Four platforms, four different shapes of project
We get asked this every other RFP: "should we use Jetson, Hailo, Coral, or RK3588?". The honest answer is "depends on the project shape", and the project shape questions matter more than the chip's TOPS rating. Here's the framework.NVIDIA Jetson — the ecosystem play
Strengths: CUDA. The entire ML research world ships CUDA-first. If your project depends on a library that wasn't ported (PyCUDA pipelines, certain medical imaging stacks, custom kernels), Jetson is the answer.Weaknesses: cost, power, ARM-Linux quirks. Even an Orin Nano dev kit is 4-6x the cost of an RK3588 board with similar inference throughput.
Use when: CUDA dependency, multi-model heterogeneous workloads, deep PyTorch / TensorRT integration, customer expects "the standard".
Hailo — the throughput-per-watt play
Strengths: inference performance per watt is leadership-class. The dataflow architecture is a meaningful differentiator on convolutional workloads.Weaknesses: proprietary toolchain, smaller ecosystem, M.2 form factor means you need a host CPU board. Custom or unusual operators are harder.
Use when: convolutional vision pipeline, low-power requirement, fanless enclosure, throughput-critical deployment. Hailo + a small ARM board (CM4, x86 NUC) is a common pattern.
Google Coral — the cheapest entry
Strengths: cheap (~$60 for the USB stick), well-documented, quantizes cleanly for many MobileNet-style models, low power.Weaknesses: limited operator support relative to Hailo or Jetson. The Edge TPU is INT8-only, fixed precision. Throughput ceiling is real (~4 TOPS).
Use when: hobbyist or low-volume project, MobileNet-class model, hard cost ceiling, no need for multi-model. We rarely ship Coral in industrial deployments anymore — Hailo dominates the same niche with more headroom.
Rockchip RK3588 — the dark horse
Strengths: cheap (~$100-200 for a board), built-in 6 TOPS NPU, eight ARM cores, full Linux distribution support. The bang-for-buck winner in 2026 for budget-conscious projects.Weaknesses: toolchain (RKNN) is workable but less mature than Hailo or NVIDIA. Operator support has gaps. NPU performance is real but not best-in-class.
Use when: budget is the binding constraint, model fits in the operator support matrix, customer is okay with a less-mature toolchain, deployment doesn't need bleeding-edge performance.
The decision matrix we actually use
- Customer says "we want Jetson" → Jetson. Don't fight on hardware brand, fight on requirements.
- Throughput / watt is the main constraint → Hailo
- Single-model, fixed budget, hobbyist context → Coral
- Cost ceiling is hard, model is standard CNN → RK3588
- Multi-model, custom kernels, research-track → Jetson
- Multi-camera, single model, industrial cell → Hailo (8 or 15 depending on model size)
One non-obvious thing
The chip's TOPS rating is the marketing number. The real performance number is "throughput on your specific model, after quantization, end-to-end". We always benchmark the candidate model on candidate hardware before promising performance numbers to the customer. A YOLOv8s might run at 60 fps on Hailo-8 and 35 fps on Orin Nano even though the Nano has more nominal TOPS — because the architecture matters more than the spec sheet.What's everyone shipping in 2026? Curious whether anyone has displaced Jetson with RK3588 in serious deployments.
