A Chip CEO Calls Out the Industry’s Inflated “Equivalent Compute” Numbers

October 30 — Horizon Robotics founder and CEO Yu Kai posted on Weibo tonight, responding to a question about whether the company could produce an assisted-driving chip with thousands of TOPS of compute — and mocking the way the industry quotes those numbers.

Yu said driving chips are a serious business, but China’s tech industry has “pushed PPT compute to a bottomless limit”. He noted that everyone in the chip industry knows Nvidia’s driving chips are quoted at twice their dense compute: Orin X has 127 TOPS of dense compute, for example, but is quoted at 254 TOPS.

He said other driving-chip companies largely use the same framework so that customers can compare. But the industry has now “broken through cosmic scale”: quoted figures can be inflated to 6–7 times dense compute and labelled “equivalent compute”, he said — “so much that we’re dazzled and can’t keep up”.

So Yu proposed that Horizon’s products carry two figures from now on: dense compute multiplied by two, called “NV compute”, and the other, called “equivalent compute” — or, as he put it, “bragging compute”. By his numbers, Horizon’s Journey 6P has 560 TOPS of NV compute and 1,960 TOPS of equivalent compute.

TOPS means trillions of operations per second. Dense compute is a chip’s peak without exploiting sparsity, while sparse compute doubles the quoted figure by skipping zero-value calculations.

Yu said this “easily completed the task netizens gave me”. One commenter suggested adding a line of small print to the next launch event’s slides: “Equivalent compute is for peers to align with, not for consumers to get excited about.” Yu replied: “Haha, suggestion accepted.”