DeepX plans to spend more than $100 million on a Samsung-made AI chip for robots
The South Korean startup sent final DX-M2 designs to Samsung's foundry and expects first silicon by March 2027
South Korean edge AI chip firm DeepX has completed a prototype of its second-generation AI accelerator, the DX-M2, incorporating custom processor-in-memory (PIM) technology from Samsung. The DX-M2 aims to bring powerful generative AI capabilities to edge devices such as robots, drones, and IoT gadgets by enabling inference of models with up to 100 billion parameters at less than 5 watts of power.
DeepX's first-generation chip, the DX-M1, is already being mass-produced for vision applications. The DX-M2 is designed to run models with 20 to 100 billion parameters (approximately 3 to 20 billion active parameters) at 10 to 30 tokens per second within the same 5 W power budget as the M1. DeepX founder and CEO Lokwon Kim attributes the DX-M2's improved performance to a threefold increase in memory bandwidth from 50 to 150 GB/s, along with doubling the power efficiency of the NeuroProcessing Unit (NPU) and a process node shift to 2nm, which also doubled overall efficiency.
The DX-M2 can handle large mixture-of-experts models and supports new techniques such as speculative decoding and TurboQuant. By offloading GEMV (general matrix-vector multiplication) operations to LPDDR5x PIM, DeepX asserts that the DX-M2 can achieve up to 6.7 times the tokens per second for Llama3.1-8B compared to the M2 alone.
However, this increased performance comes at a cost, as it requires 2.7 times more power consumption. DeepX is already collaborating with Samsung Memory to integrate the DX-M2 with LPDDR5x PIM, and they are simulating the potential 6.7x performance boost. The DX-M2 is targeted at various sectors, including AI PCs, robotics, automotive, and IoT devices.
Robotics companies are particularly interested in the DX-M2 due to its low power consumption and higher on-device compute capabilities. DeepX is working on a transition path for robotics customers moving from Nvidia's Isaac robotics platform to their own Newton platform, allowing them to switch APIs with minimal effort. The company is also partnering with Intel to integrate OpenVINO with the DeepX SDK, enabling AI workloads on x86 Intel edge CPUs.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
