the chip design process.
The three leading companies in the market—Mentor Graphics, Cadence Design Systems, and Synopsys—all have headquarters & the bulk of their staff in the United States.
28 Chinese chip design & supercomputing organisations were already on a US “parties of concern” list which meant such software—or designs made using it—couldn’t be exported to them.
New policy adds extra enforcement powers: now if Chinese companies use such software, their designs won’t be allowed to be manufactured outside of China (foreign-direct product rule).
Additionally, the US is prohibiting semiconductor manufacturers worldwide from doing any Chinese AI chip design work.
Significant limitation, as China has much less advanced semiconductor manufacturing than other countries around the world.
With the new rules, no Chinese chip design company will be allowed to outsource manufacturing abroad for advanced AI and supercomputing chips. For those Chinese organizations on the [”parties of concern”] list, they will be blocked from outsourcing the manufacturing of any types of chips at all.
These enforcement hooks are very similar to those that the United States levied upon Huawei’s chip design subsidiary, HiSilicion, which devastated HiSilicon and forced it out of the market for several years. To state it bluntly, these recent actions are designed to put the roughly 28 Chinese chip design and supercomputing firms on the parties of concern lists out of business.
Oddly, some high-end Chinese AI chip companies, such as Cambricon, are not yet on the list. Adding companies to the list is a (too) slow process. New rules have been introduced to speed up this process.
do not be surprised if Cambricon and every other Chinese chip design company of any significance is added to the backlist in the near future
Even with these steps in place, there would still be a path for China to escape the United States’ semiconductor chokepoints—albeit a complicated, expensive, and risky one. This hypothetical path would be to design the chips in China, pirate older EDA software, and manufacture the chips in China using older manufacturing technology. This “all-Chinese” semiconductor supply chain strategy is in fact a logical extension of the one Huawei announced that it intends to pursue this year with a resurrected HilSilicon business.However, the Biden administration has taken steps to stomp out that approach as well, which brings us to the third and fourth chokeholds of the policy.
Chokepoint 3: access to US-built semiconductor manufacturing equipment
US dominance is such that no advanced fab in the world is not critically dependent on US tech. This tech is deeply integrated (e.g. constant advice & repairs required).
New rules significantly restrict supply of this manufacturing equipment to China - they will have to make their fabs from the ground up.
Previous US policy was to try and keep US semiconductor tech 2 process nodes ahead of China.
Now it appears policy is not only to increase gap, but actively degrade China’s current capability. How?
China now effectively banned from fin field-effect transistor (FinFET) logic manufacturing equipment. This is used for all 16nm and below chips
For comparison, U.S. chipmaker Intel first began mass-producing 14 nm processors in 2014, but the most advanced Chinese logic chip manufacturer, SMIC, only achieved 14 nm FinFET volume production in late 2019, a more than five year gap. The United States is blocking the export of advanced semiconductor manufacturing equipment to all of China, but in the case of SMIC’s 14 nm production facility, the United States is even blocking the export of older, less advanced equipment. In other words, the United States is attempting to put SMIC’s 14 nm production line out of business.
It appears the aim here is to put China back to where the US was in ~2013/2014
China has been stockpiling this equipment in anticipation of something like this. Not clear what impact this will have
Chokepoint 4: access to US-built components
These would be the building blocks of the manufacturing equipment in chokepoint 3.
Of all the semiconductor supply chain chokepoints where the United States has significant leverage over China, this is the one with by far the most durable competitive advantage
Without U.S. components, China’s efforts to develop a domestic semiconductor equipment industry would be starting from scratch and attempting to replicate the cumulative achievements of the U.S. semiconductor industry over the past seven decades. It is an extremely tall mountain to climb.