On Tuesday at the GPU Technology Conference (GTC), NVIDIA CEO Jen-Hsun Huang surprised the audience by revealing a new set of technologies that would launch the GPU maker into the cloud computing business. Starting later this year, NVIDIA, along with its partners, will be offering a server platform that can virtualize on-board GPUs for virtual desktop infrastructure (VDI).
The underlying hardware will be based on the newly hatched Kepler GPU technology, which apparently has baked in intelligence to share processing across multiple remote users. NVIDIA has constructed a four-GPU server board, known as VGX, which is purpose-built for GPU virtualization.,
The idea is to deploy them ino servers, where they can be used to power graphics-heavy desktop apps like 3D design tools, simulations, and other visually-oriented programs. The target audience is enterprise users — designers, knowledge workers, and other PC power users — who (like consumers) are migrating to thinner clients. And since there are hundreds of millions of these users out there, the market seems primed for such an offering.
The idea is also being driven by the number and heterogeneity of computing device people are using — everything from desktops and laptops to tablets and smartphones. Since a virtualized GPU in a datacenter is operates at a level above all that architectural noise, it’s in a position to smooth out that heterogeneity. And given that you can put a lot more GPU into a server than most personal devices, these thin client machines will soon be able to tap into a lot more graphics power.
The VGX board, along with the supporting software (a GPU hypervisor and some configuration tools) can support up to 100 sessions at a time, depending upon the particular usage. And, according to NVIDIA at least, it can do so rather seamless, with the latency you would expect on a client device.
But this is not designed for supercomputing in the cloud. The GPUs on the VGX board, although Kepler-grade, are fairly modest in performance. According to NVIDIA, they support only purely single precision floating point, and are even less powerful than the non-HPC Kepler GPUs (GK104) that went into the new GeForce GTX 680 discrete part. Also the virtualization technology can’t aggregate the GPUs in a single super-GPU, either in the board or across servers.
NVIDIA is busy gathering OEMs and has signed up all the major server makers including IBM, HP, Dell, Supermicro and Cisco, as well cloud hoster Amazon. On the client hypervisor side is Citrix, Microsoft, VMware, and Xen. If all goes as planned, these virtualized GPUs should start popping up in datacenters by the end of the year.