Since 1987 - Covering the Fastest Computers in the World and the People Who Run Them

March 21, 2013

Penguin Pushes Envelope on Compute Density

Richard L. Brandt

In the midst of the GPU Technology Conference this week, Penguin Computing served up a new high-power system, heavy on the GPU/coprocessor side, to meet the needs of HPC customers with heavy processing needs. The Relion 2808GT offers 8x double-width GPGPUs or MICs, and dual Xeon E5-2600 series CPUs with up to eight cores per processor. The company says it is now the server with the highest compute density on the market.

Penguin CEO Charles Wuischpard told HPCwire that the server was designed based on feedback from one particular high-end customer, but that it will fill the needs of many companies that need to do very fast, data-intensive computing. “More and more the opportunity seems to be to have more of a coprocessing component, whether that’s an NVIDIA GPU or an Intel Phi or even AMD APU,” he says.

The ratio of CPU to GPU that his customers want varies. “Some customers want one-to-one, some want two-to-one, and then there’s that fringe out there – we see it in oil and gas mostly – that wants a lot of GPUs, very densely packed into a couple of processors,” he says. “We’ve always struggled with trying to find a design that can support that level of density and do it in a performant way. Were really pleased with this.”

The new Relion fits the needs of that GPU-hungry fringe. It can hold eight GPGPUs or other coprocessors in two rack units. Its dual-socket platform is based on Intel’s Xeon E5-2600 CPU family. If it’s loaded up with eight NVIDIA K20 GPUs, it can get 28 teraflops of single precision floating point performance. It has 16 DIMM sockets for up to 512 GB of 1600 MHz DDR3 RAM ECC memory. It also has an on-board dual 10GbE BASE-T controller and optional support for two 10GbE SFP+ ports.

Penguin CTO Phillip Pokorny says that some organizations already have software programs that run well completely inside a GPU and scale easily as the number of GPUs increase. Those customers want as many of the graphics chips as they can get in a small space. “The key things for us were finding a form factor that cooled effectively and had appropriate power. Those two challenges, cooling and power, are the ones that we run into most often,” he noted. The Relion 2808GT also features a dual 1600W high-efficiency power supply.

Pokorny adds, however, that the server can be configured in many different ways, depending on the customer needs. An advantage of the Xeon E5-2600 CPU is that it offers PCIe Gen 3, integrated on the processor die. With that, the server can deliver full bandwidth to every GPU socket. For compute jobs that require a lot of communication, Penguin can put additional PCIe switch chips on the risers. For applications that are dominated by computation on the GPU, the switch chips can be eliminated in order to add more nodes. “We’re like the old Burger King saw,” he quips. In other words: Have it your way.

Large GPU-centric applications, of course, require a lot of memory and storage to hold both the raw data and the results of the computations. Pokorny notes that it’s now very cost-efficient to add a lot of memory to support both CPUs and GPUs, and Penguin is able to easily double the RAM. It is offering many different memory configurations for the Relion 2808GT and a wide variety of hard drive options (including spinning discs with eight spindles for up to 1.6 terabytes).

Although the spec sheet for the new Relion 2808GT remains politically correct by not specifying what GPUs or coprocessors might best suit the device, it’s not a coincidence that the server was displayed at NVIDIA’s GTC13 conference. The need for speed, memory and energy efficiency makes this server a very good candidate for NVIDIA’s latest and future generations of GPGPUs.

At that conference, NVIDIA CEO Jen-Hsun Huang emphasized the speed of his latest processors, their dense memory, and fast I/O between the GPU and the DRAM. He even included in his keynote talk a surprise acknowledgment that two generations from now, the “Volta” GPU will offer stacked DRAM. Judging by hypothetical images Huang produced at the conference, several stacked memory chips can be placed very near the GPU on the same substrate, increasing both memory density and I/O speed to the processor. Huang said that Volta will be able to move data between them at 1TB/s.

As for plans to implement such future NVIDIA designs, Penguin’s response was vague. “R2808GT has been designed to accommodate the highest density of NVIDIA Tesla K10 and K20 generation GPUs, including future cards designed for similar physical [characteristics] and power envelope,” noted a company rep.

Penguin is seeing demand for the GPU-intense version of the server from a lot of different types of companies. Oil and gas businesses can use if for seismic studies, for example. Bioinformatics companies have a need to analyze huge volumes of images generated by DNA scanners that take photographic images of the DNA. Semiconductor companies use it to generate silicon mask designs. Pokorny says there is also strong demand from “government organizations,” but they don’t tell him what they need it for.

Wuischpard says the machine is a great example of Penguin’s ability to very quickly create and release customizable servers using the latest technology. He says he mentioned to one Intel executive recently that Penguin is always concerned about getting pounded by big players such as HP or Dell, but the Intel exec’s response was that the big companies are all so distracted by the tablet market that the datacenter business is being nibbled away from them by the more nimble, local OEMs like Penguin. This server, says Wuischpard, “is one aspect of that level of nimbleness.”

Share This