(High-bandwidth) Memory Persistence: Semiconductors

On December 2, 2024, the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) released a new set of regulations for semiconductor manufacturing equipment (SME) and high-bandwidth memory (HBM) chips. The updates are part of BIS’s ongoing efforts to target the semiconductor sector and aim to slow China’s progress in artificial intelligence. In the race for general artificial intelligence, advanced node semiconductors play a huge role in a country or company’s ability to advance.

However, these rules could have a huge impact on foreign partners in the U.S. semiconductor industry.

1. Overview

Most of the updates come in the interim final rule, which creates two new rules related to the Foreign Direct Products Rule (FDPR) and changes to the minimum rules; new classifications for semiconductor manufacturing equipment (SME) and tools, SME software, and high-bandwidth memory (HBM); some new licensing exceptions; clarification on software keys; and some new red flags. Another part of the rulemaking updates the entity list, adding 140 new names and removing the Verified End User (VEU) program.

There are three key components to AI development: building the AI ​​model itself, training the model using data, and obtaining the hardware performance needed to drive AI functionality. Hardware performance requires critical logic and storage elements. U.S. export controls target each of the components listed above. However, when it comes to hardware components, export controls have been focused on the logical element—the processing unit. (1) so far. The new rules target the memory portion of AI hardware components, capturing more of the physical components required to make AI hardware.

2. New foreign direct product rules

2.1 How the rules work

There are two interrelated updates to FDPR for small and medium-sized businesses. For the uninitiated, the FDPR establishes extraterritorial jurisdiction for BIS over foreign-made products that are direct products of U.S.-origin software or technology. that’s right. Items that are manufactured entirely outside the United States, contain no U.S.-origin ingredients, and do not enter or transit through the United States. However, because that thing is based on a U.S. design, or is a product of U.S.-designed software, or even if that thing is manufactured on U.S.-designed equipment (2) (!), the foreign-made item is treated as U.S.-origin technology, “Foreign direct products” of software or equipment are controlled.

While the basic proposition of FDPR involves a fairly long jurisdictional reach, the new FDPR rules go even further—covering any foreign-made small and medium-sized enterprise that contains integrated circuits (ICs) of U.S. origin. When used in conjunction with the new Red Flag 26 (assuming all SMBs contain US-origin ICs), the new FDPR touches almost all SMBs in the world. all. (3)

The first of these new FDPRs was the SME FDPR, which covered a smaller range of SMEs and broadly restricted exports to China by these foreign-made SMEs. In contrast, the Footnote 5 FDPR covers nearly all other SMEs on the Commerce Control List (CCL), but only restricts exports by these foreign-made SMEs to entities with a Footnote 5 designation on the Entity List. These footnote 5 entities are small and medium-sized enterprise fabs that BIS has identified as participating or about to participate in advanced node IC production. Made consistent changes to minimum rule.

• SME FDPR Targets an enumerated subset of SMEs (ECCN 3B001.a.4, c, d, f.1, f.5, k to n, p.2, p.4, r or 3B002.c) and restricts access to China Exports (Group D: 5 and Macau)
• Footnote 5 FDPR targets almost all other small and medium-sized enterprises (ECCN 3B001 (3B001.a.4, c, d, f.1, f.5, g, h, k to n, p.2, p.4, r except), 3B002 (except 3B002.c), 3B903, 3B991 (3B991.b.2.a Except or 3B991.b.2.b), 3B992, 3B993 or 3B994) and restrict exports to entities specified in footnote 5.

2.2 Impact on foreign small and medium-sized enterprise manufacturers

The breadth of the rule could have a huge impact on non-U.S. small and medium-sized enterprise manufacturers. The assumption made by Red Flag 26—that all SMBs containing ICs contain U.S.-origin ICs—will create substantial new compliance obligations for those SMB manufacturers doing business in China or any other D:5 country. .

First of all, that red flag is actually a red carpet. it states if If the SME contains an IC, the IC will be presumed to be of US origin. . . This begs the question: will the BIS raise any examples of SMEs? no Contains an IC. Next, foreign SME manufacturers selling to D:5 countries will need to overcome this assumption. This will be a difficult task as foreign SME manufacturers need to conduct FDPR and minimum Audit every IC within the small business or otherwise demonstrate that its manufacturing facility does not contain U.S.-sourced technology. Even if this is possible, these foreign SME manufacturers may face some problems with BIS if they continue to do business with banned companies or countries.

Semiconductor fabs that acquire small and medium-sized companies will also feel the impact of the rule through increased compliance efforts by these manufacturers.

3. High-bandwidth memory (HBM) control

3.1 How the rules work

The new rulemaking adds high-bandwidth memory (HBM) controls to the list of commercial controls under ECCN 3A090.c. This additional feature controls all HBMs with a “memory bandwidth density” greater than 2GB/mm2. The control also states that any DRAM with a density that meets the memory bandwidth will also be controlled.

There is an HBM license exception for HBMs with “memory bandwidth density” up to 3.3GB/mm2 for exporters, re-exporters or transferors headquartered in the United States (ultimate parent company is not in D:5 or Macau). However, this really just makes an exception for HBM2, and only applies to US-based companies (some of the largest HBM manufacturers are Korean companies).

Interestingly, ECCN 3A090.c “does not cover co-packaged integrated circuits with HBM and logic integrated circuits where the primary function of the co-packaged integrated circuit is processing.” This could mean that if a chip contains both HBM and logic integrated circuits on a single chip logic IC, the chip will be subject to export controls based on logic components. Therefore, when analyzing on-chip controls, simply examine the chip’s Total Processing Performance (TPP) and Performance Density (PD).

3.2 Impact on Korean semiconductor factories

Most of the world’s HBM is produced by a small group of Korean companies. These companies may bear the brunt of the new rules, especially when working with U.S. companies. These manufacturers are doubly affected by the rulemaking because they may acquire SMEs from foreign SME manufacturers subject to the new FDPR. They may face a host of new compliance obligations, including supply chain reviews and increased end-use verification.

4. Software Key

The new rulemaking also clarifies that the release of software keys and software license keys to unlock software or update existing software is considered an export event. So if the software requires a license, the software key also requires a license. This applies to source code and object code. This does not apply to keys that unlock the “sleep function” in controlled items.

It’s worth noting that this does not eliminate the SaaS and IaaS vulnerability in the current export regime – although we may see attempts to address this vulnerability in the upcoming Artificial Intelligence Services Control Rules. Currently, there is a gap because providing SaaS or IaaS is not considered an outlet so companies can provide access to AI computing ServeExecuted via SaaS or IaaS arrangements from controlled hardware in the United States.

5. Conclusion

The new semiconductor rules reflect the nuanced and targeted approach being taken by the BIS under the Biden administration. However, nuance upon nuance can breed complexity – and there is no doubt that these rules are complex! At the same time, the targeted precision of the rules has created gaps in limiting coverage, leaving some people scratching their heads as to whether the restrictions are intentional.

Are there any other strategies? We may see someone like this in the next administration. Taking a blanket approach – imposing restrictions on every company in China or even all D:5 countries – would alleviate some of the complexity and close loopholes. But the costs to the industry of such complete decoupling could be discouraging. In the shadow of this option, the compliance costs associated with complexity may not look so bad after all.

As we have written before, the balancing act of semiconductor export controls is the hardest regulatory line to walk. The purpose of regulation is to protect U.S. national security, and the ability of semiconductors to help or harm our national security is enormous. But the conundrum regulators face in regulating semiconductors remains the project’s ubiquity. (4) As advanced chips continue to advance, it becomes increasingly difficult to draw a clear line between chips that need to be protected for security purposes and chips that should be freely shared for security purposes. Enhancements to almost every product we use.

Of course, we’ll be updating as new regulations are developed, including one specifically targeting artificial intelligence technology that we’ve heard may be released before the U.S. presidential transition.

footnote

(1) While GPUs are typically used in AI development for their parallel processing capabilities, the regulations will cover any integrated circuit that meets performance thresholds.

(2) The term is actually “plant or main component of a plant”, but that’s a bit wordy even for us, so we use “device” as a shorthand, sacrificing the precision of the essence. . . Sort of.

(3) We don’t even have a self-deprecating footnote to explain our exaggerations. This is indeed more or less the case. All small and medium-sized enterprises!

(4) As a famous saying in the semiconductor field says: “Nothing is more strictly controlled, and nothing is more common.” – Socrates, maybe.