August 9, 2012

NVIDIA’s Supercomputing Kepler GPU Lands In Workstations

Michael Feldman

The Tesla K20, NVIDIA’s Kepler GPU that will power petascale supercomputers like Titan and Blue Waters in a few short months, will also be showing up in workstations. This week, the GPU maker announced that its second-generation Maximus platform, will include the new K20 along with the Quadro K5000, the Kepler-based GPU for high-end graphics. HP, Dell, Lenovo, Fujitsu, Supermicro, and BOXX Technologies are expected to offer workstations based on the new design.

Maximus, which NVIDIA launched last November, is a workstation platform that combines Quadro and Tesla parts so that users can do visualization and computation simultaneously without having to share that functionality on a single GPU resource. The technology automagically parcels out the work to the appropriate device, allowing the application to perform the visuals and number-crunching in parallel. Typical applications include computer-aided design (CAD), computer-aided engineering (CAE), computational fluid dynamics (CFD), seismic analysis, and image rendering.

Having the K20 in a deskside machine will certainly give professionals a lot more personal flops. The first-generation Maximus workstations were based on the Fermi C2075 GPU cards, which delivered about 500 double precision gigaflops. Although NVIDIA is not yet revealing the performance numbers on the upcoming K20, which won’t be generally available until December, it’s likely to be well above a double precision teraflop. On-board memory capacity is also TBA, but since the Fermi C2075 provides 6GB, it’s a good bet that the new K20 card will have more than that.

NVIDIA did reveal the Tesla K20 will have a retail price of $3,199. However, the ones going into HPC cluster and supercomputers might end up being more expensive. Those server parts could have faster clocks, more CUDA cores, and more memory than the K20s destined for workstations. In fact, there might be a few variants of this product.

Nonetheless, having a teraflop in a box is going to be very enticing for a lot of tech professionals. Some vendors are even likely to offer multiple-Tesla configurations. BOXX, for example, currently sells a first-generation Maximus workstation with three Tesla C2075 GPUs, so if it followed suit with the K20, a 5-teraflop personal computer is a real possibility.

At this point, one might wonder what’s the point of waiting in line to get cycles on an HPC cluster if you can get the same computational horsepower beside your desk. According to a report in Cadalyst, some of the NVIDIA folks seem to be thinking along those same lines. David Watters, the GPU maker’s senior director of the manufacturing and design industries, says Maximus is essentially designed to bring HPC back onto the desktop, reversing the trend that has relegated many of these applications to the server room.

Ignoring for a moment that the byte-to-flops ratio is apt to be rather lean on a 5-teraflop workstation, offering the K20 on a workstation could theoretically cannibalize some server-based revenue for NVIDIA. The company probably doesn’t see it that way though. The conventional wisdom is that users will get technical workstations for personal use in addition to the cluster, and use the latter when they need to scale up their application or for codes that don’t fit in a deskside machine. And this is most likely to occur because of memory capacity; it’s hard (and expensive) to stuff more than a hundred gigabytes or so into a workstation.

The other reason NVIDIA is sharing its top-of-the-line Kepler with the non-server crowd is that the company doesn’t want to cede the market to AMD, which this week announced their own teraflop GPUs for workstations. And since there are an increasing number of applications being ported to GPUs, teraflop workstations will look all that more attractive. Now if they could just figure out a way to get a terabyte of memory into the same box…