All news

DeepX tapes out 5-watt DX-M2 chip for local models with up to 100 billion parameters

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

If the claimed specifications are confirmed in samples, large language models will be able to run directly in robots and laptops with modest energy consumption.

South Korea's DeepX has completed the design of its second-generation DX-M2 accelerator and sent it to manufacturing. According to the company, the 2nm chip is designed for models with 20–100 billion parameters and 3–20 billion active parameters, generation at 10–30 tokens per second and 5 W power consumption. It delivers 80 TOPS and 150 GB/s of memory bandwidth. Samples are expected in March 2027, with mass production planned for autumn 2027. DeepX and Samsung are also testing LPDDR5x-PIM memory, which performs some operations inside memory chips. Simulation showed Llama 3.1 8B running 6.7 times faster than on a single DX-M2 while increasing power consumption 2.7 times; the company expects a twofold improvement without increasing the power budget. These are simulation results for the future combination, not measurements of a finished device.

Sources 1