§ Guide
TOPS per watt is the metric that matters.
Peak TOPS sells modules. TOPS per watt ships products. At the edge, where power and cooling are constrained, efficiency decides what you can actually run.
Edge devices rarely run in a server rack. They run in a sealed enclosure on a pole, in a drone with a power budget measured in watts, or on a battery that has to last a season. In all of those, the question isn't "how many TOPS?" — it's "how much inference per watt, sustained, without a fan?"
Why peak TOPS misleads
- Peak figures are hit briefly, then throttled by thermal or power limits.
- They ignore memory bandwidth, which often bottlenecks real models well below peak.
- They say nothing about the power and cooling your enclosure must provide.
Design for efficiency, then scale
Pick the most efficient part that meets yourTOPS target, confirm it sustains that throughput within your thermal envelope, and leave headroom. TheE1M lineup scales from sub-1 mW always-on perception to 33 TOPS — all NPU-accelerated — so you can size for efficiency first and still grow without changing platform. Power-constrained verticals likedrones andagriculture live on this metric.
Frequently asked questions
- What is TOPS per watt?
- TOPS per watt measures AI throughput (tera-operations per second) per watt of power consumed. It captures efficiency — how much inference you get for your power budget — which is the real constraint in fanless, battery, or solar-powered edge devices. A high TOPS figure means little if reaching it needs active cooling.
- Why does TOPS per watt matter more than TOPS?
- Peak TOPS is achievable only briefly before thermal or power limits throttle it. In a sealed industrial enclosure or a battery device, sustained efficiency decides real performance. Two parts with identical TOPS can differ severalfold in TOPS-per-watt, and the efficient one wins in the field.
- How do NPUs improve TOPS per watt?
- A dedicated Neural Processing Unit runs the multiply-accumulate math of inference far more efficiently than a general CPU or GPU, often delivering several times the TOPS-per-watt. That is why every E1M module pairs Arm CPU cores with an NPU rather than relying on the CPU alone.