DeepX says it has completed the design of its DX-M2 edge AI accelerator and submitted the final plans to Samsung’s foundry. The chip is designed for Samsung’s 2-nanometer process and aims to run generative AI directly on robots, drones and other devices with tight power and cooling limits. SourceSource
Bloomberg’s November 13 report says DeepX plans to spend more than $100 million developing the chip and expects its first silicon—the first manufactured chips—by March 2027. Tape-out is a design milestone, not a shipping notice. Source
DeepX’s simulations project 6.7× the tokens per second for Llama 3.1-8B when DX-M2 is paired with Samsung’s LPDDR5x processor-in-memory technology, compared with DX-M2 alone. CEO Lokwon Kim told EE Times that this configuration also requires 2.7× the total power. The speedup has an electricity bill. Source
Letting memory do some of the work
As EE Times explains, processor-in-memory technology moves selected calculations into the memory chip rather than repeatedly transferring data to a separate processor. DeepX’s simulation offloads general matrix-vector multiplication, or GEMV, operations into Samsung’s LPDDR5x-PIM memory. That memory technology is distinct from the 2-nanometer process used to manufacture DX-M2. Source
DeepX separately projects a 2× throughput improvement within the same power budget as non-PIM operation. Source
DeepX is targeting roughly 5 watts for DX-M2. The absolute power draw of the higher-power 6.7× PIM configuration remains unspecified. SourceSource
In EE Times’ November 9 report, Kim said DeepX had received LPDDR5x-PIM memory samples and was testing them internally. But the PIM compiler was still a simulation-based prototype, and DeepX was building hardware modules for physical testing. Source
First silicon, then the production ramp
EE Times reported on November 9 that samples were expected in March 2027, with mass production targeted for fall 2027. Yonhap’s November 10 account distinguishes initial production in the fourth quarter of 2027 from full-scale production in 2028. Bloomberg’s November 13 report describes customer evaluations followed by mass manufacturing in 2028. Those accounts may describe different stages, but they don’t explicitly reconcile the schedules. SourceSourceSource
Comments