Nvidia takes its GPU to the enterprise cloud

Nvidia has introduced a handful of technologies based on its new Kepler GPU architecture, designed to extend GPU-accelerated computing beyond supercomputing to enterprise clouds.

Nvidia boss Jen-Hsun Huang has unveiled two new Tesla GPUs, the K10 (pictured) and K20. Image credit: Nvidia

On Tuesday, the graphics chip specialist unveiled two new Tesla GPUs, the K10 and K20, and its Nvidia VGX platform for virtualised desktops. It also showed off a preview of its Cuda 5 parallel programming platform.

Nvidia said it has designed the products for large-scale datacentres, where the GPUs' fast streaming and virtualisation capabilities can "accelerate cloud computing". The Kepler architecture is the follow-up to Fermi, which delivered computational GPUs used in supercomputers for scientific research.

Cuda and Kepler-based products will continue to put GPU-accelerated computing into more hands, according to the company's chief executive Jen-Hsun Huang.

"The democratisation of high-performance computing is happening right in front of us," Huang told an audience at Nvidia's GPU Technology Conference in San Jose, on Tuesday.

Parallel computing

With roughly three times the performance at the same power as the earlier Fermi processors, Kepler introduces two key features that improve its parallel computing performance.

Dynamic parallelism adds support for run-time dependencies in Cuda applications. This means that work can be created on the GPU, rather than passing from CPU to GPU each time new threads are needed. In doing this, it simplifies the algorithms needed for data-dependent applications, making them more efficient — for example, allowing finite-element analysis applications to use adaptive meshes.

Similarly, Hyper-Q, a new queuing technology, gives Kepler 32 concurrent work queues, allowing applications to utilise GPU cores more effectively.

"Kepler is a very big deal for Nvidia," Huang said. "Graphics to the next level and GPU in the cloud."

Kepler processors

Nvidia's new Tesla products are "computing accelerators built to handle the most complex HPC problems in the world", according to the company. Both are based on the same Kepler technology found in the GeForce GTX 690 graphics card, part of the Nvidia's GPU line for consumer devices.

The Kepler processors have 192 cores and, initially, 3.8 billion transistors. In its second generation, the K20 will have over seven billion transistors, making it the most complex processor in the market, according to the manufacturer.

Kepler is a very big deal for Nvidia. Graphics to the next level and GPU in the cloud.

– Jen-Hsun Huang, Nvidia

Tesla K10 will offer single-precision floating point performance of 4.58 teraflops and 320GB per second bandwidth, and it is intended for imaging, signal processing and seismic processing applications. Available in May, it will be followed towards the end of 2012 by the K20, the flagship model in the Tesla family.

The Tesla K20 has been built for intensive high-computing. It is tuned for double-precision operations. It also has support for dynamic parallelism and Hyper-Q.

Some of the new capabilities include real-time ray-tracing support and an improved physics engine. While these are "a big deal for graphics, they're a bigger deal for high-performance computing", Huang said.

Virtual GPU

On stage, Huang also introduced the Nvidia VGX platform, which is targeted at BYOD scenarios in enterprises. The support for GPU virtualisation means that businesses can get "the largest and most powerful GPU, that's also the smallest and most power efficient; one we can all share", Huang said.

The virtual GPU capabilities in Kepler will improve remote workers' experience with virtual desktops, he said, noting that the platform allows several desktops to share the same GPU.

"Employees can now access a true cloud PC from any device — thin client, laptop, tablet or smartphone — regardless of its operating system, and enjoy a responsive experience for the full spectrum of applications previously only available on an office PC," Nvidia said in a statement.

Nvidia is partnering with server manufacturers including HP, Dell and IBM, as well as with VDI providers to add GPU virtualisation support to Citrix XenDesk, VMWare's VDI tools, and Microsoft's Hyper-V and RemoteFX. Nvidia demonstrated various use cases, delivering complete Windows desktops with workstation applications to tablet PCs and lightweight notebooks.

Beyond business computing, Nvidia unveiled a cloud game-streaming platform at the conference. The GeForce Grid uses Kepler virtualised GPUs to deliver games over a network with low latency.