Edge NPU and AI accelerator · INT8 TOPS, power envelope and TOPS per watt, host interface (PCIe/M.2), memory, temperature grade, framework support, package
A neural processing unit is a chip or module built to run neural-network inference efficiently beside a host processor; the reviewed products quote performance in INT8 TOPS. Ask three things first: what models and throughput the application needs, what power, interface and form factor the host system allows, and which temperature grade, memory and software stack the product requires.
Edge AI accelerator chips and modules (NPUs) for inference in embedded, industrial and automotive systems: chips, M.2 cards and board-to-board modules. Data-centre GPUs and training accelerators, complete AI boxes and single-board computers (see single board computer) are outside this page except where compared.
01
DEEPX lists its DX-M1 chip at 25 TOPS (INT8), power consumption from 1 W minimum to 5 W maximum, LPDDR4x up to 4,266 MT/s or LPDDR5 up to 6,000 MT/s with up to 8 GB, and support for Ultralytics, TensorFlow, PyTorch, ONNX and Keras on x86 and ARM hosts; it states INT8 quantisation keeps accuracy loss under 1% against FP32. Its M.2 module carries 4 GB LPDDR5 at 5,600 MT/s. [1]
Geniatech lists its AIM-B-H8 module, based on the Hailo-8 NPU, at up to 26 TOPS INT8 with a 2.5 W TDP, supporting concurrent multi-stream, multi-model inference with TensorFlow, TensorFlow Lite, Keras, PyTorch and ONNX; the same page also states 26 TOPS 'with just 5W typical power' and lists 5 W in its specification table, so the power figure should be confirmed for the target workload. [2]
Editorial calculation: 25 TOPS at 5 W is 5 TOPS/W for the DX-M1 at its maximum; 26 TOPS at 2.5 W is 10.4 TOPS/W and at 5 W is 5.2 TOPS/W for the Hailo-8 module. [1] [2]
The chip maker's own brochure frames the same silicon differently. Hailo's brochure for its Hailo-8 M.2 AI acceleration module (2020, distributed at Hot Chips 33) gives 26 TOPS with a power efficiency of 3 TOPS/W, an NGFF M.2 M-key board of 22 × 42 mm with breakable extensions to 22 × 60 and 22 × 80 mm, a 4-lane PCIe Gen 3.0 interface, TensorFlow and ONNX support through its Dataflow Compiler, and x86 or ARM AArch64 hosts under Linux. [3]
Editorial calculation: 26 TOPS at Hailo's 3 TOPS/W implies about 8.7 W (26 ÷ 3), against the 2.5 W TDP Geniatech lists for a module built on the same chip; the two figures cannot both describe full 26 TOPS load, so ask each supplier for measured power on your model at your frame rate. [3] [2]
Give the supplier the target models and input sizes, required frames per second or latency, precision (INT8, other), host processor and OS, interface and form factor, power and thermal budget, temperature grade, memory needs, SDK and toolchain requirements, longevity and quantities.
02
DEEPX lists a PCIe Gen3 x4 host interface (supporting Gen 1/2/3 and x1/x2/x4), a 17 × 17 mm 625-ball FC-BGA chip package, operating ranges of 0 to 70 °C commercial, −40 to 85 °C industrial and AEC-Q100 Grade 3, with Grade 2 (−40 to 105 °C) on request, and an M.2 2280 M-key module (22 × 80 mm) with optional heatsink; it supports Windows 11/10, Debian-based Linux, Yocto and Docker. [1]
Geniatech's AIM-B-H8 uses a board-to-board connector with 4-lane PCIe Gen3, measures 40 × 55 mm and operates from −40 to +85 °C. [2]
03
MemryX rates its accelerators in TFLOPS rather than INT8 TOPS. Its MX3 chip page lists 6.25 TFLOPS per chip, BFloat16 activation precision with INT8 weights, 10.5 MB of on-chip weight memory, 200 mW to 3 W per chip, a 9 × 9 mm FC-CSP package, host interfaces of PCIe Gen 3 and above and USB 3.2 Gen 1, and a junction temperature of −40°C to 125°C. [4]
The unit on the rating changes what the number means. DEEPX quotes 25 TOPS at INT8 for its DX-M1 [1] and Geniatech quotes up to 26 TOPS INT8 for the Hailo-8 based AIM-B-H8 [2], while MemryX quotes 25 TFLOPS for the four-chip Cascade 100M with BFloat16 activations [5]. The three figures sit close together on paper but count different operations, so ask each supplier for frames per second on the same model, at the same input size, measured on your host.
The memory figures differ in kind as well. DEEPX lists LPDDR4x or LPDDR5 memory with a capacity of up to 8 GB for its chip [1], whereas MemryX lists weight storage of 10.5 MB on-chip per MX3 and 42 MB on the four-chip module [4] [5]. A model whose weights exceed the on-chip total needs more chips or a different part, so give the supplier the parameter count of each model you plan to run.
Editorial calculation: four MX3 chips at 6.25 TFLOPS each give 25 TFLOPS and four at 10.5 MB give 42 MB, matching the Cascade 100M figures; four at 3 W each is 12 W, the top of the module's 1 W – 12 W range. [4] [5]
| Item | MemryX MX3 chip | MemryX Cascade 100M (MX3-2280-M-4) |
|---|---|---|
| AI performance (as published) | 6.25 TFLOPS per chip | 25 TFLOPS |
| Chips | 1 | 4× MX3 |
| Weight memory | 10.5 MB on-chip | 42 MB on-chip |
| Power | 200 mW to 3 W per chip | 1 W – 12 W |
| Package / form factor | 9 × 9 mm FC-CSP | M.2 2280 M-Key, 22 x 80 mm |
| Host interface | PCIe Gen 3 and above; USB 3.2 Gen 1 | PCI Express Gen 3 and above |
| Input voltage | — | 3.3V +/- 5% |
| Temperature | Junction −40°C to 125°C | Operating -40ºC to 85ºC Industrial |
| Humidity / certification | — | 10% ~80% RH non-condensing; CE/FCC Class A |
Values are from MemryX product pages; no measured benchmarks were reviewed.
04
The supplier linked to this page lists AI accelerator modules; its card links to the page reviewed.
Ask any supplier for INT8 TOPS and measured throughput on your models, power at that load, interface and form factor, memory, operating temperature grade and qualification, SDK, compiler and supported frameworks, driver support for your OS, thermal solution, and product lifetime commitment.
05
This is a functional reference built from the sources cited; it is not one maker's specification.
| Part | Role | Fix in the order |
|---|---|---|
| NPU die | Runs inference. [1] [2] | TOPS; precision. |
| Memory | Holds models and activations. [1] | Type; capacity. |
| Host interface | Connects to the CPU. [1] [2] | PCIe lanes; connector. |
| Package or module | Mechanical fit. [1] [2] | BGA; M.2; B2B. |
| SDK and runtime | Compile and deploy models. [1] [2] | Frameworks; OS. |
06
An NPU quotation needs values for your system. The table explains each field, gives published values where a reviewed source states them, and says what to obtain from the supplier.
| Field | What it means and published values | Ask the supplier for |
|---|---|---|
| Computing Power (TOPS) | e.g. 25 or 26 TOPS INT8. [1] [2] | TOPS and model FPS. |
| Power Consumption (W) | e.g. 1–5 W; 2.5 W TDP; 3 TOPS/W stated by chip maker. [1] [2] [3] | Power at your load. |
| Operating Temperature (°C) | e.g. 0–70; −40 to 85; −40 to 105 on request. [1] [2] | Grade. |
| Supply Voltage (V) | e.g. 3.3V +/- 5% input on an M.2 2280 module. [5] | Rails. |
| Process Node (nm) | Not stated in the reviewed sources. | Node. |
| Memory Bandwidth (GB/s) | e.g. LPDDR5 up to 6,000 MT/s. [1] | Bandwidth. |
| Package Dimensions (mm) | e.g. 17 × 17 BGA; 22 × 42/60/80 M.2; 40 × 55 module. [1] [3] [2] | Drawing. |
| Weight (g) | Not stated in the reviewed sources. | Weight. |
| Operating Humidity (% RH) | e.g. 10% ~80% RH non-condensing. [5] | Range. |
| ESD Tolerance (V) | Not stated in the reviewed sources. | HBM/CDM ratings. |
| Rating unit and precision | INT8 TOPS [1] or TFLOPS with BFloat16 activations and INT8 weights. [4] | Unit, precision, FPS on your model. |
07
Paste it into your enquiry below or into an email to a maker, then replace every bracket.
Models: [network; input size; FPS or latency] Precision: [INT8; other] Host: [CPU; OS] Interface and form: [PCIe lanes; M.2; B2B; chip] Power and thermal: [W; heatsink] Grade: [commercial; industrial; automotive] Software: [frameworks; SDK]
This company lists AI accelerator modules.
Lists Hailo-8, Hailo-10 and other AI accelerator modules and M.2 cards.
Product scope checked 2026-10-11
Explore company information and manufacturing capabilities, then discuss the scope of your project with a relevant supplier.
DEEPX lists 1–5 W for its 25-TOPS DX-M1; Geniatech lists 2.5 W TDP and 5 W typical for its 26-TOPS Hailo-8 module. [1] [2]
Both reviewed parts use PCIe Gen3 x4, as an M.2 2280 card or a board-to-board module. [1] [2]
Both list TensorFlow, PyTorch, ONNX and Keras; DEEPX adds Ultralytics and Geniatech TensorFlow Lite. [1] [2]
Sources checked 2026-10-12. DEEPX designs edge NPUs; Geniatech builds AI accelerator modules around Hailo NPUs.
Ask for use case, specification boundaries, supplier type, and RFQ preparation information for this product.
Compare manufacturer profiles with relevant product and process capability.