DEEPX and AMD Explore Edge AI Cooperation After Lisa Su Meeting

South Korean fabless chipmaker DEEPX said Nov. 3 that it discussed potential technology cooperation with AMD, putting its low-power neural processing units on the table as candidates for integration with AMD's Kria system-on-module platform and ROCm software stack.

The talks moved from conference room to headquarters in less than a month. Keith Strier, AMD's senior vice president of global AI markets, and Jang Jung-hee, head of AMD Korea, visited DEEPX's offices in Pangyo on Nov. 3. There they met DEEPX CEO Kim Nok-won to discuss collaboration in two fields the source names precisely: physical AI and heterogeneous computing.

What did DEEPX propose?

DEEPX put two concrete ideas in front of its visitor. Neither involves money yet; both involve architecture.

• NPU plus Kria. DEEPX proposed integrating its low-power neural processing unit with AMD's Kria system-on-module platform. The aim is a single module capable of handling real-time robotic control and AI inference at the same time — a combination that matters for robots, where sensing and decision-making must run concurrently rather than in sequence.

• ROCm plus DEEPX's SDK. The second proposal links AMD's ROCm software platform, built for graphics processors, with DEEPX's own software development kit. The stated goal is a computing environment that distributes workloads across central processors, graphics processors and neural processors according to what each does best.

The second idea is the more strategically interesting one. Heterogeneous computing — dividing tasks among CPU, GPU and NPU by strength rather than defaulting to one general-purpose chip — has become the standard answer to power constraints at the edge. A vendor that ships both the silicon and the software plumbing to route workloads stands a better chance of being designed in.

How did the talks start?

The chain of meetings is short and unusually specific for a cooperation discussion at this stage. Discussions between the two companies began at the AMD-Korean AI Semiconductor Companies Roundtable, hosted by the Ministry of Science and ICT on Oct. 7. AMD CEO Lisa Su met with DEEPX CEO Kim at that event.

AMD Korea representatives then visited DEEPX on Oct. 21, followed by the Nov. 3 visit by Strier and Jang. A second visit within two weeks, this one including AMD's senior vice president for global AI markets, suggests the discussions have moved past introductory courtesies.

What is at stake for each side?

For DEEPX, the appeal is clear. AMD's Kria modules already serve embedded and robotics customers, and ROCm gives the company a credible software ecosystem in accelerated computing. A partnership would place a small Korean fabless supplier inside a much larger platform's distribution channel.

For AMD, the logic runs the other way. DEEPX specializes in low-power NPUs — exactly the class of silicon that large, power-hungry GPUs do not cover well. Pairing the two lets AMD address edge workloads where thermal budgets rule out discrete graphics processors, without building that silicon itself.

Is this a deal?

Not yet. The Nov. 3 announcement describes discussions and proposals, not an agreement, and neither company has disclosed terms, timelines or product commitments. That distinction is worth keeping in view: corporate meetings are cheap, and roundtable introductions hosted by a government ministry do not guarantee procurement.

What the episode does confirm is intent on both sides, and a cadence — three contacts between Oct. 7 and Nov. 3 — that is faster than diplomatic courtesy usually requires. Whether the Kria-plus-NPU combination or the ROCm-SDK link graduates from proposal to product will say more about AMD's edge AI strategy than any roadmap slide.