# Penguin Targets Up to 16 TB of DDR5 in Photonic AI Memory Appliance

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/60569
Published: 2026-11-18T15:13:46.000Z
Updated: 2026-11-18T15:13:46.000Z
Section: Technology

> The planned design also targets 576 GB of HBM3E and 57.6 Tbps of bandwidth, with select-customer sampling planned for the second half of 2027.

Penguin Solutions unveiled its planned MemoryAI Photonic Appliance on Nov. 16, a system designed to pool memory across CPU- and GPU-based servers for AI workloads that need more capacity.

The planned design targets up to 16 TB of DDR5, 576 GB of HBM3E and 57.6 Tbps of total bandwidth. It would use Marvell’s Photonic Fabric, CXL 3.0 connectivity and photonic network interface cards to make pooled memory accessible to both CPUs and GPUs.

The shared memory tier is intended to supplement HBM attached directly to GPUs, whose capacity is constrained by GPU package size. The appliance targets inference workloads that need more memory, including workloads that reuse key-value cache data.

The appliance is part of a broader [cross-rack memory-pooling project](<https://www.tokenpost.com/news/technology/59180>) designed to let data centers expand memory separately from compute and share capacity across racks. Penguin plans to sample the appliance with select customers in the second half of 2027.

Preet Virk, senior vice president and general manager of Marvell’s Photonic Fabric Business Unit, said memory transactions account for more than 80% of scale-up network traffic. He also said the photonic design could use one-quarter the energy per bit of copper-based data movement. Penguin sponsored the interview segment in which Virk made those claims.



## Links in this article

- [cross-rack memory-pooling project](https://www.tokenpost.com/news/technology/59180)
