• Skip to main content
  • Skip to secondary menu
  • Skip to footer

Technologies.org

Technology Trends: Follow the Money

  • Technology Events 2026-2027
  • Sponsored Post
  • Technology Markets
  • About
    • GDPR
  • Contact

Penguin Solutions and Marvell Bet on Photonics to Break the GPU Memory Wall

November 17, 2026 By admin

GPU memory has a geography problem. High-bandwidth memory has to sit within millimeters of the GPU die, which caps how much any one accelerator can have. Agentic AI and long-context inference keep asking for more, and once the KV cache outgrows HBM, expensive GPUs sit around waiting on memory.

Penguin Solutions thinks the answer is light. At SC26 in Chicago the company unveiled the MemoryAI Photonic Appliance, which it bills as the first photonics-based, HBM-class memory appliance. The same day, Marvell and Penguin announced a broader collaboration on AI memory, and the two pieces fit together.

The appliance

The box sits meters away from the GPUs and connects over silicon photonics, using CXL 3.0 and Marvell photonic NICs. Penguin says optics draw less power than copper over those distances. The targeted specs:

  • 57.6 Tbps of total appliance bandwidth
  • Up to 16 TB of DDR5 RDIMM and 576 GB of HBM3E per appliance
  • Three access modes: HBM Direct, HBM-DDR5 Cached and DDR5 Direct
  • Memory pooled between CPUs and GPUs, with static or dynamic allocation handled by Penguin software

Sharing is transparent, so applications don’t need special access routines or data copies. The appliance adds to the HBM already on the GPU rather than replacing it. Penguin is also showing a MemoryAI Photonic Memory Module built on Marvell’s Photonic Fabric. Sampling to select customers is planned for the second half of 2027.

“Photonics is the step-function change that makes this possible,” said Andy Mills, VP of advanced product development at Penguin.

The Marvell side

The wider collaboration runs on two tracks. The near-term one is unglamorous and probably useful sooner. Marvell Structera X memory-expansion controllers go into Penguin’s SMART CXL add-in cards, which reuse high-density DDR4 modules pulled from decommissioned servers. They connect over PCIe 5.0 and CXL 2.0, and the same card can build CXL memory appliances that pool DDR4 across several hosts. Penguin expects these cards ready for qualification and enterprise deployment by the end of 2026.

The longer track brings Marvell Photonic Fabric into Penguin’s MemoryAI appliances and ComputeAI systems, for memory shared across multiple racks in pools reaching petabytes. Those servers are due to sample in the second half of 2027. The pitch is a single upgrade path, from direct-attached expansion to multi-rack sharing, without re-architecting anything.

Old DDR4 getting a second life next to HBM3E is neat economics, given how tight memory supply is right now. Kash Shaikh, Penguin’s president and CEO, said memory is becoming increasingly critical to the performance and economics of AI factories. Preet Virk, who runs Marvell’s Photonic Fabric business, put it in terms of tokens: inference has to deliver them more cheaply before it can scale.

The photonic products are 2027 hardware, so treat the numbers as targets. The direction is clear enough. Memory is turning into its own tier of the data center, with its own boxes and its own network.

Filed Under: News Tagged With: AI memory, CXL, HBM, Marvell, Penguin Solutions, SC26, Silicon Photonics

Footer

Recent Posts

  • Saltgrain and NTT Research Show Post-Quantum Attribute-Based Encryption Fast Enough for a Laptop
  • Quantum Moves Into the Supercomputer Queue as IQM, Altera and Infleqtion Report Progress
  • Penguin Solutions and Marvell Bet on Photonics to Break the GPU Memory Wall
  • TensorWave Signs Six-Year Deal to Supply OpenAI With 457MW of AMD Helios Capacity
  • LangChain Raises $300 Million at a $3 Billion Valuation to Build Out LangSmith Agent Platform
  • OQFT Shows Where the Quantum Fourier Transform Can Get Shallower
  • MemryX MX3+ Targets 1 POPS for Edge AI as Tianji Shows a 21-Joint Robot Hand at CoRL and Appear Starts Solid-State Battery Pre-Production
  • Adobe MAX 2026 Opens Photoshop and Premiere to Outside AI Agents and Deepens Its Google Partnership
  • Zendesk Launches Outcome AI Platform and Drops Seat Limits From New Plans
  • Otto One Is a $599 Screenless Computer Built for an AI Agent to Live On

Media Partners

  • Market Analysis
  • BareProxy.com
  • App Coding
AI Market Analyzer Rises to 33.5 Cooling as CoreWeave Lifts Breadth and the Nvidia CEO Crisis Looms
AI Market Analyzer Slips to 30.4 Cooling as Chips Crash and Leveraged Bulls Keep Buying
Netlist (NLST) Revenue Jumps 149 Percent on Samsung License Cash as Silicon Labs (SLAB) Grows 22 Percent and QuickLogic (QUIK) Doubles
Omdia Says AI Software Revenue Will Top 200 Billion Dollars in 2026 as Generative AI Grows 139 Percent
Dell (DELL) Reports November 24 With a 95 Billion Dollar AI Backlog to Turn Into Cash as Supermicro (SMCI) Takes Its Record Quarter on the Road
CoreWeave (CRWV) Grows Revenue 168% to 3.66 Billion Dollars and Lifts Its Backlog to 132.5 Billion as the Neocloud Selloff Meets Real Numbers
Marvell vs Lumentum, Which AI Infrastructure Stock Is the Better Investment?
Lucid Grows Revenue 4% While the US EV Market Shrinks 41% and Adds 1.75 Billion Dollars of Saudi-Backed Credit to Reach Its Midsize Car
Entegris (ENTG) Targets 15 Percent Annual Growth and Onto Innovation (ONTO) Posts a Record Quarter as Chipmakers Struggle to Measure AI Returns
SovAI Joins NVIDIA Inception With Two Grid-Connected AI Factory Sites Built to the DSX Reference Design
BareProxy Adds AI Crawler Control That Tells Real GPTBot and ClaudeBot Requests From Fakes
BareProxy Adds Bot Challenges That Make Scrapers Posing as Browsers Solve a Puzzle First
BareProxy Adds Cookie Consent That Holds Back Tracking Cookies Until a Visitor Says Yes
BareProxy Adds Markdown Serving That Turns Markdown Files Into Web Pages With No Build Step
BareProxy Ships Its First Nine Plugins, From Maintenance Pages and CORS to Geo Rules and Rate Limits
BareProxy 0.2.0 Runs WebAssembly Plugins Through the Proxy-Wasm Interface
BareProxy 0.3.0 Writes Its Plugins in Rust, Where a Plugin Adds 15 Microseconds Instead of 42
BareProxy Is an Operating System for Web Traffic: A Small Core and a Long Stream of Plugins
BareProxy Plugins Go WebAssembly, Starting With RenderCache and AI Crawler Control
BareProxy Runs on Linux, macOS and Windows, and One Plugin File Runs on All of Them
Testing Optimistic Quantum Fourier Transform Circuits With OQFT and Qiskit
Bluehost AI Builder Turns a Prompt Into a Website or a Booking App and Takes the AI App Builder Fight to Hosting Customers
Claude Code on GitLab vs GitHub: Thinner Integration and a Fifth of the Free CI Minutes
How k4i.com Publishes Posts Unattended With Small AI Models, Groq and GitHub Actions
How to Ship In-Browser AI Apps With No Server and No Inference Bill
Microsoft Authenticator Android Flaw Shows Why Apps Must Verify Who Is Calling Them
Zero-Budget AI Apps Work Best as Self-Feeding Sites, Not Open Tools
A Useful Software Primitive Should Be Easy to Explain and Hard to Replace
Apple's US Link-Out Fee Is Zero Until a Judge Sets It, and Apple Is Asking for Up to 15%
20 Primitive Project Ideas: Tiny Go Engines That Do the Essential 10% of Redis, Kafka and Etcd

Media Partners

  • AltSql.com
  • Technology Conferences
  • API Coding
AltSql Is a Native Hybrid of Key-Value and SQL, From the Sensor to the Gateway
AltSql Core and AltSql DB Are Now Open Source Under the Apache License 2.0
Introduction
AltSql DB 0.3 Adds Secondary Indexes and Statement Savepoints: A Lookup on a Million Rows Ran 198 Times Faster Than a Full Scan
How It Works
AltSql DB 0.2: SQL and Direct Calls Wrote Byte-Identical Files Over 100,000 Random Steps
AltSql DB Keeps a Whole Fleet in One File and Reads by Key 5 Times Faster Than SQLite Through SQL
AltSql Core and Six Prototypes Built Around It
AltSql Ask Puts One SQL Question to 500 Simulated Machines and Brings Back Only the Answers
Machine Monitoring Case Study: Overheat Alarm Handled on the Sensor While Wi-Fi Was Down
Why People Still Attend Technology Conferences, and What Organizers Must Remember
SC26 Supercomputing 2026, November 15-20, McCormick Place, Chicago
MWC Doha 2027, November 15–17, Doha, Qatar
A3 Business Forum 2027, January 18-20, Signia by Hilton Orlando Bonnet Creek, Orlando
IAMPHENOM 2027, April 20-22, Pennsylvania Convention Center, Philadelphia
Channel Partners Conference and Expo 2027, March 15-18, The Venetian, Las Vegas
Fintech NerdCon 2026, November 18-20, San Diego Convention Center, San Diego
HLTH 2026, November 15-18, The Venetian Expo, Las Vegas
electronica 2026, November 10-13, Messe München, Munich
KIOXIA Forum 2026, November 26, TODA Hall and Conference Tokyo, Tokyo
RFC 9457 Problem Details Gives API Errors One Shape Instead of Fifty Ad Hoc JSON Bodies
Idempotency Keys: How One Header Stops a Retry From Charging a Customer Twice
A Tiny ETL Binary Competes With curl, jq and SQLite in a Cron Job, So Build It That Small
An AI Agent Flight Recorder Belongs in One Portable File, the Way HAR Did It for HTTP
Benchmarking MCP Servers and Stateful API Workflows: Latency, Throughput and Tokens per Call
Diagnose a Slow API From One Request: DNS, Connect, TLS, Server Wait and Transfer
Finding API Endpoints, Tables and Flags Nobody Uses Takes Code and Traffic Together
Fuzzing MCP Servers: Generate Bad Arguments From the Tool Schema and Watch What Breaks
Give Any API a History: Poll It, Hash It and Query Old Versions With SQL
Infer Your API's Real Contract From Traffic, Then Diff It Against the Docs

Copyright © 2026 Technologies.org

Media Partners: Market Analysis · Market Research · Referently · Photography