US chip controls pair new performance tests with customer screening

The US has revised its advanced-computing export controls, changing both the technical tests for chips and the customer relationships that can trigger a licence requirement. NVIDIA says the changes cover the A100, A800, H100, H800, L40, L40S and RTX 4090, as well as systems incorporating covered chips.

File photo: US Flag Backlit, ©2009 Joshua Nathanson (Jnn13), via Wikimedia Commons. CC BY-SA 3.0. Cropped/resized; responsive adaptations carry the same licence. Photographed March 12, 2009, not at the announcement.

The package was announced on October 17. The advanced-computing rule filed for Federal Register public inspection on October 18 sets November 17 as its effective date. That differs from the November 16 date in the initial BIS release. The separate Entity List additions took effect on October 17.

Chip performance gets two routes into the main control

BIS is removing interconnect bandwidth as a criterion for identifying restricted advanced-computing chips. The revised 3A090.a classification covers integrated circuits with total processing performance of at least 4,800, or at least 1,600 combined with performance density of at least 5.92.

Those figures use the rule’s technical definitions. Total processing performance is twice the theoretical peak multiply-accumulate rate in trillions per second, multiplied by the largest bit length of the multiplication inputs and added across the chip’s processing units. Density divides that result by the area of logic dies made with non-planar transistor architecture, measured in square millimetres. Where supported bit lengths produce different total-processing-performance results, the highest result applies. For chips supporting both sparse and dense matrices, the calculation uses dense-matrix performance. A headline accelerator throughput figure alone doesn’t describe this test.

A separate 3A090.b category covers lower combinations of performance and density. The rule also excludes from 3A090 chips that aren’t designed or marketed for data centres and have total processing performance below 4,800. Consumer positioning therefore matters, but high-performance consumer chips can still enter the control framework.

A licence exception has its own review path

The new Notified Advanced Computing exception covers specified lower-performance data-centre chips and qualifying non-data-centre chips. It excludes data-centre products meeting the main 3A090.a parameters. For eligible exports and reexports to Macau or Group D:5 destinations, including China, notification is due at least 25 calendar days before shipment. BIS says it will tell applicants within 25 calendar days of a complete notification whether they may use the exception.

That distinction helps explain why a product appearing in NVIDIA’s affected list doesn’t establish that every shipment has the same outcome. Classification, intended market, destination and the conditions of any available exception all matter. Other applicable restrictions remain separate.

The customer connection can cross borders

The revised rules expand destination-based licensing for covered advanced-computing items across Country Groups D:1, D:4 and D:5, excluding destinations also listed in Groups A:5 or A:6. They also reach covered transactions involving entities headquartered in Macau or Group D:5 countries, or with an ultimate parent headquartered there, even when the transaction is elsewhere.

That creates a second screening layer alongside the chip’s specifications. Moving a delivery address to a different market doesn’t by itself settle the licensing question when the relevant customer or parent-company connection remains.

In its October 17 SEC filing, NVIDIA says covered systems include DGX and HGX products containing affected chips. It may seek customer licences, while warning that approval and timing aren’t assured. The company says strong worldwide demand means it doesn’t expect a meaningful near-term financial effect; that remains its forecast. NVIDIA had already distinguished near-term financial expectations from longer-term China-market risks in its August earnings discussion.