roobriefIT news for developers
Search

By Hardware

DEEPX NPU HAT for Raspberry Pi 5: What It Can Run

Raspberry Pi’s blog on Oct. 5 featured the Sixfab AI HAT+, which carries an NPU from South Korea’s DEEPX. The version you can buy today is the 25-TOPS DX-M1M at $90.

An NPU is an on-chip accelerator built for neural-network math, and on-device AI means running models on the board instead of in the cloud. TOPS counts trillions of operations per second and works as a rough speed gauge for accelerators. Sixfab and DEEPX announced this board on June 26, so the Raspberry Pi post is a showcase rather than a launch.

What on-device AI can do here

Light models already run on the Pi’s CPU. Raspberry Pi treats an accelerator as the answer to the few jobs the CPU can’t carry, such as an always-on camera, a robot, or industrial equipment. An NPU drawing about 3W can keep that kind of work going without a cloud round trip. The 3W is the NPU’s typical draw, though. Under load, a Pi 5 with the HAT pulls 13–15W by Sixfab’s figures, and the official 27W power adapter is recommended.

The range is narrow. The board can pick out people, cars, and poses, attach a short caption, or turn a spoken phrase into a command. Questions that need world knowledge and long conversations belong in the cloud, and Raspberry Pi draws the same line.

What each TOPS tier covers

At 13 TOPS, one camera and one model is the baseline: is it a person, where are they looking, what is their pose. The official AI HAT+ with a Hailo-8L sits here at $70. DEEPX’s 13-TOPS version (DX-M1ML, $63) is due in the fourth quarter, so you can’t buy it yet.

At 25 TOPS, you stack several models: detection plus pose or segmentation on the same frame, or multiple cameras. The DEEPX HAT lives here, and the official Hailo-8 is 26 TOPS at $110. When it announced the first AI HAT+ in 2024, Raspberry Pi said 13 and 26 TOPS do the same kinds of work, and 26 handles larger networks and several models at once.

At 40 TOPS, short text enters the picture. The official AI HAT+ 2 (Hailo-10H) is rated 40 TOPS at INT4 and carries 8GB of onboard memory, and its docs cover short captions and voice-to-command. It launched at $130, but the official page now lists $200.

Tier What it does Example
13 TOPS One-camera vision Official AI HAT+ $70
25–26 TOPS Several models, several cameras DEEPX $90, official $110
40 TOPS (INT4) Short text and voice too Official AI HAT+ 2 $200

TOPS figures aren’t measured on one ruler. DEEPX’s 25 is INT8, while the AI HAT+ 2’s 40 is INT4. The Orange Pi 6’s 45 TOPS adds up CPU, NPU, and GPU, and its NPU alone is 28.8 TOPS. The Banana Pi AIM7’s NPU tops out at 6 TOPS.

How developers actually choose

In practice, the choice follows how short the software path is.

For cameras and home security, it’s the official Hailo. Raspberry Pi OS’s rpicam-apps and Picamera2 use the Hailo NPU automatically. Frigate, the open-source recorder that picks out people and cars, supports Hailo-8 and 8L as detectors. Jeff Geerling wrote that a $70 Hailo-8L ran two cameras with detection around 12ms and low CPU use. Frigate no longer recommends Coral for new installs.

If you only need a language model, an NPU HAT may not be the answer. Geerling measured text generation on the AI HAT+ 2 and found the Pi 5’s CPU faster. The HAT wins when it frees the CPU so a camera and short text can run together, but at launch that combination failed for lack of examples. Hailo, for its part, reports that time to first token (a token is a small piece of text, roughly part of a word, that AI models read and write) is much shorter on the HAT. The two numbers measure different things.

On the DEEPX side, you export a model to ONNX (a common model file format that training frameworks can exchange), quantize it to INT8 (shrinking its numbers to fewer bits), and compile it to a .dxnn file for the chip. Compilation happens on an x86_64 PC, and the runtime executes it on the board. In Raspberry Pi’s benchmark on a Pi 5 with 8GB, mobilenet_v2 hit 2,361 FPS and deeplabv3plus 155 FPS. Sixfab quotes 30–35 FPS for YOLOv8n at 640×640. It isn’t wired into rpicam or Frigate the way Hailo is.

Orange Pi and Banana Pi boards carry the NPU in the chip, so there’s no HAT to buy.

What we couldn’t confirm

  • Raspberry Pi’s post shows a Qwen3-1.7B benchmark on this HAT (4.64 tok/s on the DX-M1M), but Sixfab’s product-page FAQ says the DX-M1M is currently vision-only and doesn’t support language models. We couldn’t tell which is right.
  • We couldn’t confirm that Frigate has added DEEPX to its release docs. The stable docs list only Hailo.
  • The DEEPX HAT is still going through certifications such as CE and FCC, and it works only with the Pi 5 (not the Pi 4, CM4, or CM5 IO Board).

What to check now

What follows is our estimate and advice, based on public specs and reviews.

  • For cameras only, start with the $70 Hailo-8L. The 26-TOPS Hailo or the DEEPX 25-TOPS part is for when you need several models or streams.
  • If you need text, weigh memory and generation speed together. Before buying the DEEPX HAT for that, sort out the mismatch between the blog and the FAQ.
  • Check that your model compiles to the chip’s format first. Hailo and DEEPX files aren’t interchangeable.

What to watch

Whether the DEEPX 13-TOPS version actually ships, how Raspberry Pi and Sixfab reconcile their language-model claims, and when Frigate puts Hailo-10H or DEEPX into its release docs.

Sources

Source: www.raspberrypi.com

← Back