GPU & AI · Edge inference
An edge AI accelerator is a small chip that does a trained model's arithmetic, so your processor does not have to. Add a Hailo-8 or a Google Coral Edge TPU to a dedicated server of your own, and the video frames, the model and the answers all stay on one machine.
It runs compact vision models fast and on very little power. It does not run language models, and we say so before you order.
TOPS figures are Google's and Hailo's own. Prices are read from our order catalogue.
01Before you order
These modules were made for one kind of job: looking at images and video and saying what is in them. Point them at that job and they are quick and cheap to run. Point them at anything else and you will be disappointed.
Outgrown the module? The next step is a full GPU in a server built for one. Every GPU we fit is priced beside its machine on GPU dedicated servers.
02Choosing the module
For a new project, pick the Hailo-8. It is rated at more than six times the Coral's throughput, it takes models from more frameworks, and its maker still ships new runtime releases for it. Choose the Coral when you already have models compiled for the Edge TPU and want them to keep working.
| Hailo-8 | Google Coral Edge TPU | |
|---|---|---|
| Rated throughput | up to 26 TOPS int8 | 4 TOPS int8 |
| Power, per the maker | No figure published on the module page | About 2 TOPS per watt, so around two watts flat out |
| Models from | PyTorch, ONNX, Keras and TensorFlow, including TensorFlow Lite | TensorFlow Lite, quantized to int8 |
| Compiled with | Hailo Dataflow Compiler | Edge TPU Compiler |
| Runtime on your server | HailoRT 4.x, from the hailo8 branch | libedgetpu with the PCIe driver |
| Software in 2026 | Maintained: HailoRT 4.24.0 came out in June 2026 | Google archived the repositories; community builds keep it running |
| Added to your server | $13.99/mo | $13.99/mo |
The main branch of HailoRT and its 5.x releases are for the newer Hailo-10 and Hailo-15. The Hailo-8 runs on the 4.x line from the hailo8 branch. Install a 5.x package and the module will not be found.
Google has made its libedgetpu, PyCoral and Edge TPU repositories read-only on GitHub. Community forks still build the runtime and the PCIe driver for current kernels and TensorFlow releases, so a working Coral setup stays workable. Expect no new features from Google.
03The servers
The accelerator goes into the M.2 slot of an HPE Moonshot cartridge, a compact server with its own processor, memory and NVMe drive. Pick the cartridge for the work around the model, then add the module as one line in its configurator.
AVideo and camera work
A quad-core Intel Xeon E3-1585L v5 with Intel Iris Pro P580 graphics, whose Quick Sync engine can decode the video so the cores stay free. The natural home for camera analytics.
Intel Xeon E3-1585L v5 Quad Core 3.00 Ghz · 32 GB DDR4 · 1 × NVMe-SSD 240 GB · 1 Gbps
BMore cores around the model
An Intel Xeon D with 8 or 16 cores and room for more memory. Pick it when the code around the model is the bigger job: tracking, databases, many streams at once.
Intel Xeon D-1548 Octa Core 2.00 GHz · 32 GB DDR4 · 1 × NVMe-SSD 240 GB · 1 Gbps
Both servers come in New York, Miami, San Francisco, Amsterdam and Bucharest. Orders in Amsterdam and Bucharest are billed in euros, with the same figures.
Each includes a 1 Gbps port that nothing meters and no other customer shares at the access switch, root access, and a console you open from the control panel if the network setup goes wrong.
Billing is monthly with no contract. You see the full monthly total in the configurator before paying; where US sales tax applies, checkout adds it.
04Getting started
Open the m710x or the m510, choose memory, storage, city and operating system, then pick the module under Tensor Processing Unit. One module per server.
A machine with a module added is a custom build, delivered in 4 to 24 hours. Our installer finds the module on the PCI bus and adds it to the server's hardware record, so our support can see it is there.
Install the maker's driver and runtime on your system, then load your compiled model. The module shows up on the PCI bus straight away.
# Hailo-8 shows as 1e60:2864, the Coral Edge TPU as 1ac1:089a
lspci -nn | grep -E '1e60:2864|1ac1:089a'Stuck on a driver? Our AI support on live chat answers in seconds and can look at your server and account while you talk. Ask for a person at any point and one of our team takes over.
05Questions
A TPU server is a computer with a tensor processing unit, a chip built only for the math inside neural networks. Here that means a dedicated server with a Google Coral Edge TPU in its M.2 slot. We also fit the Hailo-8, which does the same job and is rated faster.
No. Cloud TPUs are large chips in Google's own data centers, and only Google Cloud rents them. What we fit is the Edge TPU, Google's small inference chip, on a server you have to yourself.
No. A language model needs gigabytes of fast memory even when quantized hard, and both modules are built around a few megabytes on the chip. For a language model you need a GPU: what it takes to run an LLM works out the size, and GPU dedicated servers lists every GPU we fit with its price.
On paper, yes. Hailo's own figure for the Hailo-8 is up to 26 TOPS; Google's for the Edge TPU is 4 TOPS, both counted on 8-bit integers. Real speed depends on your model and how well it compiles, so test your own network before you plan capacity around either figure.
No. Each cartridge takes one module. For more inference, order several cartridges and spread the camera feeds across them; each is a separate server with its own port and its own address.
Not as one product. What you can do is order as many accelerator cartridges as you need in the same city. When the job needs more than small modules can give, several GPUs in one machine or across machines are on multi-GPU servers.
A recent Linux. Ubuntu is the default build in the configurator, and both makers publish their drivers for Linux first. Hailo also supports Windows. Install the runtime yourself after hand-over, at the version your model was compiled for, and it stays at that version until you change it.
Run lspci after you log in: look for 1e60:2864 (Hailo-8) or 1ac1:089a (Coral). Our installer checks the same PCI bus and adds the module to the server's hardware list, which our support can read if you ever need help.
06Somewhere else to look
Choose the cartridge, add a Hailo-8 or a Coral, and the server is yours in 4 to 24 hours.