Data center controls
 Products > MemoryAI™ Photonic Appliance

The Industry’s First Photonics-based HBM-class Memory Appliance for Agentic AI

The new Penguin Solutions MemoryAI™ Photonic Appliance, built on the Marvell Photonic Fabric™ technology platform, introduces a step-function change in technology and architecture for shared memory.

Let's Talk
Read Press Release

Extend High-Bandwidth Memory Beyond the GPU

As agentic AI workloads grow more demanding, memory capacity can become a constraint on AI performance and scale.

Using silicon photonics, which has significantly higher bandwidth than traditional copper interconnects, the memory appliance will enable transparent sharing of disaggregated high-bandwidth memory (HBM) across multiple CPU- and GPU-based servers, providing greater flexibility to scale memory capacity for AI workloads.

HBM provides the high-speed memory access needed for demanding AI workloads, but GPU-attached HBM is constrained by the physical limits of the GPU package. The MemoryAI Photonic Appliance will extend access to memory at HBM-class data rates across a photonic memory fabric, delivering terabytes-per-second data rates and alleviating memory wall constraints.

Designed to complement GPU-attached HBM, the appliance will expand available capacity with shared HBM and DDR5 memory that can be accessed across AI infrastructure. Applications will be able to access this shared memory as if it were directly attached, without specialized access routines or data copying.

AI on chip

A More Flexible Approach to AI Memory

Data center
By disaggregating memory from individual processors, organizations gain more flexibility in how memory capacity is accessed, shared, and allocated across an AI data center.

Scale Memory Beyond the GPU

Expand available memory capacity beyond the physical limitations of directly attached HBM to support growing agentic and inference workloads.

Share Memory Across AI Infrastructure

Enable CPUs and GPUs to transparently access shared memory, creating greater flexibility in how memory resources are used across AI infrastructure.

Allocate Memory Based on Workload Needs

Support static or dynamic memory allocation, making memory capacity available where workloads need it and enabling a more efficient memory architecture.

Preserve HBM for Active AI Workloads

Complement GPU-attached HBM with additional shared memory capacity, preserving high-speed GPU memory for active token generation as inference workloads scale.

A placeholder Image

Speak with the
Memory Experts

Availability: The MemoryAI Photonic Appliance is expected to be generally available in the second half of calendar year 2027.

Features and specifications are subject to change.

Memory Innovation Designed for Agentic AI

Penguin Solutions brings 30 years of experience in advanced memory to today’s growing AI infrastructure demands. In our AI Factory Foundations workshop, our experts work with your business and technical teams to identify the AI opportunities worth pursuing and develop a practical roadmap for bringing them into production.

Schedule a discovery conversation with our team to learn more about our MemoryAI Photonic Appliance and discuss your current AI initiatives, business priorities, technical environment, and where we can help accelerate your path to agentic AI.