CXL vs. RDMA over Ethernet for Disaggregated Memory
If memory disaggregation is the destination, there are two roads being built toward it. One is Compute Express Link (CXL), a new memory interconnect. The other is Remote Direct Memory Access (RDMA) running over standard Ethernet. Both aim to let servers reach memory that does not live inside them — but they get there very differently, and on very different timelines.
The practical question for infrastructure teams is not which one wins. It is which one is deployable now.
The Shared Goal
Both approaches exist to solve the same problem: memory is trapped inside servers, stranded where it cannot be shared and over-provisioned where it cannot be avoided. Each wants to turn memory into a pool that compute can draw from on demand. Where they differ is in how the compute node talks to that pool.
CXL: A New Memory Interconnect
CXL is a cache-coherent interconnect that rides on the PCI Express physical layer. It lets a CPU address external memory as if it were part of the system's own memory map, with hardware maintaining coherence. That is a powerful model: at its best, remote memory looks and behaves like local memory to software.
The trade-off is maturity and reach. CXL depends on the CPU, platform, and device generation all supporting the right revision, and broad memory-pooling and switching capabilities are still ramping across the ecosystem. It is a compelling long-term direction — but adopting it often means aligning your hardware roadmap to it and waiting for the pieces to arrive.
RDMA over Ethernet: Using the Fabric You Already Have
RDMA takes a different route. It lets one machine read and write another machine's memory directly over the network, bypassing the remote CPU and the operating system's copy path. Run over a lossless Ethernet fabric using RoCEv2, it delivers high throughput and low latency on infrastructure that already exists in data centers today.
RDMA does not require a new interconnect or a new CPU generation. It uses Ethernet — the fabric already running through the building.
That is the decisive advantage for teams with a problem now: RDMA over Ethernet is a proven, widely deployed technology, not a future capability gated on a hardware refresh cycle.
The Trade-offs at a Glance
- Programming model: CXL exposes remote memory through the CPU's coherent memory map; RDMA exposes it through a networking model applications and runtimes drive explicitly.
- Hardware dependency: CXL requires matching CPU, platform, and device support; RDMA over Ethernet runs on RDMA-capable NICs over a standard, lossless-configured Ethernet fabric.
- Availability: CXL memory pooling is still maturing across the ecosystem; RDMA over Ethernet is deployable today.
- Reach: CXL is oriented around rack-scale coherence; Ethernet fabrics already span the data center.
Why TORmem Builds on Ethernet and RDMA Today
TORmem's memory-centric architecture is built to be deployed now, not later. Its memory appliances present a shared pool that GPU and CPU servers reach over standard Ethernet using RDMA and RoCEv2 — proven technology that lets enterprises expand memory capacity and raise utilization without waiting for a future interconnect to arrive across the fleet.
This is not a bet against CXL. It is a recognition that the memory bottleneck is here today, and the fabric to relieve it is already in the rack.
The Bottom Line
CXL and RDMA over Ethernet are not mutually exclusive, and the industry will likely use both. But they answer different questions. CXL asks what memory should look like a few hardware generations from now. RDMA over Ethernet asks how to share memory on the infrastructure you already own, this year.
For teams whose GPUs are gated by memory today, the fabric that is already deployed is the one that solves the problem now.