Penguin Solutions, Inc. has unveiled the MemoryAI™ Photonic Appliance, which it describes as the industry’s first photonics-based, HBM-class memory appliance. It is designed to address GPU memory constraints in agentic AI workloads.

Built on Marvell’s Photonic Fabric technology, the appliance connects via CXL 3.0. It uses silicon photonics to let CPUs and GPUs share a disaggregated pool of memory at high bandwidth over distances of meters, compared with the millimeter-scale limits of current GPU-attached HBM.

The appliance aims to address the “AI memory wall” by expanding available memory capacity and supplying data to memory-constrained GPUs. It is designed to complement existing HBM by offloading data, with the goal of improving accelerator utilization and efficiency for large-scale AI inference.

The system targets 57.6 Tbps of total bandwidth. Each appliance supports up to 16 TB of DDR5 memory. It also supports up to 576 GB of HBM3E memory.

Penguin is showcasing the appliance at the SC26 conference in Chicago. Sampling for select customers is planned for the second half of 2027.