Remote GPU Across the STRATUM Continuum

A physical GPU can live on one STRATUM worker while a VM executing somewhere else in the Continuum consumes that accelerator through its local STRATUM GPU PCI device. A CPU-only execution environment can therefore gain GPU capability without physically containing a GPU.

The provider can be in the same rack, another datacenter, an edge site, an Amazon EC2 GPU instance, another cloud, or another location reachable through STRATUM Continuum. The workload can remain where its CPU, storage, network, data-sovereignty, or mission requirements place it.

GPU provider

The provider is a STRATUM worker with access to a physical accelerator. In AWS, this can be a GPU-enabled EC2 instance such as a G5 instance with an NVIDIA A10G or a P3 instance with an NVIDIA V100, where that family is available in the selected region. The worker enrolls into STRATUM normally and exposes its accelerator resources through AI Accelerator Fabric.

GPU consumer

The consumer is the VM that needs GPU capability. It may execute on a CPU-only EC2 worker, an on-premises worker, a laboratory system, another cloud, or an edge platform elsewhere in the Continuum.

Attach STRATUM GPU to the VM and install STRATUM Guest Tools. Inside the guest, STRATUM presents the synthetic GPU PCI device and the CUDA/NVML path used by ordinary GPU applications. The application does not need to know which physical server contains the accelerator.

Accelerator driver activity is transported between the VM and the physical GPU provider across the STRATUM Continuum. The transport supports HTTP/2, Zstd compression, and the Continuum's encrypted mesh.

From the VM's point of view, the accelerator is local. Physically, it may be remote. For supported providers, the same architectural model can expose CUDA or ROCm capability through STRATUM GPU.

1. GPU worker

Launch a GPU-enabled EC2 instance, install STRATUM as a worker, enroll it into the controller, make the Continuum path reachable through the required AWS security-group policy, and confirm that the physical accelerator appears in AI Accelerator Fabric.

2. CPU worker

Launch a compatible CPU-only EC2 instance, or use an existing STRATUM worker elsewhere in the Continuum. Enroll the worker normally and place the VM there.

3. VM

Attach STRATUM GPU, install STRATUM Guest Tools, verify the synthetic PCI device and /dev/stratum-gpuctl, validate CUDA/NVML or the applicable accelerator stack, and run the application normally.

The CPU worker and GPU worker do not need to be the same machine. They do not need to be the same physical site.

Remote GPU execution sends accelerator operations across a network instead of a local PCIe bus. Performance therefore depends on bandwidth, latency, request patterns, and physical distance.

For latency-sensitive workloads, place the VM and GPU provider close together, such as within the same AWS VPC or Availability Zone where appropriate. STRATUM still allows a much wider placement model when the workload can tolerate the network characteristics.

This creates a placement option that local PCI attachment cannot provide: accelerator scarcity and workload location can be managed independently.

The STRATUM GPU shim intercepts accelerator driver API activity, so the supported guest runtime and physical provider stack must remain compatible. For CUDA environments, validate the supported CUDA path and normal STRATUM GPU checks before placing production AI workloads on the remote accelerator.

Do not install a conventional NVIDIA kernel driver inside a STRATUM GPU VM. The synthetic STRATUM PCI device is owned by the stratum_gpu driver and STRATUM Guest Tools.

Cloud security groups and network policy must allow the STRATUM Continuum path between participating nodes. When traffic crosses untrusted infrastructure, the Continuum provides the encrypted mesh used by the GPU transport.