NVIDIA Virtual Compute Server

GPU sharing (fractional) is only possible with NVIDIA vGPU technology. It allows multiple VMs to share a GPU and maximize utilization for lighter workloads that require GPU acceleration.

With GPU aggregation, a VM can access multiple GPUs, which is often required for compute-intensive workloads. vCS supports both multi-vGPU and peer-to-peer computing. In multi-vGPU, the GPUs are not directly connected. In peer-to-peer, they are connected through NVLink to achieve higher bandwidth.

vCS supports monitoring at the app, guest, and host levels. In addition, proactive management capabilities provide the ability to perform live migration, pause and resume and create thresholds. These thresholds show consumption trends that impact the user experience. All of this is possible in the vGPU Management SDK.

NVIDIA GPU Cloud (NGC) is a hub for GPU-optimized software that simplifies workflows for deep learning, machine learning, and HPC, and now supports virtualized environments with NVIDIA vCS.

NVIDIA® NVLink™ is a fast, direct GPU-to-GPU connection that provides higher bandwidth, more connections, and improved scalability for system configurations with multiple GPUs - now virtually supported with NVIDIA virtual GPU (vGPU) technology.

Error Correction Code (ECC) and Page Retirement provide increased reliability for computing applications that are susceptible to data corruption. They are especially important in large cluster computing environments where GPUs process very large data sets and/or run applications for long periods of time.

Multi-instance GPUs (MIG) represent a revolutionary technology that can extend the capabilities of the data center, allowing each NVIDIA A100 Tensor Core GPU to be split into up to seven fully isolated instances and backed up at the hardware level with dedicated memory, cache, and high-bandwidth compute cores. With vCS software, a VM can run on each of these MIG instances, allowing organizations to take advantage of the management, monitoring and operational benefits of hypervisor-based server virtualization.

GPUDirect® Remote Direct Memory Access (RDMA) allows network devices to access GPU memory directly, bypassing CPU host memory, reducing GPU communication latency, and fully offloading the CPU.