Analysis
Memory modules and the cables between racks belong in the same capacity discussion. That is Marvell's argument in its November 13, 2026 preview of Supercomputing 2026: the company says growing AI workloads encounter memory and connectivity bottlenecks that leave hardware underused. More compute alone is not its proposed answer.
Marvell plans demonstrations of Photonic Fabric memory modules, Structera memory expansion using CXL connections, near-memory computing and network switching. It also lists active electrical cables rated at 1.6T and 800G. Those advertised cable rates are specifications, not measured throughput for a complete cluster running a customer's job.
The company describes optical links extending connected memory and compute across racks, alongside separate networking work within and between data centers. This remains a component preview. The announcement supplies no measured utilization gains, complete system power budget or installation cost. Its DDR4 and DDR5 memory demonstrations do not establish how every proposed combination will behave together.
For an operations team, the useful question is whether waiting on data is actually the expensive part of its workload. Ask the supplier to identify the stalled path and the configuration needed to improve it. Then test one representative job on the intended cluster, changing one path at a time and measuring wait time, throughput and power together. A conference demonstration can start that conversation; your workload has to finish it.
Source: Marvell Scales Connectivity to Address AI Infrastructure's Next Bottleneck at Supercomputing 2026 · 2026-11-13 · Primary source context · attributed account
Why it matters
Memory and connectivity need evaluating together, with advertised component rates kept separate from measured gains on a complete working system.
