Skip to content

§ Guide

How many TOPS do you actually need?

Datasheet TOPS is a marketing number. What you need is set by your model, your resolution, and your frame rate. Here's a practical way to size it.

Rule of thumb:operations-per-inference × frames-per-second × streams, plus ~30–50% headroom. Match that to a module's sustained throughput — not its peak — because utilisation and memory bandwidth, not the headline figure, decide real frame rate.

Sizing table

WorkloadTypical bandExample E1M module
Keyword spotting / always-on audio< 0.5 TOPSE1M-AEN
Single-camera classification, 720p @ 15 FPS1–2 TOPSE1M-AEN / E1M-X V2N
Object detection (YOLO-class), 1080p @ 30 FPS2–8 TOPSE1M-X V2N / V2H
Multi-stream detection or segmentation10–30 TOPSE1M-X V2N+M1 / V2H+M1

Bands are starting estimates — always confirm with a benchmark on target hardware.

Then check power

Once you have a TOPS band, the next constraint is efficiency: a high TOPS number is useless if it needs a fan your enclosure can't fit. ReadTOPS per watt, then browse theE1M lineup (0.25–33 TOPS) to find the fit.

Frequently asked questions

How many TOPS do I need for object detection?
For single-stream 720p–1080p object detection (YOLO-class) at 15–30 FPS, plan on roughly 2–8 TOPS. Multiple cameras or higher resolutions push you toward 10–30 TOPS. Frame rate, input resolution, and model size drive the number far more than the headline TOPS figure.
Is more TOPS always better?
No. Beyond what your model and frame rate require, extra TOPS mostly burns power and budget. Size to the workload with headroom for growth, then verify on real hardware — utilisation and memory bandwidth often matter more than peak TOPS.
How do I estimate TOPS from my model?
Take the model’s operations-per-inference (GFLOPs/GOPs), multiply by target frames per second and number of streams, then add ~30–50% headroom for pre/post-processing and real-world utilisation. Match the result to a module’s sustained (not peak) throughput.

Ready to build on an open edge AI standard?