· 6 min read · EdgeAI.computer Editorial
Reading TOPS ratings: what edge silicon actually publishes
Published throughput figures use different precisions and sparsity assumptions. Here is what the current generation of edge parts states, and what those numbers do not tell you.
A single headline number sits at the top of almost every edge AI datasheet, and no two vendors define it the same way. NVIDIA rates the Jetson Orin Nano Super at up to 67 TOPS and the Jetson AGX Orin 64GB at up to 275 TOPS, both at INT8 with sparsity. Jetson Thor is quoted differently again, at up to 2,070 TFLOPS at FP4. Hailo publishes 26 TOPS at INT8 for Hailo-8, and for Hailo-10H gives two figures: 40 TOPS at INT4 and 20 TOPS at INT8.
Comparison therefore starts by normalising precision. An INT4 figure is not interchangeable with an INT8 figure, and a sparse figure assumes a model whose weights have been pruned in the way the accelerator expects. Google takes a third approach with Coral, publishing 4 TOPS alongside an efficiency figure of 2 TOPS per watt, which is the more useful number when the constraint is a battery or a sealed enclosure.
Memory then decides what is actually possible. Jetson AGX Orin ships with 64 GB of LPDDR5; Jetson Thor with 128 GB of LPDDR5X; Hailo-10H carries its own 4 GB or 8 GB of LPDDR4. DGX Spark publishes both capacity and bandwidth — 128 GB unified at 273 GB/s — which is the pairing that matters for generative workloads.
Power and temperature are the constraints that quietly eliminate candidates. Jetson AGX Orin is configurable from 15 W to 60 W and Jetson Thor from 40 W to 130 W, while Hailo-10H is rated at 2.5 W typical and offered in industrial and automotive grades from -40 °C to 85 °C. A part that fits your accuracy target but not your enclosure is not a shortlist entry.
Every specification on this platform links to the manufacturer document it came from. Use it to shortlist, then benchmark your own model on the exact build you intend to ship.
Last updated .