On-Prem

Systems

Raspberry Pi unveils Hailo-powered AI Kit to make the model 5 smarter

13 TOPS for $70 has PiBoss Eben Upton predicting an ML explosion


Raspberry Pi has created a machine-learning addition for its single board computer that features the Hailo-8L AI accelerator.

This AI Kit is an M.2 Hat board with a preinstalled M.2 stick that carries the Hailo-8L accelerator, marking the first time Raspberry Pi has ever sold a first-party accessory for AI. The Hailo-8L is a cut down version of the Hailo-8 chip announced earlier this year, and offers 13 TOPS of performance at INT8 rather than the 26 TOPS of the full-fat version.

For comparison, Qualcomm, AMD, and Intel have recently announced SoCs packing NPUs at 45, 50, and 48 TOPS respectively.

So while the Hailo-8L is therefore not Earth's mightiest accelerator it is – like all Raspberry Pi products – keenly priced at $70, plus ~$80 for its Raspberry Pi 5 host. The combo of an Arm-compatible Pi and the accelerator will therefore be well below the cost of just an NPU-equipped SoC.

The AI Kit: M.2 Hat board with Hailo-8L M.2 stick ... Let's face it, this is a bargain – Click to enlarge

Plus, the Hailo-8L has remarkably low power consumption, Hailo's CTO Avi Baum told The Register. The chip consumes about a watt of power for every three TOPS, capping out at roughly five watts, and usually uses just a watt or two for a "typical" workload – such as processing 60 FPS video in real time.

Relying on a separate accelerator rather than an integrated NPU is an intentional design choice, Raspberry Pi CEO Eben Upton told The Register. The computer designer already started to focus on a disaggregated architecture with the Raspberry Pi 5, which has the CPU and GPU cores on a 16nm chip while I/O functions have been moved to a cheaper 40nm chip. An NPU integrated into the 16nm silicon slab would have been expensive.

Upton observed that including the Hailo-8L or some other accelerator on the Raspberry Pi 5s main board wouldn't have been ideal, as it would have made it more expensive. By offering the AI Kit, users can still buy the Raspberry Pi 5 for a lower price, and add the Hailo-8L if they are keen on exploring machine learning. The modular approach also meant rapid development: Upton revealed the AI add-on moved from concept to final product in just six months.

Out of the cloud

Although the AI Kit represents a big performance leap for the Raspberry Pi 5, the single-board machines already have history in the field.

Upton pointed out that long before AI became the Next Big Thing, Raspberry Pi users were already using the technology for things like object detection. The CEO mentioned one particular instance of a Japanese cucumber farm using a Raspberry Pi 3 to sort thorny cucumbers with about a 70 percent accuracy rate.

However, that cucumber sorter relied on AI processing in the cloud. The debut of this kit means some AI workloads can run on an Pi – reducing reliance on connectivity.

"It's a good host for smaller models, and more optimized and quantized large language models (LLMs) are running at tokens per second on the Raspberry Pi 5," Upton told The Register, before noting that some work will still need to be run in the cloud. LLaMa 3 70B is tough to run locally on a PC unless you have an impressive setup.

While there aren't a ton of use cases right now, Upton expects the lower entry price of Raspberry Pi 5 AI to create an "exponential increase" in AI applications and solutions.

When asked whether the kit represented a long-term partnership between Raspberry Pi and Hailo and if anything other AI hardware was in the works, Upton replied that he would be "disappointed if this is the last Raspberry Pi-Hailo collaboration." ®

PS: Raspberry Pi is targeting 260 to 280 pence per share when it debuts on the stock market this month, making it worth up to £540 million ($690 million).

Send us news
34 Comments

OpenAI develops AI model to critique its AI models

When your chatbots outshine their human trainers, you could pay for expertise ... or just augment your crowdsourced workforce

So much for green Google ... Emissions up 48% since 2019

AI datacenters blamed for the increase, even as Chocolate Factory bets on AI to fix it

AMD buys developer Silo AI in bid to match Nvidia's product range

First it comes for market leader's GPUs ... now it's nibbling at software

Cloudflare debuts one-click nuke of web-scraping AI

Take that for ignoring robots.txt!

64% of people not happy about idea of AI-generated customer service

Not unreasonably, nearly half worried it would give them the 'wrong answers'

Meta training AI models on citizen data gets a hard não from Brazil

Zuckerborg's justification isn't good enough, says watchdog

A friendly guide to local AI image gen with Stable Diffusion and Automatic1111

A picture is worth a 1,000 words... or was that a 1,000 TOPS

Palantir, Oracle cosy up to offer Karp firm's tech across Big Red's cloud

Foundry and AI plaform available in OCI

China’s homebrew openKylin OS creates a cut for AI PCs

Devs of OS named for a mythical beast join in the 'local models will will deliver legendary productivity' trope

SoftBank buys struggling UK AI chipmaker Graphcore

Potential for combo with Arm tantalizes

With users mostly happy to keep older kit, Macs just ain't selling like they used to

And the AI PC revolution probably won't help, we're told

Babel fish? We're getting there. Reg reviews the Timekettle X1 AI Interpreter Hub

A handy standalone translator, but you'll need deep pockets, both figuratively and literally, if you want one