Tera Operations per Second (TOPS)
Tera operations per second (TOPS) is a peak throughput rating equal to one trillion arithmetic operations per second under specified conditions.
What counts as an operation
TOPS is commonly used for neural-network accelerators in automotive computers. It describes how many low-precision operations the hardware can theoretically issue, but the number depends on the counting convention. A multiply-accumulate calculation may be counted as two operations, and results for INT8, INT4, FP8, or other data types are not equivalent.
Some ratings also assume structured sparsity, where known zero values allow parts of a calculation to be skipped. A dense and a sparse TOPS figure therefore describe different conditions. The data type, sparsity rule, clock, and system configuration must accompany the headline number.
What it means in an EV
ADAS computers use acceleration for tasks such as object detection, lane interpretation, free-space estimation, occupant sensing, and combining several perception models. Higher peak throughput can provide room for larger models or more concurrent work, but only if the rest of the platform supplies data and executes the software efficiently. ADAS processor
Useful vehicle performance also depends on memory bandwidth, sensor input rates, latency, scheduling, software optimization, thermal limits, and power consumption. Safety mechanisms, redundancy, and fallback can consume resources without increasing the advertised feature list.
Why TOPS numbers are hard to compare
Two processors with the same TOPS rating can produce different inference times for the same model. One may spend more time moving data, support a different numerical precision, or achieve a lower percentage of its theoretical peak. Conversely, a processor with fewer TOPS may run a particular optimized network faster or with less energy.
TOPS is therefore a compute-capacity indicator, not a measurement of driving intelligence, feature quality, or vehicle safety. Meaningful comparison requires the same workload, precision, accuracy target, latency method, and power boundary. For the wider sensing and compute context, see Sensors for automated driving.