Marvell to Begin Sampling Optical Memory Modules in 2027
The company plans to launch optical memory hardware that allows processors to share pools of data across server racks.
Updated on Oct. 20, 2026 in Semiconductors
Live Poll
Do you believe new AI hardware designs will significantly improve overall computing efficiency?
Marvell Technology is set to begin customer sampling of a new optical memory module in 2027. This system utilizes optical connections to enable processors to access shared memory across distances of up to 50 meters.
Why it matters
The architecture addresses memory capacity limitations by allowing processors to access larger shared pools when their individual capacity is insufficient. This could result in two to three times higher token throughput within existing data center space and power constraints.
The module configuration supports 7.2 terabits per second of optical fabric bandwidth with memory access latency under 350 nanoseconds. Each module handles up to 2 TB of DDR5 and 72 GB of HBM3E memory.
The players
Marvell Technology
Marvell Technology is a semiconductor company that specializes in data infrastructure products for enterprise and cloud computing.
Celestial AI
Celestial AI is a technology firm specializing in optical interconnect and memory hardware that was acquired by Marvell in 2026.
The details
The system features an appliance design using 16 modules to distribute memory among 16 hosts. It is designed for rack-to-rack sharing over distances up to 30 meters, leveraging the broader 50-meter reach of the architecture.
Timeline
February 2, 2026: Marvell acquired Celestial AI.
August 2026: Shared memory architecture was introduced.
October 6, 2026: Marvell displayed the module at an investor day.
2027: Customer sampling of the optical memory module is scheduled.
The Tech Race
This development represents a shift toward optical interconnects in data centers to replace traditional electrical copper connections that face reach and bandwidth limits. Marvell is positioning this hardware to outpace the scaling limitations of existing electronic memory architectures.
Enterprises and cloud providers could see significant performance improvements for compute-heavy tasks like AI processing without requiring additional data center floor space. By increasing token throughput, these modules may improve the efficiency of large-scale computational workloads.
The takeaway
The move to optical memory suggests that physical distance will become less of a constraint for high-performance computing clusters in the near future. Data centers are likely to prioritize these modular, high-bandwidth architectures as demand for memory-intensive AI models continues to grow.
Further reading
For more information on the current landscape of hardware architecture, see United States Semiconductors.
Source note: This article includes information reported by TokenPost.
Live Poll
Do you believe new AI hardware designs will significantly improve overall computing efficiency?
