Marvell Technology, Inc. plans to provide samples of its Photonic Fabric Memory Module to multiple customers in 2027, seeking to ease memory constraints as AI models and conversations grow.
In a technical paper, company researchers describe a proposed appliance built from 16 modules. It would give as many as 16 hosts access to 32 TB of DDR5 and 1.125 TB of HBM3E, extending the memory available beyond what is attached to each processor.
The design targets key-value caches, data kept for reuse while a model produces an answer. Larger models and longer conversations expand those caches, Marvell says. The company proposes optical links that would let processors in separate server racks draw on a common memory pool.
The module specification pairs as much as 2 TB of DDR5 with optical bandwidth rated at 7.2 terabits per second. Marvell proposes two roles for the HBM3E: it could serve as a cache for the DDR5 or add to the memory available in the shared pool.
The paper compares the proposal with earlier designs that pool memory using electrical CXL links. Marvell's optical approach is intended to give cloud operators access to shared capacity across racks when memory beside an individual processor is insufficient.
Marvell outlined the broader architecture in August. At its Oct. 6 investor day, it presented the module and set out its plan to sample it with multiple customers in 2027. The company did not identify those customers.
Marvell projects up to two to three times the token throughput within the same data center space and power limits. That is the company's expectation for the proposed architecture; the technical paper's evaluation does not establish that multiplier.
The researchers evaluated aspects of the design through hardware emulation and AI-serving simulations. They wrote that validation on a physical appliance remained pending when the paper was prepared. Customer sampling in 2027 is Marvell's next announced milestone, with no narrower timetable given.
