The threshold is rewritten, the workaround goes

In October 2023 the United States rewrote the accelerator threshold. The chip-to-chip bandwidth criterion was removed; what remains is total processing performance of 4800 or more, plus a new quantity, performance density - that same performance divided by die area. The reach was extended to Country Groups D:1, D:4 and D:5.

The string "600 Gbyte/s" appears nowhere in the rule's text: the bandwidth criterion is gone from ECCN 3A090 entirely. The new wording is total processing performance of 4800 or more, or else of 1600 or more together with a performance density of 5.92 or more. A second tier, 3A090.b, is added at 2400 to 4800 with density 1.6 to 5.92, or 1600 or more with density 3.2 to 5.92. Total processing performance is defined as twice the multiply-accumulate operations per second times the bit length of the operation; performance density is that divided by the area of the logic dies. The rule states the reason for the new parameter outright: it "prevents the workaround of simply purchasing a larger number of smaller datacenter AI chips". There are four dates. On 17 October 2023 the government announced it had submitted the rule for publication. The document is dated 18 October. The Federal Register carried it on 25 October, generally effective 17 November. But on 23 October the government told NVIDIA that for products with a total processing performance of 4800 or more designed or marketed for datacenters the requirement was effective immediately: the thirty-day wait was cancelled. The rule itself names no product, only numbers. The list comes from NVIDIA, which on 17 October named the A100, A800, H100, H800, L40, L40S and GeForce RTX 4090, plus the DGX and HGX systems containing them. The countries come from there too: China and Country Groups D:1, D:4 and D:5, among them Saudi Arabia, the United Arab Emirates and Vietnam, but not Israel. The requirement also follows the buyer whose ultimate parent is headquartered in Group D:5.