A chip marketed at 1,000 TOPS might deliver the effective equivalent of 200 to 300 TOPS in real driving conditions. That is the claim a senior NIO executive made in an interview reported by 36Kr, accusing the auto industry of systematically inflating the specifications of smart-driving chips – commonly by three to six times – and proposing four standards the industry should use to compare systems: high memory bandwidth, real computing power, efficient coordination, and stable mass production.
The comments are the most direct attack yet from a Chinese automaker on the marketing war that has defined the smart-driving chip market. Automakers and suppliers have spent the past two years competing on headline TOPS numbers – trillions of operations per second – and NIO is now saying those numbers do not mean what buyers think they mean.
TOPS is not what it seems
TOPS, the unit used to market almost every smart-driving chip, measures a chip’s theoretical peak throughput for a specific kind of math – typically integer operations. The number says nothing about what the chip can actually do inside a vehicle: how much memory bandwidth it has to feed its compute cores, how efficiently its software stack uses the silicon, whether it can run multiple tasks at once, or whether it can do all of that while staying within automotive power and thermal limits.
According to the NIO executive, the gap between advertised and real performance is not a rounding error. Published figures are commonly inflated by three to six times, the executive said. The inflation comes from multiple sources: peak theoretical throughput instead of sustained throughput, sparse-matrix claims that real workloads cannot exploit, and benchmark conditions that do not include the software stack that actually runs in the car.
NIO’s four standards
The executive’s response was not just criticism but a proposed framework. The first standard is high memory bandwidth: a chip that cannot move data fast enough will starve its compute cores, no matter how many TOPS it advertises. The second is real computing power – sustained, measurable throughput under realistic workloads rather than peak theoretical numbers. The third is efficient coordination: the ability of the chip to run perception, planning, and control tasks in parallel without contention, which depends as much on the software architecture as on the silicon. The fourth is stable mass production: a chip that cannot be manufactured at automotive quality and volume is a demo, not a product.
The four standards read like a checklist for procurement, and that is exactly their purpose. NIO has been buying smart-driving chips from suppliers while developing its own in-house silicon, and its executives have a direct interest in forcing the market toward specifications that favor engineering substance over marketing. The company’s in-house Shenji NX9031 chip, announced for its flagship ET9, is NIO’s answer to the suppliers it is now publicly criticizing – and the comparison is more favorable to NIO under its own standards than under the TOPS marketing war.
Why the argument matters beyond the marketing
The stakes are safety and cost. Automakers certify vehicles for sale around autonomous-driving features, and regulators increasingly ask what a system can actually do, not what its chip’s spec sheet says. A driver-assistance system built on a chip that delivers a quarter of its advertised performance may behave differently from how it was tested – a gap that has consequences in the real world. The inflation also distorts procurement: automakers comparing chips on inflated numbers will pay for performance they never receive, and suppliers with honest specifications lose bids to competitors with aggressive ones. If buyers adopt NIO’s four standards as a checklist, suppliers will be forced to publish memory-bandwidth figures, sustained-throughput measurements, and mass-production data alongside the TOPS headline – a disclosure regime that most suppliers have so far avoided.
The debate is not unique to China. Automakers and suppliers worldwide – from Tesla’s own silicon program to NVIDIA’s Thor platform, Qualcomm’s Snapdragon Ride family, and Huawei’s Ascend-based systems – have all played the numbers game at some point. NIO’s complaint is that the game has gone so far that the numbers are no longer comparable across vendors, and it is trying to establish a common vocabulary that would let buyers compare systems on terms that matter.
NIO has picked a fight with the entire smart-driving chip marketing apparatus, and its four standards – memory bandwidth, real compute, coordination, mass-production readiness – are a credible attempt to change how the industry measures itself. Whether the suppliers accept the framework or not, the argument will force every automaker buying chips to ask what the TOPS number actually means. In an industry where the difference between 300 and 1,000 advertised TOPS can be a $500 per-vehicle cost difference, the question was overdue.
