Marvell Uses Light to Create Huge GPU Memory Pools

Marvell is building a new system that uses light to let 16 GPUs share a massive 32TB pool of memory.

Semiconductor company Marvell has revealed a new architecture designed to solve one of the biggest bottlenecks in artificial intelligence: memory. According to reporting from TechRadar, the company is developing a Photonic Fabric Memory Module. This system uses light-based connections to allow up to 16 GPUs to share a massive, centralized pool of 32 terabytes of DDR5 memory and over a terabyte of high-bandwidth memory (HBM). The technology, which builds on designs from Celestial AI, is still in the early stages. Marvell is not expecting to provide samples to customers until 2027, but the announcement signals a major shift in how future AI data centers could be built.

The core innovation is memory disaggregation. In today's AI servers, expensive, high-performance memory is tightly attached to each individual GPU. This creates isolated pools of resources. If one GPU runs out of memory, it cannot easily borrow from a neighbor, leading to inefficiency and forcing developers to buy more hardware than they might otherwise need. Marvell’s photonic fabric breaks this model by creating a high-speed network that connects many GPUs to a shared memory appliance. Using light traveling through fiber optics instead of electricity over copper wires allows for extremely fast, low-latency communication over longer distances. This enables the system to trick the GPUs into seeing the shared, remote memory as if it were directly attached, effectively creating a single, enormous memory space for an entire cluster of processors.

This development fits into a broader industry trend toward disaggregated infrastructure. As AI models grow exponentially larger, the traditional approach of building monolithic servers with fixed ratios of compute, memory, and storage is becoming too rigid and expensive. The memory requirements for training and running models like GPT-4 already exceed what a single GPU can hold, forcing complex and costly workarounds. By decoupling memory from the processor, data center operators can scale each resource independently. This allows for more flexible, efficient, and cost-effective hardware configurations tailored specifically to AI workloads. Marvell’s announcement is one of the most concrete proposals for applying this disaggregated model to the demanding world of high-performance AI compute.

For CTOs and infrastructure teams, this technology could fundamentally change data center economics and design in the long term. If successful, it would allow companies to purchase compute and memory as separate, scalable building blocks, reducing waste and lowering the total cost of ownership for large AI clusters. However, the path to adoption is not simple. The AI hardware market is dominated by Nvidia's tightly integrated ecosystem of GPUs, networking, and software. Marvell will need to prove its open-fabric approach can deliver comparable performance and ease of use. Meanwhile, Nvidia is not standing still. The company holds a significant stake in Marvell, suggesting it is hedging its bets, while also developing its own advanced interconnect technologies. The next few years will be critical in determining whether this disaggregated, light-based future can challenge the current, integrated status quo.

This technology represents a shift toward disaggregated memory, where RAM is no longer trapped inside a single server. For engineers and architects, it promises to break the memory capacity bottleneck that currently limits the size and complexity of AI models, enabling more powerful systems without custom hardware.

By allowing expensive memory resources to be shared efficiently across many GPUs, Marvell's fabric could significantly reduce the capital expenditure (CAPEX) for building and scaling AI data centers. This directly challenges the current integrated hardware model and could lower the barrier to entry for large-scale AI training.