Marvell's optical memory module to deliver 1 TB HBM3e and up to 32 TB DDR5 in shared memory pool
The company's Photonic Fabric Memory Module is designed to connect HBM and DDR memory optically to accelerators, with a memory pool spanning 16 hosts. Every performance figure comes from Marvell itself, relayed through a single secondary source — and between the 50-meter headline and practical designs of 25–30 meters lies an important caveat.
Marvell has put forward a move that could change where AI memory can physically sit: the company's Photonic Fabric Memory Module uses optical links to let HBM and DDR memory operate up to 50 meters away from a GPU or XPU, according to the company. By comparison, the usual electrical limit is around 1.5 millimeters — a difference that explains why today's memory has to sit squeezed right next to the processor.
The technology was first presented in August 2026 at the industry events FMS and Hot Interconnects. Marvell elaborated in mid-to-late October 2026, including through an interview with theCUBE and the company's investor day, reported on October 20, 2026. This is thus a fresh case, and the technology has just been framed as part of Marvell's AI data center push.
What Is Actually New
The core of the story is not a new memory type, but a new way of connecting memory. Today, HBM — high-bandwidth memory — is soldered directly onto the same package as the accelerator, because electrical signals cannot travel very far. Marvell's module converts the connection to optics, allowing the memory to sit in its own rack or a separate cabinet.
The Numbers — and Who Owns Them
All key figures in this story are company-stated specifications, not verified measurements:
- 50 meters is the outer limit for reach. According to the source, practical designs built without retimers target 25–30 meters — a substantially lower and more grounded limit than the headline suggests.
- Latency is said to be under 200 nanoseconds — Marvell's own claim.
- The memory pool can span 16 hosts, with roughly 1 TB of HBM3e alongside 16–32 TB of DDR5.
- Dave Lazovsky, executive vice president at Marvell, has claimed the technology can lift inference throughput by 2–3x — but only "in certain contexts" that the source does not specify. We therefore cannot say which scenarios this applies to, and it would be pure speculation to guess.
Why the 1.5 mm Limit Has Shaped AI Infrastructure
The electrical limit of around 1.5 mm is the reason HBM today must be placed tightly next to the GPU. It limits both how much memory each accelerator can have and how flexibly the data center can be organized: the memory follows the chip, not the other way around. Optical breakout turns this on its head — in principle, the memory can sit wherever space and cooling allow, and multiple hosts can share a common pool. This is what Marvell points to when the company highlights pooling across 16 hosts.
Who It Affects
The target audience is hyperscalers and AI infrastructure operators who need more memory per accelerator than today's HBM capacity allows, or who want to use memory more efficiently across machines. No independent benchmarking or signs of adoption appear in the source — there are as yet no confirmed customers.
Open Questions
The source says nothing about whether the module is in production, in sampling, or merely a demonstration, and no availability date is given. It is also unclear how the latency and throughput gains hold up outside controlled environments. And "2–3x in certain contexts" remains undefined until Marvell specifies which workloads underlie it.
Verification Status
It is worth being explicit: every performance figure in this story — 50 meters, latency under 200 ns, the pool sizes, and the 2–3x gain — originates from Marvell and is relayed through a single secondary source, Cryptobriefing (source ID b2af35c0-0d8f-4742-9b89-97f7747686a7). No primary source from the company or independent confirmation exists yet. Until such measurements appear, the 50-meter claim should be read for what it is: a marketed outer limit, not a measured property.