Eigen RadarAI
Analysis

Marvell and Penguin plan shared memory across AI servers

Marvell and Penguin Solutions have announced a partnership to expand memory for AI inference. Their plans combine reuse of memory from retired servers with a later optical system that shares capacity across racks. The first expansion cards are expected by the end of this year, followed by sample systems using optical connections in the second half of next year.

Artificial Intelligence··Evening
Low-angle wide view of unbranded CXL memory-expansion add-in cards with copper heatsinks standing upright beside reclaimed DDR4 memory modules on a maple worktable in a sunlit hardware laboratory.

Servers will gain access to pooled memory

Marvell, a semiconductor company, and computing systems specialist Penguin Solutions announced a memory-development partnership on 16 November. It combines direct expansion of server memory with a later optical system for sharing capacity across multiple racks. The partners are addressing memory used during AI inference, when a trained model processes a request and generates an answer.[1], [2]

Larger models, longer inputs and growing key-value caches are increasing memory requirements, the companies say. Those caches retain information about a model’s processing state. The proposed expansion path separates the growth of memory capacity from computing capacity, so operators can add memory without increasing both resources together.[1]

Retired-server memory is part of the first stage

The first stage puts Marvell’s Structera X memory-expansion controllers into Penguin’s SMART CXL add-in cards. CXL is an interconnection technology for computing devices and memory. The partners intend to reuse high-density DDR4 memory modules from decommissioned servers. The cards could also form dedicated appliances that allocate pooled memory across multiple hosts. Qualification and enterprise deployment are expected by the end of this year.[1]

Optical-system samples are due in the second half of next year

The second stage would integrate Marvell Photonic Fabric optical links into Penguin’s MemoryAI appliances and ComputeAI systems, extending memory pools across racks. Jointly developed servers are expected to sample in the second half of next year. Yole Group memory analyst Jihoon Han says new memory production capacity takes time and substantial investment, while higher memory prices make AI infrastructure expansion more expensive.[1]

References

  1. News sourceMarvellMarvell and Penguin outline CXL and optical memory plans for AI↩1↩2↩3↩4
  2. News sourceThe RegisterMarvell and Penguin Solutions aim to end the RAMpocalypse with photonic memory fabric↩