DeepX claims inference up to 100B parameters at less than 5 W

South Korean startup DeepX has taped out its second-generation DX-M2 edge AI accelerator, designed to run large language models locally in robots, drones, vehicles, and other power-constrained devices. The company says it can handle models with 20 billion to 100 billion total parameters, typically 3 billion to 20 billion active parameters, at 10-30 tokens per second while consuming under 5 W. Compared with its mass-produced vision-focused DX-M1, the M2 triples memory bandwidth to 150 GB/s, doubles NPU efficiency through design changes, and gains another efficiency increase from a move to a 2-nanometer process, reaching 80 TOPS.

DeepX is also developing an arrangement with Samsung that moves memory-heavy calculations into LPDDR5x processor-in-memory chips. Simulations for Llama 3.1 8B project 6.7 times as many tokens per second as the M2 alone, but at 2.7 times the total power; the company says a twofold speed gain may be possible without exceeding the original power budget. Samples are expected in March 2027, with production planned for fall. DeepX reports more than 100 purchase orders for the M1 and is building software and industry partnerships to ease migration from Nvidia robotics systems. Its planned 2028 DX-M3 will use chiplets and LPDDR6 PIM, targeting automotive and humanoid-robot applications. The company’s broader case is that efficient on-device inference can reduce reliance on data centers and keep AI responsive when connectivity is limited.