Penguin Solutions and Marvell Announce a Photonic Memory Appliance for 2027
Penguin Solutions and Marvell announced the MemoryAI Photonic Appliance at SC26. The proposed memory system connects to GPU servers over fiber, giving systems a separate pool for workloads that need more capacity than fits in onboard GPU memory.
What was announced
The companies describe up to 16 TB of DDR5 and 576 GB of HBM3E per appliance and 57.6 Tbps of aggregate bandwidth. These figures describe different memory tiers and an appliance-level bandwidth target. They are not evidence that a GPU can access all of the memory at that rate, or with the latency of local HBM.
Availability and evidence
Customer samples are expected in the second half of 2027. The original coverage found no pricing or independent performance tests. Remote memory adds network and integration costs that depend on how the workload accesses data.
When to evaluate it
The architecture is especially relevant to workloads under pressure from model weights, long contexts, or growing inference caches. Look for measured latency, bandwidth per host, utilization, and total cost on a representative workload before planning procurement.