Skip to content
LiveNext 15:49:53

Definition

What is Total processing performance?

Total processing performance (TPP) is Total processing performance (TPP) is the US export-control measure of an AI chip's raw math speed: twice its peak multiply-accumulate rate times the bit length of the numbers, summed across the chip.

Total processing performance, or TPP, is a single number the US Commerce Department uses to say how powerful an AI chip is. It decides which chips need an export license.

How it is calculated

The Bureau of Industry and Security introduced TPP in an October 2023 rule, replacing an older "bits times TOPS" measure that commenters called ambiguous. The formula is 2 × MacTOPS × bit length, added up over every processing unit on the chip.

Each piece has a plain meaning:

  • MacTOPS is the theoretical peak number of trillions of multiply-accumulate operations the chip can do per second. A multiply-accumulate is the basic move in neural-network math: multiply two numbers, add the result to a running total.
  • Bit length is how many bits the numbers have. Math at 16 bits counts for more than the same math at 8 bits.
  • The 2 reflects the industry habit of counting one multiply-accumulate as two operations on a datasheet.

For chips that run both sparse and dense math, the rule uses the dense figure, so a vendor cannot inflate the score with tricks that only help on mostly-empty data.

Why it matters

Think of TPP as a horsepower rating. Regulators needed one number they could write into law, so a chip is either above the line or below it. Because it measures peak speed on paper, it says nothing about memory, networking or how well software uses the chip. That is why rules built on it often add other tests, such as how much performance is packed into each square millimeter of silicon.

TPP also scales. Rules can add up the TPP of every chip a customer can reach, which turns a per-chip limit into a limit on a whole cluster. That is how a rule can cap the total compute a company gets from an overseas AI data center, even when no single chip crosses a threshold.

What it means if you buy or sell compute

If you build with AI, TPP rarely touches you directly. If you sell cloud capacity, ship servers or resell GPU time across borders, it can decide whether a sale needs paperwork. Thresholds change as rules change, so check the current text of the regulation instead of relying on any figure you remember.

TPP sits inside the wider system of export controls. It is related to, but not the same as, FLOPS, the usual way chipmakers advertise speed, and it applies to every AI accelerator, not only GPUs.

A caution

TPP is a regulatory yardstick, not a benchmark of how good a chip is at training or running models. Two chips with the same TPP can perform very differently on real work, and a chip well under a threshold can still be useful when thousands are combined.

Questions people ask

What does TPP stand for in chip export rules?

TPP stands for total processing performance. The US Bureau of Industry and Security defines it as 2 × MacTOPS × bit length, aggregated over the chip's processing units.

Is TPP the same as FLOPS?

No. FLOPS counts floating-point operations per second. TPP multiplies peak multiply-accumulate speed by the bit length of the operation, and it is used to decide whether a chip needs an export license.

Total processing performance in the news

#05

Microsoft builds a 2.8T-parameter model of its own

Microsoft AI finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, and says it is 4 to 12 times more compute-efficient than DeepSeek's base model. Nothing ships yet. The signal is that Microsoft wants its own frontier model.