Resources

In the Lab: 8x AMD Radeon AI PRO R9700, One Fabric

We ran eight AMD Radeon AI PRO R9700s as a single fabric in a PRU 2500. Every GPU-to-GPU pair holds ~111 GB/s bidirectional — roughly 2x what NVIDIA's own reference architecture targets for RTX PRO 6000 fleets.

Every GPU Loan is Also a Topology Loan

Software portability lets a workload move across silicon. Hardware composability lets the silicon move across workloads. A lender's recovery scenario needs both.

The Commodity Optical Switch Thesis: Why the Moat in Photonic Infrastructure Is the Control Plane

The optical switch will be essential, but essential does not mean differentiated. The durable value in photonic infrastructure sits in the control plane — the software that composes heterogeneous fleets into one adaptable system.

Scale-Across: The Architecture the Inference Era Actually Needs

The inference era is being scored on the wrong metric. Scale-across composes GPUs, memory, and I/O per workload — then reclaims it — so operators stop paying peak-case prices for average-case demand.

The Corespan Mercury Rack — a vendor-neutral composable rack platform with 72 double-wide PCIe Gen5 GPU slots, up to 1.2PB of composable NVMe per PRU 2500, and up to 111 kW of rack power.

The Mercury Rack: One Rack, Every GPU, Yours to Compose

72 double-wide PCIe Gen5 slots, any vendor's accelerators, 1.2 PB of composable NVMe per PRU, and 111 kW in a standard air-cooled 48U rack. The Mercury Rack makes the compute-to-storage ratio a scheduling decision instead of a purchasing one.