GPU and AI · Edge inference

TPU server on bare metal.A Coral Edge TPU or a Hailo-8, in a Xeon machine that is yours alone.

A TPU server at Server Room is an HPE Moonshot cartridge with a Google Coral Edge TPU or a Hailo-8 module on its M.2 slot, rented by the month with root access. It runs int8 vision models beside your data. It is not a Google Cloud TPU, which only Google rents, and it does not train or serve large language models.

One tenant per machine
M.2 option on m710x and m510
Root and remote console
Own network, AS19624

Choose the host cartridge, tick the module in the configurator, then install the vendor runtime on the system you picked. Hailo rates the Hailo-8 at up to 26 TOPS and keeps its toolchain current. Google rates the Coral at 4 TOPS, and it suits pipelines already built for the Edge TPU.

Compare the two modules

Definition

What a TPU server is, and what it is not

TPU stands for tensor processing unit, the name Google gave its own matrix accelerators. Two very different parts carry the name. Cloud TPUs are large chips for training and serving models, and Google rents them inside Google Cloud only. The Edge TPU is the small one: about two watts, integers only, built to run a compiled vision model next to a camera or a sensor feed.

We rent the small kind and its closest rival, the Hailo-8, as an M.2 module inside a dedicated server. Both are edge AI accelerators: they run a neural network that has already been trained and quantized to int8, and they do nothing else.

Sized for

Object detection on video frames, image classification, segmentation and pose estimation. Networks from the MobileNet, EfficientDet-Lite and YOLO families, quantized to int8 and compiled for the module, score a frame without the data leaving your machine.

Not sized for

Training of any kind. Language models, at any quantization: neither module can hold billions of parameters. Float32 models that will not quantize, and layers the vendor compiler does not support, which fall back to the processor or fail to compile.

Where those go instead

Sizing a language model is the job of LLM hosting. Training and large inference run on GPU dedicated servers, with a dedicated GPU per machine. An NPU and a large pool of memory in one box is Ryzen AI servers.

The two M.2 modules

Hailo-8 or Coral Edge TPU

Both modules take an int8 network compiled ahead of time by the vendor's own compiler. They differ in rated throughput, in the frameworks they accept and, since Google archived the Coral software, in who still maintains the stack.

Hailo-8 M.2 AI acceleration module, top view

Maintained toolchain

Hailo-8 M.2 module

Rating (Hailo's): up to 26 TOPS, int8
Form factor: M.2 Key M, 22 × 42 mm, extends to 2280
Interface: PCIe Gen 3
Model sources: TensorFlow, TensorFlow Lite, ONNX, Keras, PyTorch
Toolchain: Dataflow Compiler, HailoRT, Model Zoo
Runtime for Hailo-8: HailoRT 4.x, hailo8 branch (4.24.0, June 2026)
Option price: from $13.99/mo

The Dataflow Compiler quantizes your network and compiles it to a HEF file, and HailoRT runs that file from C++ or Python. Use the hailo8 branch: HailoRT 5.x and the master branch support only the Hailo-10 and Hailo-15.

Archived by Google

Google Coral M.2 Accelerator

Rating (Google's): 4 TOPS, int8, at 2 TOPS per watt
Processor: Google Edge TPU
Model sources: TensorFlow Lite, int8, through the Edge TPU Compiler
Runtime: libedgetpu and PyCoral, last PyCoral release 2.0.0 (July 2021)
Maintenance: Google repositories archived; community forks build it today
Option price: from $13.99/mo

A model either compiles for the Edge TPU or it does not, and the established vision models do. Google has archived libedgetpu, PyCoral and the Edge TPU repositories, so current kernels and Python versions rely on community forks of the runtime and the PCIe driver. Choose it when your models already run on an Edge TPU. For a new project, start on the Hailo-8.

Throughput figures are each maker's own rating. We publish no benchmark of ours: frames per second depend on your model, its input size and the batch.

Data path

How the offload works inside one machine

The Xeon still decodes the video, resizes each frame and runs your application. The module takes the neural network and nothing else, so the cores stay free for the work around it. On the m710x, the processor's Iris Pro P580 graphics can decode video in hardware with Quick Sync, which leaves its four cores for the rest.

Live from the order catalogue

Two host cartridges, priced as the checkout prices them

The module is an option line on two HPE Moonshot cartridges: the m710x with a quad-core Xeon E3, and the m510 with an 8- or 16-core Xeon D. Each price below is the entry build of that line, with the module as one line on top.

HPE ProLiant m710x

Intel Xeon E3-1585L v5

Intel Xeon E3-1585L v5 Quad Core 3.00 Ghz · 32 GB DDR4 · 1 × NVMe-SSD 240 GB · 1 Gbps

Machine, entry build: $62.58/mo
+ Google Coral Edge TPU: $13.99/mo
+ Hailo-8: $13.99/mo
Setup, once: $29
Cities: New York, Bucharest, San Francisco, Amsterdam, Miami

Billed in euros in Bucharest, Amsterdam, at the same figures.

Configure the m710x

HPE ProLiant m510

Intel Xeon D-1548 / Intel Xeon D-1587

Intel Xeon D-1548 Octa Core 2.00 GHz · 32 GB DDR4 · 1 × NVMe-SSD 240 GB · 1 Gbps

Machine, entry build: $62.58/mo
+ Google Coral Edge TPU: $13.99/mo
+ Hailo-8: $13.99/mo
Setup, once: $29
Cities: New York, Miami, Bucharest, Amsterdam, San Francisco

Billed in euros in Bucharest, Amsterdam, at the same figures.

Configure the m510

A module changes the machine, so the order is a custom build and carries the one-time setup fee shown. Memory, disks, port speed and the billing term are all set in the same configurator, which shows the final monthly total before you pay.

Setting it up

From order to first inference

We install the operating system and hand you root. The vendor runtime is not preinstalled, so its version is yours to pick and pin.

Configure the host. Open the m710x or m510 configurator, tick the Hailo-8 or the Coral, and choose the operating system.
Check the module is on the bus. Run lspci -nn | grep -Ei '1e60|1ac1'. The Hailo-8 shows as 1e60:2864 and the Coral as 1ac1:089a.
Hailo-8: install the PCIe driver and HailoRT from the 4.x line (the hailo8 branch of hailo-ai/hailort, or Hailo's Developer Zone packages), then run hailortcli fw-control identify to read the module's firmware and serial.
Coral: build the gasket and apex PCIe driver and libedgetpu from a maintained fork, load the driver, and check that /dev/apex_0 exists before you run your first model.
Pin the versions. Hold the kernel, the driver and the runtime at versions that work together. Nothing on our side updates the machine behind your back.

Answers

TPU server questions

What is a TPU server?

A server with a tensor processing unit attached to it. At Server Room that is a dedicated HPE Moonshot cartridge with a Google Coral Edge TPU or a Hailo-8 on its M.2 slot, for int8 inference on vision models. Google's Cloud TPUs are a different and much larger chip, rented only inside Google Cloud.

Can I rent a Google Cloud TPU from you?

No. Cloud TPUs are available only in Google Cloud, and no hosting company can rack one. For training or serving large models on hardware you control, see our GPU dedicated servers, where each machine has a dedicated GPU.

How much does a TPU server cost?

A complete machine with a Hailo-8 or a Coral starts at $76.57 a month. The module is one line on top of the machine, from $13.99 a month for the Hailo-8 and from $13.99 for the Coral, and the order carries a one-time setup fee from $29. The host cards above show which machine each price belongs to.

Hailo-8 or Coral Edge TPU: which should I choose?

Hailo rates the Hailo-8 at up to 26 TOPS and Google rates the Coral at 4 TOPS, both on int8. The Hailo-8 also has maintained software: HailoRT 4.24.0 for it came out in June 2026, while Google has archived the Coral repositories. Choose the Coral when your models already run on an Edge TPU.

Can a TPU server run an LLM?

No. Neither module can hold a model with billions of parameters, at any quantization. Size the model on our LLM hosting page, then run it on a GPU dedicated server or on one of our Ryzen AI machines.

Which operating systems and drivers work?

Any system the configurator offers for the cartridge can be installed. HailoRT is built for Linux and Windows; the Coral runtime and its community forks target Linux. We install the system, and you install the runtime at the version you choose.

Where are your TPU servers located?

TPU servers are sold in these cities: New York, Bucharest, San Francisco, Amsterdam, Miami. They run on our own network, AS19624. Orders in these cities are billed in euros, at the same figures: Bucharest, Amsterdam.

Is the TPU shared with other customers?

No. The module sits inside your cartridge and the machine has one tenant. You get root and the remote console, and no other customer's workload runs on it.

Pick a cartridge, tick the module

Questions before you order? Our chat AI answers in seconds and can check your account and servers, and a person takes over whenever you ask.

Configure an m710x