Question 1

Does DynamicXcelerator require specialized hardware?

Accepted Answer

The DynamicXcelerator™ is designed to run on industry-standard servers and components—including GPUs, NVMe storage, FPGAs, NICs, within the complete system architecture provided by Corespan Systems.  It uses industry-standard technology.

Question 2

How does DynamicXcelerator improve GPU utilization?

Accepted Answer

DynamicXcelerator™ improves GPU utilization by disaggregating GPUs and other PCIe devices from traditional server boundaries. Resources can be dynamically reassigned in real time, ensuring each workload receives exactly the GPUs it needs -- no more, no less. This composability also enables hosts to be configured with virtually any number of GPUs, independent of their physical proximity, maximizing overall system efficiency.

Question 3

What is DynamicXcelerator?

Accepted Answer

DynamicXcelerator™ is an infrastructure composition system that disaggregates GPU infrastructure, it is essentially the name given for the overall system architecture from Corespan Systems, which comprises PRU, iFIC and Corespan Composer

Question 4

Can Corespan Composer be used in on-prem or sovereign environments?

Accepted Answer

Corespan Composer directly programs the photonic switches used in the DynamicXcelerator™ architecture to establish predictable, workload-specific connectivity between disaggregated resources.

Question 5

Who is DynamicXcelerator™ designed for?

Accepted Answer

It is designed for enterprises, cloud providers, and Neo-clouds who require a secure, flexible, and efficient GPU deployment architecture.

Question 6

Can the 2500 Series integrate with existing infrastructure?

Accepted Answer

Yes. The 2500 Series is designed to integrate seamlessly with standard servers, existing GPUs, and legacy environments while providing a clear path for future expansion and modernization.

Question 7

How is the 2500 Series managed?

Accepted Answer

The platform is orchestrated through Corespan Composer, a centralized fabric and resource management software suite that enables real-time provisioning, monitoring, and optimization across the entire disaggregated system.

Question 8

What is the FIC 2500?

Accepted Answer

The FIC 2500 is a PCIe Gen5 fabric interface card that connects hosts and device chassis to the photonic interconnect.

Question 9

What role does the FIC play in composable infrastructure?

Accepted Answer

The FIC abstracts physical connectivity and enables software-controlled attachment of remote PCIe resources.

Question 10

Where is the FIC deployed?

Accepted Answer

FICs are installed in host machines (i.e. servers) and PRU chassis to establish end-to-end PCIe connectivity across the fabric.

Question 11

Does the FIC affect performance or latency?

Accepted Answer

The FIC is designed to preserve native PCIe performance with predictable, deterministic latency.

Question 12

How is the FIC managed?

Accepted Answer

FICs are managed through Corespan Composer as part of system orchestration and lifecycle control.

Question 13

What is the PRU 2500?

Accepted Answer

The PRU 2500 is a high-density PCIe Gen5 chassis that pools GPUs and other PCIe devices for dynamic allocation across multiple systems.

Question 14

How many devices does the PRU 2500 support?

Accepted Answer

The chassis supports up to 12 PCIe devices, including up to 10 double-width GPUs, depending on configuration.

Question 15

Is the PRU 2500 limited to specific vendors or device types?

Accepted Answer

No. The PRU 2500 supports multi-vendor, multi-generation GPUs and standard PCIe devices.

Question 16

Can devices be attached or detached while systems are running?

Accepted Answer

Yes. The PRU 2500 supports hot-swap and real-time attach/detach, enabling workload-driven reconfiguration.

Question 17

What workloads is the PRU 2500 designed for?

Accepted Answer

It is well suited for AI inference, HPC, storage acceleration, and financial computing workloads that require high bandwidth and flexible resource allocation.

Question 18

How does the PRU 2500 improve infrastructure efficiency?

Accepted Answer

By pooling devices and allocating them only when required, the PRU 2500 reduces idle hardware and avoids overprovisioning.

Question 19

What is the Corespan Photonic switch (PXC)?

Accepted Answer

The DynamicXcelerator™ architecture integrates optics, single-mode fiber, and an all-optical active photonic switch (PXC) to deliver scalable, low-loss interconnects for PCIe serdes.

Question 20

Why use a photonic fabric instead of electrical switching?

Accepted Answer

Photonics offers lower power consumption, lower insertion loss, and better scalability than electrical switching, particularly at large port counts as well as reach over copper based interconnects.

Question 21

Are the optical interconnects blocking or oversubscribed?

Accepted Answer

No. The photonic interconnects are deterministic, non-blocking connectivity based on explicit system composition and completely private.

Question 22

Does the fabric require proprietary cabling or optics?

Accepted Answer

No. It uses standard single-mode fiber and optics, with support for both standard and custom cable lengths.

Question 23

What is the 2500 Product Series?

Accepted Answer

The 2500 Series is a PCIe Gen5–based disaggregated infrastructure platform that enables dynamic composition of GPU, storage, and accelerator resources using high-speed photonic interconnects. It provides a scalable foundation for next-generation AI, HPC, and cloud infrastructure.

Question 24

What problems does the 2500 Series solve?

Accepted Answer

The 2500 Series eliminates the constraints of fixed server architectures by enabling modular, component-level resource pooling. This flexibility improves utilization, simplifies upgrades, and delivers stronger long-term ROI through adaptive infrastructure design.

Question 25

What types of workloads are the 2500 Series designed for?

Accepted Answer

The 2500 Series is optimized for AI inference, enterprise AI and GPU-as-a-Service deployments.

Question 26

How does the 2500 Series improve performance?

Accepted Answer

It enables direct GPU-to-GPU and GPU-to-storage communication over PCIe Gen5, reducing the latency and bottlenecks associated with traditional server architectures, and unlocking higher throughput for data-intensive workloads.

Question 27

What is Corespan Composer?

Accepted Answer

Corespan Composer is the unified software control plane for our DynamicXcelerator™ architecture. It orchestrates system composition, resource attachment, photonic interconnect control, and bare-metal lifecycle automation through a single platform.

Question 28

What problems does Corespan Composer solve?

Accepted Answer

Corepsan Composer is the software interface into the DynamicXcelerator™ system that replaces static server configurations with dynamic composition of CPUs, GPUs, storage, and accelerators. This increases utilization, reduces stranded capacity, and shortens time-to-deploy.

Question 29

Is Corespan Composer a virtualization platform?

Accepted Answer

No. Composer operates at the bare-metal level. Resources are composed into physical systems without hypervisors or virtual machines, preserving native performance and deterministic behavior.

Question 30

How does Corespan Composer interact with the photonic switches?

Accepted Answer

Corespan Composer directly programs the photonic switches used in the DynamicXcelerator™ architecture to establish predictable, workload-specific connectivity between disaggregated resources.

Question 31

Does Corespan Composer integrate with existing workflows?

Accepted Answer

Corespan Composer supports automated hardware discovery, OS provisioning, and lifecycle management, enabling integration with existing operational and DevOps processes.

Question 32

What is Corespan 2500 with 5090 Solution?

Accepted Answer

It is a PRU 2500-based solution that combines 8 to 12 RTX 5090 GPUs, hybrid liquid cooling, FIC 2500 connectivity, and Corespan software into a composable platform for AI inference and related accelerated workloads.

Question 33

Is it a traditional server?

Accepted Answer

No. The PRU 2500 is a resource chassis, not a standalone server with its own CPU and operating system. It is designed to make accelerator resources available to external hosts.

Question 34

How many GPUs does it support?

Accepted Answer

The solution is designed for 8 to 12 RTX 5090 GPUs in a single PRU 2500 configuration.

Question 35

Does it support Docker and Kubernetes?

Accepted Answer

Yes. The platform is designed to work with standard Docker and Kubernetes tooling, including the NVIDIA Container Toolkit and the NVIDIA Kubernetes device plugin.

Question 36

How are the GPUs connected to hosts?

Accepted Answer

Hosts connect through one or more FIC 2500 cards, which present remote GPUs to the operating system as PCIe devices.

Question 37

What makes the cooling approach different?

Accepted Answer

The system uses direct-to-chip liquid cooling on the GPUs and airflow for ancillary components, enabling dense GPU deployment in practical data center environments.

Question 38

How does a 4x RTX 5090 PRU 2500 compare to an H100 SXM on real inference workloads?

Accepted Answer

On Qwen2.5-32B-Instruct at FP8 with vLLM, pipeline parallelism, and chunked prefill, a 4x RTX 5090 PRU 2500 delivers ~5,345 tokens/sec aggregate, which beats a fully optimized single H100 SXM running the same model class on TRT-LLM (~4,285 tok/s). The H100 still wins per-GPU, but the 5090's case is aggregate throughput per dollar.

Question 39

What is the cost-per-million-tokens advantage versus H100 cloud pricing?

Accepted Answer

At sustained utilization, a 4x RTX 5090 PRU 2500 delivers Qwen-32B-class FP8 inference at roughly $0.04 per million tokens, versus $0.30 to $0.60 per million tokens on H100 cloud pricing. That is 7.5x to 15x cheaper for the same model class.

Question 40

Can the PRU 2500 mix GPU types in a single chassis?

Accepted Answer

Yes. The split-side architecture supports 5x liquid-cooled RTX 5090s on one side with the other side configurable for additional 5090s, RTX 6000 PRO cards, AMD Instinct MI350 accelerators, or up to 600 TB of NVMe storage.

Question 41

What inference frameworks are supported?

Accepted Answer

The benchmark above runs on vLLM, the most widely deployed open-source serving stack. The platform also supports TensorRT-LLM, SGLang, and any other Docker- or Kubernetes-native inference runtime.

Question 42

Can the Mercury Rack mix GPUs from different vendors in one rack?

Accepted Answer

Yes. The 72 slots take standard double-wide PCIe Gen5 accelerators, so NVIDIA, AMD, and Intel silicon can share the same rack and fabric while serving different workloads, with CUDA and ROCm available side by side.

Question 43

Does composability hurt GPU-to-GPU performance?

Accepted Answer

The PCIe Gen5-over-optical fabric provides full Gen5 bandwidth between any GPU pair in the rack, with no proprietary interconnect domain limiting which GPUs can communicate. Pooled devices are addressable like local ones.

Question 44

Does Mercury require direct liquid cooling?

Accepted Answer

No. Mercury delivers up to 111 kW in a standard air-cooled 48U rack. In-rack direct liquid cooling is available as an option, not a prerequisite, so it fits data centers that are air-cooled today and staying that way.

Question 45

How does Composer work with Kubernetes and Slurm?

Accepted Answer

Composer sits beneath your existing schedulers rather than replacing them. It assigns GPUs, hosts, and NVMe to workloads at runtime without cabling changes or reboots, so Kubernetes and Slurm keep operating as they do now.

Frequently Asked Questions