The Pulse
Marvell plans AI memory and networking demos at SC26
Marvell will demonstrate memory expansion, optical interconnect and networking technologies at Supercomputing 2026. The display includes a Photonic Fabric memory module and a simulator using Teralynx T100 switch silicon.
Marvell will demonstrate optical interconnects, memory expansion and networking hardware for AI infrastructure at Supercomputing 2026. Its November 13 announcement focuses on connecting compute and memory across racks as workloads place more demand on both.
The conference runs November 15–20 at McCormick Place in Chicago. Marvell says connectivity and memory constraints are leaving compute and memory-intensive infrastructure underused, and is presenting technologies aimed at both.
Memory across racks
At the show, the company plans to display a Photonic Fabric memory module alongside CXL memory products, including a Structera add-in card and near-memory compute and expansion or compression solutions. Partner demonstrations will feature Structera X memory expansion controllers with DDR4 and DDR5 memory.
The company says optical interconnects allow memory and compute to scale together across multiple racks. Its approach covers scale-up connections, scale-out networks using switches, pluggable optics and electrical interconnects, and scale-across connections between data centers using coherent pluggables.
In an August announcement, Marvell described its Photonic Fabric memory modules, network interface card and chiplets as parts of a multi-rack optical shared-memory architecture for AI inference. That architecture creates a shared-memory tier across multiple XPUs and racks up to 50 meters, with up to 32TB of warm KV cache offload, according to the company.
The company also claimed the approach delivers up to 2–3 times higher token throughput within existing data center footprints and power envelopes. It said the design allows KV cache to be loaded from a shared-memory tier rather than storage.
Switching and storage
The demonstrations will also cover Bravera solid-state drive controllers and OCTEON data-processing-unit-based storage solutions. In August, Marvell said its Bravera SC6 PCIe 6.0 SSD controller was expected to begin sampling in the fourth quarter of 2026 and supports NAND from multiple suppliers.
A separate demonstration will use Teralynx T100 switch silicon in a data center simulator. A virtual rack will feature Teralynx T100 switch trays from original design manufacturers and CXL solutions, while another demonstration will show multipath reliable connection using Teralynx 10 switch silicon.
The company introduced the 102.4 Tbps Teralynx T100 in June for scale-out and scale-up deployments. It supports up to a 512-port radix for scale-out use and a range of interconnect standards and emerging scale-up fabric protocols, according to the earlier announcement.
The display will include 1.6T and 800G active electrical cables based on Marvell Alaska A AEC digital signal processors. Marvell also lists its coherent pluggables for front-end and back-end connections between data centers as part of its scale-across offering.
Conference panel
Preet Virk, senior vice president and general manager of Marvell’s Photonic Fabric business unit, will participate in a conference panel on photonics for AI and high-performance computing clusters. The session is scheduled for November 18 from 12:15 to 1:15 p.m.