AIINEWS

AIINEWS

Global intelligence on artificial intelligence and computing

Marvell says memory transactions exceed 80% of scale-up AI traffic

AIINEWS Editorial Board•hardware
19 November, 2026•2 min read
Marvell says memory transactions exceed 80% of scale-up AI traffic

Executive Briefing

“A Marvell executive says memory transactions account for more than 80% of scale-up AI...”

Key Takeaways

  • 1.Penguin’s Andy Mills described an appliance that places memory tens of meters from GPUs and connects it through Marvell’s optical interconnect. Each appliance contains a total of 16 TB of HBM and DDR5 memory for servers across multiple racks to share.
  • 2.Preet Virk, who leads Marvell’s Photonic Fabric unit, said memory transactions make up more than 80% of scale-up network traffic. He argued that shared memory could keep much of that data from moving.
  • 3.Virk estimated that the approach could use about one-quarter as much energy per bit moved as copper-based data movement. That comparison concerns data-transfer energy, not total data-center power consumption.
  • 4.The interview appeared in a theCUBE program at SC26. SiliconANGLE disclosed that Penguin sponsored the segment and said sponsors had no editorial control.

A Marvell executive says memory transactions account for more than 80% of scale-up AI network traffic. Marvell and Penguin Solutions are presenting an optical connection that lets servers share memory, aiming to reduce data movement.

Product Features & Key Highlights

Penguin’s Andy Mills described an appliance that places memory tens of meters from GPUs and connects it through Marvell’s optical interconnect. Each appliance contains a total of 16 TB of HBM and DDR5 memory for servers across multiple racks to share.

Preet Virk, who leads Marvell’s Photonic Fabric unit, said memory transactions make up more than 80% of scale-up network traffic. He argued that shared memory could keep much of that data from moving.

Virk estimated that the approach could use about one-quarter as much energy per bit moved as copper-based data movement. That comparison concerns data-transfer energy, not total data-center power consumption.

The interview appeared in a theCUBE program at SC26. SiliconANGLE disclosed that Penguin sponsored the segment and said sponsors had no editorial control.

Practical Assessment & Evaluation

A shared pool could let servers access the same data without each keeping its own copy. Its practical value depends on whether reduced data movement outweighs the latency of accessing remote memory.

The more-than-80% traffic share and the one-quarter energy estimate describe different measures. Neither establishes how much power an entire AI system would save.

Pricing, Availability & Next Steps

The source provides no independent measurements, test configuration or methodology for Virk’s traffic and energy figures.

Latency, deployment cost and performance when multiple servers run AI workloads concurrently remain unclear.

Independently verified, edited, and published by AIINEWS Editorial Board
3-tier verified
Editorial standards
Share this intelligence report:
Expanded Perspectives & Intelligence

Related Dispatches & Deep Dives

Curated 8 specialized reports

Explore AIINEWS Intelligence Vault

Over 2,000 in-depth dispatches structured across 18 specialized tech & AI categories.

Browse all 18 categories