DeepX tapes out $50 edge AI chip to run big models at 5W
What's the deal? South Korean fabless startup DeepX has taped out DX-M2, a second-generation edge AI chip built on SamsungDealroom has a profile for this one. Try Dealroom → Electronics' 2nm process. The company says the single-die part will run generative AI models locally at roughly 5 watts, targeting a price under $50 a chip. Mass production is targeted for fall 2027.
Why now? The tape-out hands the design over for fabrication, with first silicon expected in March 2027 and customer evaluation samples in the second quarter. Reports differ on timing beyond that: initial shipments in the fourth quarter of 2027, with full-scale production planned for 2028.
What's the endgame? DeepX is positioning DX-M2 for robots, drones, AI PCs, and industrial devices that need to run models on-device without sending work to remote data centres. The chip can hold mixture-of-experts models of 20 billion to 100 billion total parameters — with roughly 3 billion to 20 billion active per token — decoding at 10 to 30 tokens per second within the power budget.
The bet rests on memory, not compute. Chief executive Lokwon Kim told EE Times the biggest change between generations is a threefold jump in memory bandwidth, from 50GB/s to about 150GB/s. DeepX is also using a custom Samsung LPDDR5X processing-in-memory part, which it says could lift tokens per second by up to 6.7 times on Llama 3.1 8B — at 2.7 times the power.
Why it matters: If the $50 target holds, it is roughly a quarter of what rival edge accelerators are said to cost per part, according to ZDNet Korea. That could reset the bill of materials for battery-powered devices currently forced to choose between a cheap chip that cannot hold a useful model and an expensive one that can.
What could go wrong? The tape-out is a design milestone, not proof of performance. The 80 TOPS and 6.7x figures are pre-silicon specs and simulations, and on-chip numbers remain a year away. DeepX reports development spending will exceed $100 million, per Bloomberg.
The signal: DeepX's first chip, the DX-M1, has entered mass production for vision applications, giving it some commercial footing. But the announcement lands amid active competition for low-power edge and robotics chips, where the decisive milestones will be working benchmarks, customer evaluations, and repeatable production — not the tape-out alone.
Read more: eetimes.com
Image credit: DEEPX
