South Korean edge-AI startup DeepX has taped out its second-generation AI accelerator, the DX-M2, a chip designed to run large language models of up to 100 billion parameters on less than five watts of power. The milestone moves generative AI off the cloud and onto robots, drones, laptops and cars that cannot afford a data center's power budget.
Built on Samsung Foundry's 2-nanometre process, the DX-M2 is aimed at models with 20 billion to 100 billion parameters, delivering 10 to 30 tokens per second depending on the model. DeepX says the chip reaches 80 TOPS, with memory bandwidth tripled from 50 to 150 GB/s compared with its first-generation DX-M1, which is already in mass production.
Memory that computes
The DX-M2's headline trick is processor-in-memory, or PIM. Most AI inference is limited not by raw compute but by shuttling model weights between memory and the processor, especially during the matrix-vector operations that dominate language-model decoding. PIM lets some of those operations run inside the memory chip itself.
DeepX spent a year and a half working with Samsung Memory on the technology, pairing the DX-M2 with Samsung's LPDDR5x-PIM. Its simulations show offloading those operations into memory delivers 6.7 times the tokens per second on Meta's Llama 3.1 8B model versus the DX-M2 alone, at the cost of 2.7 times the power. Within the same power envelope, DeepX says it can still get roughly double the performance. The company has received LPDDR5x-PIM samples and is testing them in-house, with a prototype compiler running on simulations while the hardware modules are built.
Robots, drones and AI PCs
The chip targets the fast-growing market for physical AI: robots, drones, AI PCs and in-vehicle agents. DeepX is working with Taiwanese manufacturers on AI laptops and workstations, and says robotics customers now prototyping on Nvidia's Thor platform want lower power and more on-device compute. Its DX-Newton software libraries are designed to let customers move pipelines from Nvidia's Isaac robotics platform with little more than an API change, and the company claims switching can deliver 20 times longer battery life.
Hyundai Motor Group's robotics lab already uses the first-generation DX-M1, and DeepX says it holds more than 100 purchase orders for that chip across drones, robots, smart factories, surveillance, aerospace and defence, healthcare and the internet of things. It is also collaborating with Intel to integrate OpenVINO with its software kit, and has discussed a robotics platform with AMD based on its Kria boards.
What comes next
Engineering samples of the DX-M2 are expected in March 2027, with mass production targeted for autumn 2027. DeepX's third-generation chip, the DX-M3, is planned as a chiplet-based design for 2028, targeting 1 ePOPS of INT8 performance using LPDDR6-PIM, aimed at carmakers and humanoid-robot builders.
Founder and chief executive Lokwon Kim argues the shift matters beyond the edge. "This is important because with 5 W power consumption we can go everywhere, drones or any other battery-powered device," he told EE Times. He believes that if models of up to 100 billion parameters can run on devices, AI data center traffic could fall by more than 80 percent. Whether the DX-M2 lives up to its simulated numbers will only be clear once silicon returns from the fab — but the tape-out puts DeepX's bet on the table.
