Nvidia on Tuesday announced technologies that could make its upcoming Tesla graphics processors more accessible to cloud deployments in enterprises, while also reversing a trend of relegating highly parallel chips to specialized math and scientific calculations.
The company announced new Tesla graphics processors with hardware and software hooks that allow the chips to be self-sufficient in deploying virtual machines and executing programs. Analysts said the new technologies could open the floodgates for cloud deployments through servers with virtualized graphics processors, which will rely less on components such as CPUs for task execution.
Graphics processors are generally considered faster than CPUs, which are more relevant for everyday tasks such as productivity applications. A larger number of servers are now harnessing the parallel-processing capabilities of CPUs and GPUs to scale performance, especially in supercomputing. The world's second-fastest supercomputer, the Tianhe-1A system at the National Supercomputer Center in Tianjin, China, uses Nvidia's Tesla GPUs and Intel's Xeon processors to deliver 2.5 petaflops of performance.
The new graphics processors include the Tesla K10, which has started shipping, and the Tesla K20, which will ship in the fourth quarter. The new chips are based on the Kepler architecture, and are faster and more power-efficient than the older Fermi architecture, which was considered power hungry by design. Chip prices were not available.
Nvidia has added to Kepler a new virtualization technology called VGX, which virtualizes the GPU and makes it a resource that can be shared by multiple CPUs and threads, said Jeff Brown, general manager of the professional solutions group at Nvidia. Nvidia has built a memory management unit into the GPU, ensuring a straightforward deployment of a virtual machine.
GPUs have in the past been used for virtualization. For example, Nvidia and its rival Advanced Micro Devices have offered professional graphics cards for deployment of Windows 7 virtual desktop from servers to client devices. But with VGX, now the GPU can skip CPU cycles and directly deploy and manage virtual machines.
The new virtualization technology has interesting implications in server designs and the deployment of cloud services to thin clients and devices like tablets, said Dean McCarron, principal analyst at Mercury Research.
"We can see some shifting going on. For one, GPUs haven't shown up in the server environment outside the high-performance computing space." McCarron said. "It opens the door for playing very complex, visually detailed games on a thin-client."
The new technology also makes it easier and inexpensive to add GPUs to general server environments, McCarron said. For example, a server-side virtualized GPU will be able to able to render a high-definition game and deliver it over the cloud while take advantage of GPU acceleration features.
"Now you can start doing some interesting things with your workload in terms of a client-server architecture," McCarron said.
The VGX architecture removes a couple of major bottlenecks that kept hybrid -- CPU and GPU --- systems from achieving maximum power and performance efficiency, said Dan Olds, principal analyst at Gabriel Consulting Group.
"After these changes are implemented fully, users will see much higher CPU, GPU, and thus overall system utilization, which makes the already compelling hybrid computing story even stronger," Olds said.
A VGX hypervisor is connected to hypervisors from Citrix or VMware, and graphics boards are being developed based on VGX that can be plugged into PCI-Express 3.0 slots. Server makers including Dell, Hewlett-Packard, IBM and Cisco are adopting the new VGX technology, Nvidia's Brown said.
Multiple workloads require GPUs in servers to execute programs more efficiently. Nvidia has introduced new technology called HybridQ to improve the parallelism and utilization of GPUs, said Sumit Gupta, senior director for the Tesla Business Unit.
The GPU helps execute multiple tasks simultaneously and more efficiently through a hardware scheduler, which ensures the GPU can prioritize and execute tasks more efficiently instead of looping back to the CPU, Gupta said.
"As soon as you start going back to the CPU too much, you lose the benefit of the GPU," Gupta said.
A part of CPU design involves a scheduler that makes sure the workload and its branches get handled correctly. That kind of capability wasn't present in parallel processing units, and Nvidia's implementation of HybridQ will scale GPU performance, Mercury Research's McCarron said.
The GPU already a lot of capabilities of the CPU, and the HybridQ will bring them closer in features, McCarron said. Rather than having specially tuned software, HybridQ will scale the number of workloads virtualized GPUs handle, McCarron said.
The HybridQ technology can be applied to a number of tasks that include computational fluid dynamic, bioinformatics, genome sequencing and circuit simulation, Gupta said. Programmers don't have to make changes in code to benefit from the HybridQ technology and the technology is compatible with Nvidia's CUDA framework, which enables programmers to write highly parallel programs.
HybridQ will come with the Tesla K20, which is targeted at high-end servers and will become available in the fourth quarter.