First SoM with DeepX AI chip
Virtium Embedded Artists in Sweden has developed the industry’s first system on module (SoM) card with the latest DeepX AI accelerator.
The DX-M1 SoM pairs the 5W, 25TOPS DeepX M1 accelerator with the i.MX8M mini from NXP Semiconductor to provide AI inference at the edge, and Virtium is planning version with the M2.
The module, which measures 82mm x 50mm and uses the MXM3 300pin connector, has been developed by Embedded Artists which was last year acquired by US storage and embedded module maker Virtium.
“We chose DeepX for this first turn as they are a low power solution with 25TOPs and that increases reliability and that’s key. We have development kits ready and production can be through EA or Virtium, so we can be in volume production in a matter of weeks,” said Dave Beasley, executive VP of worldwide sales and marketing tells eeNews Europe.
“We think we have a first mover advantage at the moment and we have a good relationship with DeepX and they have a good roadmap,” he said. That includes the M2 which aims to have performance of 40TOPS in the same 5W power envelop by using a 2nm process technology. “We are planning a version with the M2 for sure,” said Beasley.
NXP has added a neural processing unit (NPU) to its i.MX9 family and is in the process of acquiring AI chip designer Kinara to add more AI capability, but the AI accelerator is still key.
“Embedded NPUs will handle a certain amount of applications but you will run out of power and memory and then you need an accelerator,” says Paul Martin, VP technology strategy at Virtium.
”DeepX covers from the low levels of sensor data with RNNs for industrial health monitoring and then you go up in performance with CNNs for vision for drones. Then you get into transformer architectures with transformer vision models summarising videos, gesture recognition so you can really start to interact with machines.”
“We are in the early days of edge AI but it is growing fast, and people will customise and optimise.”
The SoM provides 4GB of LPDDR5 memory for the AI processor accessed via a 64-bit, 4-channel data bus. This enables the DX-M1 to run multiple AI models such as YOLO at the same time.
“The SoM has 4GBytes for the DeepX M1 and 2GBytes for the i.MX8mini and that takes you up into the vision models and allows you to cover a lot of cameras at once,” said Martin. “Even YOLO v12 the models aren’t getting bigger, there’s a lot of optimisation.”
The SoM, which sits on a carrier board with customer specific interfaces, is aimed at drones, security and surveillance and automated inspection and monitoring as well as transportation, medical and smart farming applications. Virtium Embedded Artists provides pre-designed standard carrier boards to ease integration into customers’ designs.
There are two versions with a 1.8GHz i.MX 8M Mini for operation at temperatures between 0°C and 70°C, or with a 1.6GHz i.MX 8M Mini with an operating-temperature range of -40°C to 85°C.
“We will stay focussed on DeepX but we have the flexibility8 to use other accelerators so when we see the right solutions we will engineer something around those. This is the first release and we have several more planned for this year more additional application processors,” said Beasley.
Virtium will manufacture the modules in the US and Vietnam. “We have some manufacturing with EA in Europe in a pilot line but when we go to scale it will be US or Vietnam,” he said. “We were finding ways to get into the SoM market with Embedded Artist in Sweden who are creating development kits so we acquired them last year. We wanted to leave their brand intact but add in Virtium,” he added.
Shipments of the module are expected to start in August.
Raspberry PI has worked with DeepX to support the M.2 PCI Express module measuring 22 x 80mm running on the Raspberry Pi 5 single board computer.
www.embeddedartists.com/products/imx8m-mini-dx-m1-som
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
