2 min read
Add as a preferred source on Google

Marvell and Penguin Solutions Plan Cross-Rack Photonic Memory Pools

The design would pool memory across racks for AI inference. Sampling of the joint expansion server is expected in the second half of 2027.

Fiber-optic connectors rest beside a compact memory appliance / TokenPost.ai
Fiber-optic connectors rest beside a compact memory appliance / TokenPost.ai

Marvell and Penguin Solutions are developing a photonic memory system to pool capacity across data center racks, giving AI inference workloads a way to scale memory independently of compute.

The companies announced the project Monday, Nov. 16, as part of a two-track memory expansion effort. The rack-scale design aims to support low-latency, high-bandwidth access to shared memory and petabyte-scale pools for inference.

The companies plan to integrate Marvell Photonic Fabric-based memory modules and network interface cards into Penguin Solutions’ MemoryAI memory appliances and ComputeAI systems. The shared-memory approach is designed to let data centers add capacity separately from compute and make that capacity available across racks.

A separate server-level track pairs Marvell’s Structera X memory expansion controller with Penguin’s SMART CXL expansion card. The card is designed to reclaim high-density DDR4 modules from retired servers and connect them through PCIe 5.0 and CXL 2.0. It can also serve as a dedicated CXL memory device, allowing operators to pool recovered capacity and distribute it among multiple hosts. The card is expected to be available for qualification and enterprise customer deployment by the end of 2026.

The projects address rising memory needs as AI workloads expand. Larger models, longer context windows and growing key-value cache requirements have pushed demand to hundreds of gigabytes and, in some cases, terabytes per node. The two designs offer server-level expansion and multi-rack sharing as ways to increase memory without expanding computing capacity at the same pace.

The joint photonic memory expansion server project is expected to reach the sampling stage in the second half of 2027.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…