As AI moves into robots, vehicles and wearables, SK hynix argues that memory design matters as much as raw computing power. The company’s latest article, featuring University of Illinois Urbana-Champaign professor Nam Sung Kim, explains why local processing and moving data efficiently matter when power, heat and response time are tight.
SK Hynix Watch analysis
What happened
In the fifth part of its AI ecosystem series, SK hynix sets out how constraints change when AI leaves the data centre. A robot that must adjust its movements in near real time cannot always wait for a remote server; devices also have less room for batteries, cooling and large memory systems.
The article discusses several ways designers could respond: smaller models through quantisation, memory lookup approaches such as Engram, and hardware options including processing-in-memory, 3D-stacked DRAM and Compute Express Link. It also describes High Bandwidth Flash as a potential memory layer between high-bandwidth memory and SSDs. These are approaches and technologies in discussion, not a claim that one finished design solves the problem. Read SK hynix’s article.
Why it matters
For physical AI, a pause in data delivery can become a pause in a real-world decision. Keeping processing close to where data is generated may help with latency and connectivity, but it puts pressure on memory capacity, bandwidth, heat and power consumption too. The architecture has to make room for all of them, which is less glamorous than a faster chip headline and rather more consequential.
Our read
The useful point is that AI performance is not just a contest in compute. Memory location and data movement can shape what a device can do, especially when it must act locally. SK hynix has a commercial interest in memory being central to that story, but the engineering constraints it describes are concrete. Treat its list as a map of approaches under discussion, not a forecast of which one will win.
What to watch
- Whether physical-AI devices increasingly process sensor data locally rather than relying on remote services.
- How designers balance memory capacity and bandwidth against power, heat and device size.
- Whether approaches such as processing-in-memory or High Bandwidth Flash appear in specific products, with measurable benefits.
Discussion spark: For robots and other devices that must react quickly, should designers prioritise local memory and processing even if that raises cost and complexity, or lean on remote computing wherever connectivity allows?
Sources and evidence
- [AI Ecosystem] Completing AI: Physical intelligence and the role of memory (16 November 2026, 23:58 UTC)
Independent WittyWires tracker for public updates about SK Hynix. Not affiliated with or endorsed by SK Hynix; this is not an official account.