AMD announced a new line of GPUs today aimed at the increasingly lucrative virtual GPU business. Up until now, that’s been mostly Nvidia’s turf, and Team Green has released several iterations of its Nvidia Grid platform over the past few years. (Note: Nvidia uses the same “Grid” term to refer to both its hardware virtualization business and its cloud gaming service.The gaming version is “GeForce Grid”). The idea behind GPU virtualization is that companies don’t necessarily need to purchase a discrete card for every single user or desktop. Instead, companies can centralize their infrastructure and provide access to virtual systems. By deploying multiple GPUs in a single server, companies can provide the benefits of a discrete GPU (up to and including high-end workstation-class rendering platforms) while simultaneously cutting costs.
AMD is claiming that its new Multiuser GPU family will support up to 15 users on a single GPU. Typically this is done by dividing the GPU’s total processing power across the number of users, though Nvidia has released Grid cards with multiple GPUs on a single PCB to enable optimal scaling per-user. “The AMD graphics cards are uniquely equipped with AMD Multiuser GPU technology embedded into the GPU delivering consistent and predictable performance,” said Sean Burke, AMD corporate vice president and general manager, Professional Graphics. “When these AMD GPUs are appropriately configured to the needs of an organization, end users get the same access to the GPU no matter their workload. Each user is provided with the virtualized performance to design, create and execute their workflows without any one user tying up the entire GPU.”
AMD’s Multiuser GPU uses the SR-IOV (Single Root I/O Virtualization) standard developed by PCI SIG, and the company claims it’s explicitly designed for both OpenCL and graphics performance, with OCL 2.0 support, OpenGL 4.4, and DirectX 12 all supported as well.
With relatively few details, it’s hard to say how AMD’s Multiuser GPU family will compare to Nvidia’s Grid. The single 3D render of a Multiuser GPU shows a single-slot, fairly long PCB without any apparent cooler. It’s not unusual for these kinds of systems to be cooled by a system fan, but the GPU hardware running underneath that heatsink must be fairly svelte to fit in a single-slot form factor. A bullet point in AMD’s PDF implies that these may not be distinctly different cards at all.
Given that AMD has announced no hardware specs, it’s possible that Multiuser is a brand name for software functionality that’s rolling out to existing hardware. AMD would still have to do plenty of software work to get these capabilities rolling in its FirePro cards, but the exact situation is unclear. It’s also unclear why AMD’s PDF characterizes Nvidia Grid the way it does, by listing it as a “Software” solution compared to AMD’s “Hardware” option and by claiming that the maximum number of users supported is eight per GPU.
Nvidia’s own webpages claim that Nvidia Grid is a hardware-based solution with full GPU pass-through enabled to allow local clients to interface properly with remote virtualization stations. It also notes that the maximum number of simultaneous users on a Grid GPU is up to 32 in certain cases. We’ve reached out to AMD to clarify these questions and will update this story when we hear back. What’s likely is that AMD chose some preconditions for its table that aren’t listed within it — allocating a certain amount of RAM per user, or supporting certain resolutions could knock back the number of simultaneous users that a Grid GPU can support, for example.