AI Potluck
Infrastructure / Edge AI hardware

Hailo-8

Hailo Technologies

Edge AI inference accelerator from Hailo Technologies in Israel, delivering up to 26 TOPS at roughly 2.5W typical power. It is sold as a bare chip and in M.2, mPCIe and PCIe card form factors with on-die memory, and supports TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX through the Hailo Dataflow Compiler and the HailoRT runtime.

hailo.ai answers 403 to every automated request, so the figures were checked against Geniatech's AIM-M-H8 spec table, Waveshare's listing and Hailo's M.2 datasheet mirrored by Premio. The compiler needs registration while the runtime does not, which the openness axis weighs. Verified 2026-08-14 via those listings and the HailoRT README.

Openness

3 high confidence
3.0
schematics
none
toolchain
partial (open runtime (HailoRT) + closed/registration Dataflow Compiler)
datasheets
brief (public briefs)
blobs
required (proprietary silicon + firmware)
retail
open_market (global modules)

Proprietary silicon with an open source runtime and public product briefs, but the model-compiler toolchain requires registration and the silicon itself is closed. HailoRT's README carries the open half of the toolchain - "libhailort, pyhailort & hailortcli - distributed under the MIT license", with hailonet under LGPL 2.1 - and the gated half is named by the Model Zoo README, which instructs "Install Hailo Dataflow Compiler ... In case you are not Hailo customer please contact hailo.ai". The same README is why firmware blobs are required: the PCIe driver "includes the Hailo-8 firmware that runs on the Hailo device". The published documentation is Hailo's own Hailo-8 M.2 data sheet Rev 1.1, mirrored publicly by distributor Premio and stamped "Hailo proprietary - unauthorized reproduction prohibited", which is also why no schematics are published at all. Retail availability is Waveshare listing the module at $179.99 with an add-to-cart button. hailo.ai refuses every automated request, so all of this rests on substitute sources.

Adoption

4 high confidence
4.0

Retail-available modules and a large developer base; the vendor reports 100K+ users and hundreds of deployed customer products, and runs an active GitHub org. That 100K-users figure is not independently derivable - hailo.ai refuses every automated request and no substitute publishes it - so the number is the vendor's own. What the substitutes do establish is the standing the band was placed on: the part is stocked and sold by several independent distributors on three continents (Waveshare at $179.99, Geniatech AIM-M-H8, AAEON/UP, Advantech via Mouser), and HailoRT is a live repo at 215 stars and 80 forks. That is broad standing for an edge accelerator, so level 4 / broad holds. The band rests on that observable distribution rather than on the user count, which is why the unverifiable figure does not move it.

Capability

3 high confidence
3.0

Strong dedicated CNN inference throughput for the edge, but vision-oriented and limited to models that fit on-die memory. Every element of the recorded value reproduces against substitutes, hailo.ai being unfetchable - Geniatech's AIM-M-H8 spec table reads "AI performance 26 TOPS(INT8)", "AI Model Frameworks Supported TensorFlow,TensorFlow Lite,Keras, PyTorch&ONNX" and "Memory - Processor integration", which is the on-die claim stated as a specification rather than as marketing; Waveshare independently lists 26 TOPS at 2.5W typical. That is one band below the Hailo-10H, which adds a DDR interface and generative-model support.

Verified 2026-08-14