GPU Disaggregation Over Photonics
Pool GPUs, storage, and accelerators into shared resources connected by a photonic fabric — and compose them around workloads in real time, without overspending or overbuilding.
What is GPU disaggregation?
GPU disaggregation separates GPUs, storage, and accelerators from individual servers and pools them as shared resources. Instead of sizing every server for its worst-case workload, infrastructure teams compose exactly the resources a workload needs — and reclaim them the moment it finishes.
Why PCIe over optics?
Copper interconnects run out of reach and power headroom as clusters grow. Photonics offers lower power consumption, lower insertion loss, and better scalability than electrical switching — particularly at large port counts — with reach copper can't match. The Corespan 2500 Series connects standard PCIe Gen 5 servers directly to a photonic fabric at full x16 line rate, so pooled GPUs behave as if they were local.
How Corespan does it
The 2500 Series is a PCIe Gen 5–based disaggregated infrastructure platform: the PRU 2500 pools GPUs into a shared resource unit, the FIC 2500 attaches servers to the photonic fabric, and Corespan Composer — the unified control plane of the DynamicXcelerator architecture — orchestrates composition, resource attachment, and photonic interconnect control from a single platform. GPUs are assigned through familiar Docker and Kubernetes workflows.
The economics follow: a 4x RTX 5090 PRU 2500 out-throughputs an H100 SXM on Qwen-32B FP8 inference at one-tenth the cost per million tokens.
Why disaggregate GPUs?
Static servers strand capacity. Composable infrastructure puts it back to work.
Maximize utilization
Turn stranded GPUs into a shared pool that any workload can claim — and release — on demand.
Scale beyond the server
Full x16 PCIe Gen 5 over photonic interconnects extends the resource pool across racks, not just inside one chassis.
Keep your existing servers
The 2500 Series integrates with standard servers, existing GPUs, and legacy environments — no forklift required.
One control plane
Corespan Composer handles system composition, resource attachment, photonic switching, and bare-metal lifecycle automation.
See composable GPU infrastructure in action
Talk to our engineers about pooling your GPU estate over a photonic fabric.