Things Embedded · Performance guide

How much computer does your application need?

Pick what you're building, move the sliders to describe it, and we'll size the processor, AI performance, memory and environment, then match it against every system we supply.

AI performance ladder

Typical requirements

Worked examples from the same sizing rules the finder uses. Headline TOPS figures are the vendor's own (INT8 for NPUs, Hailo and Jetson; FP8/FP4 with sparsity for NVIDIA RTX), so treat them as a guide to the class of hardware, not a guarantee of frame rate.

Application	Workload	AI performance	Fits on

How we estimate

Vision and video AI

Each detection model needs a fixed amount of compute per frame. We multiply it by cameras × frames per second × a resolution factor, then divide by 0.3 because real models reach about 30% of an accelerator's headline TOPS.

TOPS ≈ cameras × fps × GOPs per frame × resolution ÷ 1,000 ÷ 0.3

Light ≈ YOLOv8n (8.7 GOPs), Standard ≈ YOLOv8s (28.6), Heavy ≈ YOLOv8m (78.9), Segmentation ≈ YOLOv8m-seg (110). 4K is tiled, so ×2.5.

Generative AI and LLMs

Language models are limited by memory before compute. The model's weights must fit in GPU or Jetson memory, plus working memory for each user.

Memory ≈ parameters × bits ÷ 8 × 1.15 + users × context

An 8B model at 4-bit needs about 5GB; 70B at 4-bit needs about 40GB, which means a 48GB GPU or a Jetson AGX Thor.

Processor and memory

Control, HMI, gateway and signage loads are sized on CPU cores and RAM. Video recording adds cores for decoding. Every system offering a configuration that meets the need is included, and the ranking favours the smallest one that fits.

Environment rules use our filter data: Extended Temperature means at least −10 to 60°C, Wide means −40 to 85°C. DC power choices must sit inside the system's rated input range; on AC mains every system qualifies, as they ship with or offer a mains adapter.