DeepX plans 2nm edge AI chip
Cette publication existe aussi en Français
DeepX in Korea is designing an edge AI chip with a power consumption of 5W to be built on the 2nm process at Samsung Foundries. It is also looking at a chiplet version.
The DX-M2 aims to have a performance of 40TOPS and will be optimised to run transformer models such as GPT as well as AI agents. This will be able to handle models with up to 20bn parameters such as DeepSeek to put natural language chat functions into embedded devices such as vending machines and humanoid robots. A custom RISC-V processor provides the scheduling.
This follows the 25TOPS DX-M1 that is built on a 5nm process that is entering mass production in the first half of 2025. This is optimised for video processing with a mix of floating point and 8bit integer (INT8) weights and is also being used for AI PCs and edge servers. “We are using mixed precision so we know how much 8bit, 4bit and floating point we have from the compiler so we know what we need to optimise for,” said Tim Park, strategic marketing director at DeepX.
DeepX has raised $531m in funding so far with backing from Hyundai Robotics. “Now we are at the scale up stage,” he said.
“The M1 is designed for vision applications and we worked with BMW, Mercedes, Volkswagen for vision for autonomous driving – we will have AECQ100 grade 1 in October, November. We learnt a lot about the automotive market so we are not really targeting automotive directly at this stage. Besides that we are going to the heavy vehicle market where we can work with Renesas, Texas Instruments and NXP for that market where other chips are too expensive.”
A second version of the chip, the V3, is a system-on-chip design, adding four ARM Cortex-A53 cores alongside the NPU of the M1 and an image signal processing block. This is aimed at still cameras and middle to high end IP cameras, handling vision models of 150m to 200m parameters for around $20. That will sample in August.
There will also be an SoC version of the M2
“We want to serve the lowest power NPU at under 5W, the same as the M1, and we are the first on 2nm and will sample in August 2026. People are starting to build their models now which will mature next year so it’s a perfect match for us. We are also preparing a SoC version.”
Humanoid robots are another target area for the M2.
“Humanoid robots need reasoning on the device and that’s another market we are targeting,” he said. The M2 will support the Nvidia vision language model which is used by robotic developers such as Figure.
“Figure AI have their own model and they were using Nvidia and are looking for their own silicon and M2 could be what they use.”
Longer term the company is looking at providing the M2 as a 2nm chiplet, particularly for he automotive market. “Once the chiplet market is mature we want to sell the M2 die to automotive OEMs, that’s what I expect. We have to be in the loop, as the market is not going to be open to everyone but we should be part of it.”
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
