Marvell reveals optical fabric for sharing 32TB across 16 GPUs

Marvell has shown a Photonic Fabric Memory Module designed to let AI systems share large pools of memory across multiple GPU hosts. The company’s optical interconnect platform is aimed at scale-up infrastructure, where moving data between accelerators is becoming as important as the processors themselves. ¹²

What Marvell is showing

The module is part of Marvell’s Photonic Fabric platform, which the company says can use light-based links to connect compute resources across a wider physical area than conventional in-rack designs. Reporting on the demonstration describes a proposed 16-module appliance sharing 32TB of DDR5 and 1,152GB of HBM3e across 16 hosts. Those figures describe the demonstrated architecture, not a consumer product available to buy. ¹

Marvell’s own technology page describes Photonic Fabric as an optical scale-up network for accelerated computing. It combines photonics, switching, packaging and software to move towards a more unified pool of high-bandwidth memory for AI data-centre workloads. The company says the platform is intended to extend scale-up domains beyond a single rack while keeping latency and power under control. ²

Why optical memory links matter

AI accelerators can be limited not only by compute but by how quickly they can exchange model data. Keeping more memory close to a single processor is expensive and difficult to scale, while sending data through longer conventional paths adds latency and consumes power. An optical fabric offers another route: distribute the memory while using high-bandwidth light-based connections between the hosts.

The idea is especially relevant to inference systems, where demand for memory capacity can rise faster than the number of accelerators. Marvell says its platform can support high-capacity, in-network memory and scale across multiple racks. TechRadar’s report also frames the module as a way to reduce the capital pressure of building larger AI clusters, although those savings will depend on pricing, software support and how the complete system performs in production. ¹²

This is infrastructure rather than a new graphics card. The module will matter first to cloud providers, AI laboratories and system builders that need to combine many accelerators. The harder questions are practical ones: how reliably the shared memory behaves under real workloads, how much optical switching costs, and whether customers can integrate it without redesigning their entire stack.

Our opinion

Marvell’s demonstration is interesting because it treats memory as a network resource instead of something permanently tied to one accelerator. That could become a useful way to make AI systems less wasteful, but a diagram and a trade-show demonstration are not the same as a shipping platform. The 32TB headline is compelling; the eventual test will be whether software can make that capacity feel local enough to justify the hardware.