Penguin Solutions unveiled its MemoryAI Photonic Appliance at SC26 in Chicago on 16 November 2026, alongside a collaboration with Marvell on AI memory expansion. The product is intended to let CPUs and GPUs share memory across racks using optical connections. Sampling is planned for the second half of 2027, making this a development roadmap for research computing centers rather than an immediately deployable upgrade.
A named appliance follows Marvell’s technology preview
Penguin’s 16 November announcement, distributed by Business Wire, identifies a specific product built around Marvell Photonic Fabric technology. It is intended to supplement GPU-attached high-bandwidth memory with a shared pool located farther away. Penguin says applications would reach that pool without special access routines or explicit data copying.
The product announcement and dated sampling plan are the new development beyond CASRAI’s earlier coverage of Marvell’s SC26 interconnect preview. The two companies have also announced a broader partnership covering both optical pooling and conventional server-memory expansion.
What Penguin is targeting
Penguin gives the appliance a target total bandwidth of 57.6 Tbps and a design capacity of up to 16 TB of DDR5 RDIMM memory plus 576 GB of HBM3E per appliance. Those are announced design targets, rather than independently measured results. The planned access modes include direct HBM, direct DDR5 and an arrangement using HBM as a cache for DDR5.
The company describes static and dynamic allocation to both CPUs and GPUs. Its stated sampling window is the second half of 2027, for selected customers. That timetable distinguishes the unveiling from general commercial availability.
The partnership has two different delivery tracks
In a separate 16 November statement, Marvell describes two complementary approaches. The first uses its Structera X controllers in Penguin SMART CXL add-in cards to reuse DDR4 modules from decommissioned servers. This is direct memory expansion through PCIe 5.0 and CXL 2.0. Marvell expects those cards to be available for qualification and enterprise deployment by the end of 2026.
The second approach connects memory appliances and compute systems across multiple racks using Photonic Fabric modules and network interface cards. Marvell presents that as a route toward larger shared pools whose capacity can grow separately from computing resources. Its statement places sampling of the jointly developed photonic memory-expansion servers in the second half of 2027.
Jihoon Han, a memory analyst at Yole Group quoted in Marvell’s release, frames expansion and pooling as ways to make better use of available memory while manufacturing capacity takes time to increase. This is an attributed industry assessment, rather than evidence that the new appliance has already delivered savings.
Independent reporting highlights the latency tradeoff
HPCwire’s Alex Woodie interviewed Marvell executive Preet Virk about the architecture. Virk described optical links between racks and a unified address space, while acknowledging that remote memory access adds latency compared with directly attached HBM. HPCwire reports his comparison of roughly 190 nanoseconds for the photonic mesh against about 90 nanoseconds for directly attached memory.
Investing.com’s separate product report likewise describes an appliance still headed toward customer sampling. Neither announcement establishes that more memory automatically improves every research workload. Capacity, bandwidth and latency remain different properties to assess.
What research computing centers can evaluate
The following is editorial analysis of the announced architecture. A university planning accelerator purchases can separate three questions: whether its workloads exceed local memory capacity, whether they can use shared capacity efficiently, and whether additional access latency changes completion times. A useful evaluation would compare the same model, context length and concurrency under each memory configuration.
For teams pursuing reproducible AI experiments, those settings belong alongside model and software versions in the benchmark record. Comparing an appliance capacity figure with a GPU’s local memory size alone would leave the workload behavior unresolved. The relevant outcome is whether the proposed configuration helps a defined research task under its actual constraints.
Research infrastructure proposals can also distinguish the nearer-term DDR4 expansion track from the later photonic appliance. They have different interfaces and qualification schedules. Documenting those dependencies in the computational environment would make an eventual pilot easier to reproduce and assess. The next concrete milestone in the companies’ roadmap is customer qualification for the CXL cards, followed by selected-customer photonic sampling in 2027.
Sources
- Penguin Solutions via Business Wire, MemoryAI Photonic Appliance announcement, 16 November 2026. Checked 2026-11-16.
- Marvell, Marvell and Penguin Solutions Collaborate to Address Data Center AI Memory and Accelerate Inference at Scale, 16 November 2026. Checked 2026-11-16.
- HPCwire, Alex Woodie, Marvell and Penguin Launch Photonics Memory Appliance, 16 November 2026. Checked 2026-11-16.
- Investing.com, Penguin Solutions unveils photonics-based memory appliance, 16 November 2026. Checked 2026-11-16.







