DeepX tapes out DX-M2, a 2nm chip designed to run 100B-parameter models in about 5 watts

DeepX says the DX-M2 runs models of 20 billion to 100 billion parameters, roughly 3 billion to 20 billion active, at 10 to 30 tokens per second depending on the model[1]. Memory bandwidth triples from 50 GB/s on the first-generation DX-M1 to about 150 GB/s, and the NPU reaches 80 TOPS. DeepX credits design changes with doubling power efficiency and the move to 2nm with doubling it again[1].

The 5 watt figure covers the chip, not the whole system. DeepX's own product page lists 5W for the SoC only on the standard package with up to 96GB of external LPDDR5X, and 8.5W for a Pro module with memory included[3]. The chip also supports large mixture-of-experts models, speculative decoding and TurboQuant[1].

DeepX has spent a year and a half working with Samsung to pair the DX-M2 with Samsung's LPDDR5X-PIM, memory that can run memory-bound operations itself instead of shipping weights back and forth[1]. In DeepX's simulations, offloading those operations gives 6.7 times the tokens per second on Llama 3.1 8B, though it needs 2.7 times the power. The company thinks it can get about twice the performance within the normal power budget, and says its PIM compiler is still a prototype running against a simulation[1].

DeepX is targeting robotics companies that prototype on Nvidia GPUs and then hit power and heat limits, as well as AI PCs, in-car agents and smart home hubs[1]. It has built robotics libraries called DX-Newton meant to mirror Nvidia's Isaac APIs, so developers can switch by changing the API name, though a full robotics stack is still in development. Hyundai Motor Group's Robotics Lab already uses the DX-M1[1].

The first-generation DX-M1 is in mass production. DeepX told EE Times it has more than 100 purchase orders[1], while The Korea Times reported 88 commercial orders worth about US$13.3 million as of October[2]. The DX-M2 has also been named a CES 2027 Innovation Awards honoree in the AI category[2].

DeepX expects first silicon in March 2027, customer sampling in the second quarter and initial production volumes in the fourth quarter of 2027, with full-scale mass production in 2028[2]. A chiplet-based DX-M3 using LPDDR6-PIM is on the roadmap for 2028, aimed at carmakers and humanoid robot builders[1].

Most on-device AI chips today handle vision tasks like object detection. If the DX-M2 hits its targets in real silicon, a drone or robot could run a model in the 100 billion parameter class locally, without a network link or a power-hungry GPU. The big caveats are that today's figures are design targets and simulations, and that system power will be higher than the headline 5 watts once memory is counted.