TPP
How it works
For each operation type, 2 x MacTOPS x the bit length of the operation is computed, and the results are summed over all processing units on the chip to yield TPP. 'Performance density' = TPP / applicable die area (mm^2), counting logic dies made with a non-planar transistor process. The 2023 rule controls chips (3A090.a) when: (1) TPP is 4800 or more, or (2) TPP is 1600 or more with a performance density of 5.92 or more; ECCN 3A090.b covers lower tiers (e.g., TPP 2400-4800 with performance density 1.6-5.92). Adding performance density prevents the workaround of buying many smaller chips that are collectively as powerful as a restricted one.
Problem solved
Export control of advanced AI chips needs an objective, measurable performance boundary rather than a list of product names that is easy to circumvent. TPP (with performance density) provides a parameter harder to bypass via minor modifications or by combining smaller chips.
Evolution
In the October 7-13, 2022 rule, BIS introduced TPP as an objective parameter for controlling advanced computing chips.
The October 25, 2023 rule (effective November 17, 2023) added a performance density parameter and the new 3A090.a/.b structure to prevent workarounds.
The December 2024 rules left the TPP and performance density thresholds unchanged from the October 2023 controls.