Marvell and Penguin Solutions are developing a photonic memory system that lets multiple servers share large pools of DRAM and high-bandwidth memory. The potential prize for AI operators is keeping large inference caches in faster, shared memory rather than repeatedly rebuilding them or pushing them to slower storage.
Marvell Technology Watch analysis
What happened
The Register reports that Penguin’s first systems are expected in the second half of 2027. Each appliance is planned to support up to 16 servers, with as much as 16 terabytes of DDR5 and 1 terabyte of HBM. Marvell says its photonic fabric will provide 57.6 terabits per second of bandwidth across the appliance.
The system combines memory modules containing HBM3e and DDR5 with optical connections between servers. Penguin is building the appliance and its software layer. The Register says the design can expose the shared memory through several tiering modes, including one that uses HBM as a cache for DDR5.
Why it matters
AI inference produces key-value caches that preserve a model’s conversation state. Those caches can become large, and rebuilding them when a session resumes can be costly. Marvell says moving inactive caches from HBM to its memory appliance could improve performance and memory use compared with storage-based offload.
Pooling memory could also reduce how much DRAM each server needs. That is a substantial infrastructure pitch, though the first systems are still planned for next year and the performance benefits are company claims, not results established in the report.
Our read
This is an attempt to make memory a shared data-centre resource, rather than a fixed allowance attached to each machine. If the fabric delivers useful bandwidth and operators can manage the extra layer without adding headaches, it could give AI infrastructure teams another option for handling oversized inference caches. The hardware has a timetable; the proof will come from real workloads.
What to watch
- Whether Penguin delivers the first systems in the second half of 2027.
- Real-world performance and memory-use results for AI inference workloads.
- How easily the fabric works with existing server and memory configurations. Activity teaser: Marvell and Penguin Solutions are planning a photonic memory appliance that lets up to 16 servers share pools of DRAM and HBM. The proposed system could help AI operators manage bulky inference caches, but its first systems are not expected until the second half of 2027. The bandwidth and performance case is promising; real workloads will decide whether it is more than an elegant bit of plumbing.
Discussion spark: If shared photonic memory can make AI inference faster and use server memory more efficiently, would you trust a new fabric layer in your data centre before it has proved itself on real workloads?
Sources and evidence
- Marvell and Penguin Solutions aim to end the RAMpocalypse with photonic memory fabric – The Register (16 November 2026, 13:00 UTC)
Independent WittyWires tracker for public updates about Marvell Technology. Not affiliated with or endorsed by Marvell Technology; this is not an official account.