Penguin Plans Shared Optical Memory for AI Processors
Penguin Solutions plans to give AI processors access to a shared supply of memory through optical connections.
Published November 17, 2026Intelligence / Frontier Tech / Experimental Systems

Penguin Solutions plans to give AI processors access to a shared supply of memory through optical connections. Unveiled November 16, its MemoryAI Photonic Appliance is designed to move some memory meters away from GPUs while supporting high data rates. The company plans selected-customer sampling in the second half of 2027. Penguin announcement
The intended benefit is more room for demanding inference workloads without tying every increment of memory capacity to another processor. Marvell and Penguin envision memory shared across racks, allowing computing capacity and memory capacity to grow more independently. Marvell collaboration
Long conversations give that idea a practical purpose. AI systems retain intermediate calculations in a key-value, or KV, cache so they can reuse earlier work. As conversations accumulate, retaining that cache competes for limited fast memory. Discarding it can mean computing the same information again when a conversation resumes. Research background
Penguin proposes shared HBM and DDR5 memory as an additional tier alongside GPU-attached memory. Its product description says applications would access the pool as though it were directly attached, with capacity assigned statically or dynamically. The goal is to preserve the GPU’s local high-bandwidth memory for active token generation. Product design
Moving memory beyond the package
The announced design targets 57.6 terabits per second of total appliance bandwidth, with capacity for up to 16 terabytes of DDR5 and 576 gigabytes of HBM3E. That bandwidth figure describes the appliance total, rather than the rate available to each GPU. These are design targets. Penguin announcement
Penguin lists three access modes: direct HBM access, DDR5 access with HBM acting as a cache, and direct DDR5 access. CPUs would reach the shared pool through CXL 3.0 connectivity and Marvell photonic network interface cards. Penguin announcement
The partnership also includes a nearer-term, separate product: add-in cards that connect reused DDR4 modules to servers through PCIe 5.0 and CXL 2.0. Marvell expects those cards to reach qualification and enterprise deployment by the end of 2026. Their server-level expansion differs from the optical appliance’s planned multi-rack sharing. Marvell collaboration
What the research shows
A July preprint by Marvell researchers studies a related photonic-CXL design with 32 terabytes shared across 16 hosts. Its simulation reported a 6.6-fold improvement in time to the first token at 300 concurrent conversations, compared with a baseline whose cache capacity was exhausted. That study used a different architecture from Penguin’s announced appliance; the result belongs to a modeled workload, with end-to-end validation on physical hardware still pending. Preprint
Commercial timing also needs clarification. Penguin’s product page calls the second half of 2027 a general-availability window, while the announcement specifies sampling to selected customers. The release therefore provides the more conservative milestone for planning. Product page Announcement
For operators, the eventual decision will turn on how much useful inference the extra memory enables at its installed cost. The research identifies cost comparisons and real-hardware validation as further work, making customer trials an important next test. Preprint
Sources
Penguin Solutions announcement, November 16
Sources · 4
- https://ir.penguinsolutions.com/news/news-details/2026/Penguin-Solutions-Unveils-Industrys-First-Photonics-Based-HBM-Class-Memory-Appliance-to-Break-the-GPU-Memory-Wall-for-Agentic-AI/default.aspx
- https://www.marvell.com/company/newsroom/marvell-penguin-solutions-data-center-ai-memory-inference.html
- https://arxiv.org/pdf/2607.27187
- https://www.penguinsolutions.com/en-us/products/photonics-based-hbm-class-memory-appliance