Nvidia pushes vGPU into enterprise servers

Nvidia is pushing virtual GPU software deeper into mainstream servers with a new Blackwell card built to split one accelerator across many workloads. (developer.nvidia.com) The company said on April 22 that its RTX PRO 4500 Blackwell Server Edition pairs with NVIDIA vGPU 20, the latest release of its virtualization stack for data center GPUs. The card has 32 gigabytes of GDDR7 memory, a 165-watt single-slot design, and support for up to two Multi-Instance GPU partitions. (developer.nvidia.com; nvidia.com) A virtual GPU works like carving one physical chip into slices that separate virtual machines can use at the same time. Nvidia said vGPU 20 is aimed at mixed use inside the same server, including Microsoft Office, engineering software, data science, video, and lightweight artificial intelligence development. (developer.nvidia.com) That pitch lands as enterprises try to add artificial intelligence features without turning every server into a dedicated training box. Nvidia’s post frames the change as a move away from single-purpose infrastructure toward shared systems that can run office apps, design tools, and inference jobs side by side. (developer.nvidia.com) The software layer matters as much as the silicon. Nvidia’s vGPU support policy splits releases into one-year Production Branches and three-year Long Term Support Branches, a detail that matters to information technology teams planning upgrades across large virtual machine fleets. (docs.nvidia.com) VMware vSphere is one of the platforms already built around that model. Broadcom’s vSphere 8 documentation says a virtual machine can be configured with up to 16 virtual GPU devices, and Nvidia’s support matrix says vGPU software on vSphere requires vSphere Foundation or vSphere Enterprise Plus. (techdocs.broadcom.com; docs.nvidia.com) VMware has also been widening how those shared GPU profiles can be mixed on one host. In vSphere 8 Update 3, VMware said administrators could place virtual machines with different vGPU profile types and different memory sizes on the same physical GPU, including compute and graphics profiles together. (blogs.vmware.com) That makes tenancy and isolation less of a side issue and more of an operating requirement. The more enterprises normalize shared GPUs for everyday applications, the more their hypervisors need controls for placement, lifecycle management, and keeping one tenant’s workload from interfering with another’s. (developer.nvidia.com; www.vmware.com) Nvidia’s message is that the next enterprise GPU sale is not only about bigger models or faster training. It is also about making one server card useful to many users at once, then letting the infrastructure software keep that sharing orderly. (developer.nvidia.com)