AI Accelerator Sticks Compared: Hailo, Coral, Jetson

You Don't Need a New Machine, You Need an NPU

The most common question in local AI right now is some version of "do I have to buy a new box?" Usually the answer is no. A Raspberry Pi 5, an aging mini PC, an Intel NUC in a cupboard, or a home server doing nothing but Plex can all be given real inference capability with an add-on that costs less than dinner for four. The trick is picking the right one, because the AI accelerator market is full of parts that look comparable on a spec sheet and are not remotely interchangeable in practice.

This guide covers add-on accelerators — USB sticks, M.2 modules, and HATs that attach to hardware you already own. If you're choosing a whole new machine instead, our best mini PC for local LLMs guide covers that decision, and our edge AI dev board buyer's guide covers boards with the NPU already integrated.

• Most builds should start with the silicon they already have — modern Intel iGPUs and NPUs handle a surprising amount via OpenVINO before you spend anything
• Hailo-8L (~$70, 13 TOPS) is the current value sweet spot for vision workloads on a Raspberry Pi 5 or M.2-equipped host
• Hailo-8 (~$110-200, 26 TOPS) is the step up for high stream counts and heavier per-frame enrichment
• Jetson Orin Nano (from $249, ~40 TOPS claimed) is a different category — a full CUDA computer, not an add-on

Quick Picks

• Best first purchase for a Pi 5 camera or NVR build: Hailo-8L M.2 at roughly $70 and about 1.5W
• Best for many camera streams or added classification passes: Hailo-8 at 26 TOPS
• Best if you need CUDA and the NVIDIA tooling stack: Jetson Orin Nano, from $249 — but budget for it being the whole computer
• Best for language models rather than vision: an accelerator with its own memory, such as the Hailo-10H-based ASUS UGen300 with 8GB onboard
• Best value move overall: try your existing iGPU or NPU first and only buy once you've measured a real bottleneck
• Google Coral USB (4 TOPS, ~$60): still works, still widely documented, but no longer the default recommendation for new builds — the Frigate project now advises against it for fresh installs

What Do TOPS Actually Tell You?

Less than the marketing implies, and it is worth understanding why before comparing any two numbers.

TOPS means trillions of operations per second, but the figure is always quoted at a specific numeric precision. A part rated 40 TOPS at INT4 might deliver 20 TOPS at INT8, and single-digit throughput at higher precision. Two accelerators advertising the same TOPS can differ by 2x on the same workload simply because one is quoting a precision your model doesn't use.

Three things matter more than the headline number:

1. Precision support. Vision models typically run INT8. Quantized language models increasingly run INT4. Check that the rated figure applies to the format you'll actually deploy.
2. Onboard memory. An accelerator with no memory of its own streams weights across the bus from the host. That is fine for a small detector and painful for a language model. Parts with dedicated LPDDR — like the Hailo-10H modules with 8GB — can hold a model resident.
3. Toolchain maturity. A model has to be compiled for the target. If your architecture isn't in the vendor's model zoo and the conversion tooling is thin, raw throughput is irrelevant because you'll never get your model onto the part.

Hailo-8L vs Hailo-8: Which One Should You Buy?

The 8L at around $70 and 13 TOPS is the one most people should buy, and the 8 at 26 TOPS is the one that justifies itself only under specific load.

The practical dividing line in the home-NVR community — the largest single body of real-world testing on these parts — sits around camera count. Builds up to roughly four cameras are usually fine on host silicon alone. Six to ten cameras is where the 8L has become the standard recommendation. Beyond that, or when you're running additional passes per frame such as face or license plate recognition on top of object detection, the 26 TOPS part starts earning its price difference.

Power is worth noting too: the 8L draws about 1.5W. That is low enough to sit inside a fanless case or hang off a Pi 5 without a thermal plan, which is a meaningful part of why it has become the default.

Where Does Google Coral Fit in 2026?

Honourably, and mostly in existing deployments. The Coral USB Accelerator delivers 4 TOPS for around $60 and was the part that made edge inference approachable for a generation of hobbyist projects. There is more documentation, more sample code, and more forum history for Coral than for anything that came after it.

What changed is the software surface. Coral is tied to TensorFlow Lite with a restricted set of supported operations, which means models have to be built for it rather than merely converted to it. Newer accelerators accept ONNX and PyTorch exports with far less ceremony. The Frigate NVR project — one of the largest consumers of these devices — no longer recommends Coral for new installations, which is about as clear a signal as this category produces.

If you own one, keep using it. If you're buying today for a new build, the money goes further elsewhere.

When Is a Jetson the Right Answer Instead?

When you need CUDA, and you should be honest with yourself about whether you do.

A Jetson Orin Nano starting at $249 is not an accelerator you attach to something — it is a complete Arm computer with an NVIDIA GPU, its own operating system, and access to the CUDA ecosystem. That last part is the entire argument. If your workflow depends on libraries, custom kernels, or research code that assumes CUDA, no amount of TOPS on a Hailo part will help you, because the code simply won't run.

If your workflow is "run this detection model on these camera streams," CUDA buys you nothing and you're paying three to four times as much for compatibility you won't use. Modules like the Jetson T3000 family push much higher performance for robotics and multi-sensor work, and rugged Jetson boxes take that into the field — but all of them are host computers, not add-ons.

How Do You Choose Without Guessing?

Measure before you buy. The sequence that wastes the least money:

1. Benchmark what you have. Run your actual model on the host's CPU, and then on its iGPU or NPU through OpenVINO if it's Intel silicon. Record frames per second, or tokens per second, on your real workload — not a demo.
2. Identify the bottleneck. If you're decode-bound on video rather than inference-bound, an accelerator won't help. Fix the decode path first.
3. Size for headroom, not for peak. Sustained load matters more than burst. A part at 60% utilization stays cool and leaves room for the extra model you'll inevitably want to add.
4. Check the model zoo before the price. Confirm your specific architecture has a supported conversion path on the part you're considering.
5. Prefer M.2 when the host has a slot. It's cheaper than USB for the same silicon and thermally better. Choose USB when you need the accelerator to move between machines.

What About Running Language Models?

That is a different purchase, and conflating it with vision acceleration is the most common expensive mistake in this category.

Vision accelerators are optimized for small models processing a stream of frames. Language models are large weights processing a stream of tokens, and the binding constraint is memory bandwidth and capacity, not raw operations. An accelerator with 4 TOPS and no onboard RAM will not run a 7B model no matter how the arithmetic looks.

What you want instead is a part with its own dedicated memory — the current crop of Hailo-10H-based modules ship with 8GB, enough to hold a quantized small model resident. Our LLM quantization guide covers how much a given model actually needs in each format, which is the number to size against. For anything above roughly 13B parameters, an add-on stick is the wrong tool and a unified-memory machine is the right one.

The Short Version

Start with what you own. Buy a Hailo-8L when you've measured a real vision bottleneck. Step up to a Hailo-8 for high stream counts. Buy a Jetson only when CUDA is a requirement rather than a preference. And if the goal is running language models locally rather than watching cameras, look at accelerators with onboard memory — or skip the add-on entirely and read our mini computer coverage on unified-memory machines.

Prices in this category move quickly and new parts land every few weeks, so treat the figures here as the shape of the market rather than a quote. The decision framework outlives the price list.

Sources: EdgeAIStack hardware comparison — 2026; Botmonster Tech, Hailo-8 vs Coral for Frigate — 2026; CNX Software on the Hailo-10H-based ASUS UGen300 — August 3, 2026; NVIDIA Jetson Orin Nano — 2026.