Marvell Technology has laid out its most detailed picture yet of a light-based memory system that could free AI servers from one of their tightest constraints: memory that has to sit right next to the processor. At this week's PECC Summit, the chipmaker detailed a rack-scale appliance that pools up to 32 terabytes of DDR5 memory across as many as 16 servers or AI accelerators, connected not by copper but by optical links.
The system, called the Photonic Fabric Memory Appliance, is built from 16 Photonic Fabric Memory Modules linked through a passive fiber shuffle that exposes a single shared address space to every attached machine. Alongside the DDR5 pool sits 1 terabyte of HBM3e, which can act as a cache for the larger pool.
It is a direct attack on the memory wall. Copper wires lose signal quality at the speeds AI chips demand, which is why high-bandwidth memory has traditionally been bolted within millimetres of the processor — Marvell says the usual electrical limit is about 1.5 millimetres — capping how much memory any one chip can reach.
Memory, untethered from the processor
Marvell's answer is to move data as light. Each module pairs up to 2 terabytes of DDR5 with optical bandwidth rated at 7.2 terabits per second, and the company says memory can sit up to 50 metres from the processor, with 25 to 30 metres the practical reach for designs without signal retimers. The company points to latency claims of under 200 nanoseconds.
The appliance is rack-mountable, holds 128 standard DDR5 DIMM slots, and attaches either via a PF-NIC card using PCIe Gen6 and CXL 3.x, or via a PF-Chiplet integrated directly into an accelerator package. The pitch is simple: memory should scale independently of compute. If a workload needs more memory, operators add memory — not another expensive processor they did not otherwise want. Marvell executive vice president Will Chu has put it this way: "As AI scales, memory must scale more independently of compute."
Built for the KV-cache bottleneck
The immediate target is the key-value cache, or KV cache — the data AI models keep on hand so they can reuse earlier context without recomputing it. As reasoning models and agentic applications sustain longer conversations, those caches swell, and when they outgrow local memory they are evicted or pushed to slower SSD storage, dragging down throughput. A shared pool keeps reusable cache closer to compute and lets inference software move work to whichever processor is free. The company projects up to two to three times the token throughput within the same data centre space and power limits.
At the PECC Summit, senior director Ravi Mahatme went further, presenting simulation-based analysis suggesting roughly six times as many concurrent conversations at a comparable time to first token with the same compute resources. That figure is workload-specific modelling, not a measured production benchmark, and the company labels its latency targets as pre-silicon estimates.
Still pre-silicon — sampling in 2027
The appliance is not yet hardware customers can buy. Marvell outlined the broader architecture in August, presented the module at its October 6 investor day, and plans to sample it with multiple customers in 2027. The company has not named those customers.
The technology comes from Celestial AI, the optical-interconnect startup Marvell acquired in a deal completed on February 2, 2026, valued at $3.25 billion upfront. Nvidia is hedging with its own context-memory and co-packaged optics work, and electrical CXL pooling remains an alternative path. Whether light-based shared memory becomes a standard tier of AI infrastructure will depend on what those 2027 samples prove in real clusters.



