Research Note: TOPS Is a Starting Point, Not a Device
Published May 2026
Question
Pocket-scale AI hardware is now marketed with large TOPS numbers. Pockot needs to track those numbers, but it also needs to keep them in their lane. TOPS does not specify memory, model quality, sustained watts, thermals, runtime, software support, or offline usefulness.
Source-Backed Data Points
- Apple states that the M4 Neural Engine is capable of 38 trillion operations per second. Source: Apple M4.
- Qualcomm lists Snapdragon X Elite with an integrated Hexagon NPU at up to 45 TOPS AI performance. Source: Qualcomm Snapdragon X Elite.
- Raspberry Pi documentation lists AI HAT+ variants at 13 TOPS and 26 TOPS, and AI HAT+ 2 at 40 TOPS. Source: Raspberry Pi AI HAT+ documentation.
Reading
The numbers show that useful local acceleration is no longer confined to workstation GPUs. Laptop-class and small-board devices now have public NPU specifications large enough to justify serious measurement. But Pockot should never treat TOPS as an answer. It is an input.
The missing fields are the practical ones: sustained system watts, memory shared with the model, quantization format, runtime support, tokens per second for a specific model, enclosure temperature, and battery curve. A 45 TOPS laptop NPU and a 26 TOPS HAT are not directly comparable without the rest of the stack.
Tool Rule
The feasibility calculator will keep TOPS editable but pair it with sustained watts and model size. Version 1 adds chip profiles as a convenience, not as a capability claim. A real device entry still needs measured watts and measured runtime.