SK hynix's "Roadmap for Light-Based AI Memory" Is a Review Paper in Nature Electronics

SK hynix published, with five universities, a review paper in Nature Electronics titled "Co-packaged optics for high-performance computing and artificial intelligence." The paper went online on 19 August; the company's newsroom post carries 20 August KST. Headlines have rendered it as SK hynix unveiling a roadmap for light-based AI memory.

What the common telling gets wrong

There is no product here, and no schedule and no customer. This is a co-authored academic review article whose corresponding authors are split between Seunghoon Hong, SK hynix's AI Infra team lead, and Professor Kyusang Lee of the University of Virginia. The collaborating institutions are UVA, UIUC, Nanyang Technological University, MIT and Yonsei. No process node, no capacity, no launch date and no partner silicon appears anywhere.

The figures are targets, not specifications

The paper states goals of more than 100 Tb/s of bandwidth per node, energy below 1 pJ/bit, and chip-to-chip latency of less than 10 nanoseconds. These are numbers the field must reach for the approach to work — not measurements SK hynix has demonstrated. Quoting them as product specifications, as several write-ups have, inverts their meaning.

The problem being described

The bandwidth wall, as the paper frames it: compute throughput has typically tripled every two years, while interconnect bandwidth has advanced only about 1.4-fold over the same period. The proposed path runs 2D, to 2.5D interposer-based co-packaged optics, to 3D heterogeneous stacking — and then extends optics to the memory interface.

What is actually new

Co-packaged optics is not novel; it already ships on network switches. The claim worth noticing is the extension of optics onto the memory interface, which nobody has productised. Read that way, this is the HBM leader stating publicly that it believes the electrical memory interface is running out of road, and planting a flag on the optical answer before Samsung and Micron do. That is a position, not a roadmap — and positions from the market leader are worth reading precisely because they are cheap to state and expensive to abandon.