DeepX Tapes Out 2nm DX-M2 Edge Chip, Claims 20-100B Param Inference Under 5W

Korean edge-AI startup DeepX taped out its second-gen DX-M2 on a 2nm node with 80 TOPS, 150GB/s memory bandwidth and under-5W power, saying it can run 20-100B-parameter models at 10-30 tokens per second. Paired with Samsung's LPDDR5x-PIM, simulations project a 6.7x throughput lift on Llama3.1-8B at 2.7x the power. Samples target March 2027 with mass production in fall 2027, aiming the chip at robots and drones.