SCIENCE

T-Platforms Group has unveiled the TB2-TL, a new high-density, heterogeneous computing solution.

The T-Platforms TB2-TL represents a new breed of high-density computing systems, combined with industry-leading performance-per-watt ratio. The first system to use the NVIDIA Tesla X2070 GPU, the TB2-TL is in a class by itself as the densest HPC solution coming to the market. The combination of the T-Platforms T-Blade 2 packaging and the NVIDIA Tesla 20-series GPUs enables 1 petaFLOPS+ level performance in only ten standard racks.

Architected for the high end of the supercomputer market, the TB2-TL system enables customers to reach 105TF of peak DP performance in just one standard 19" rack, while achieving a record-breaking 1450 MFLOPS/W performance-per-watt ratio -- another industry first for T-Platforms. The 7U chassis is packed with 32 Tesla 20-series GPUs and 32 Intel Xeon L5600 series CPU's, up to 3GB of memory per CPU core along with 192GB of GDDR5 RAM, 2 QDR 36-port Infiniband switches, and a management and switching module containing optional FPGA-governed Global Barrier and Global Interrupt networks to support extreme scale systems.

Extraordinary energy efficiency is achieved by using an innovative L-shaped heat sink providing the necessary heat dissipation for the entire TB2-TL blade board equipped with Tesla 20-series GPUs, CPUs and mezzanine memory modules, greatly improving on standard blade tray design used by other system manufacturers.

The single TB2-TL provides almost 4x the performance increase over the same enclosure packed with 64 Intel Xeon E5670 CPUs, while drawing the same amount of power and positioned in a similar price range. Additionally, the T-Platforms Clustrx operating system enables full support for heterogeneous hardware and software environments with near real-time monitoring of petaFLOPS-level installations. The TB2-TL's combination of price/performance, power/performance and petascale-ready operating system makes extreme scale computing available to a wider range of users than ever before.

"The T-Platforms T-Blade 2 is an exceptional accomplishment that delivers high compute density and petaflop class scalability," said Andy Keane, general manager, Tesla business at NVIDIA. "T-Platforms innovative system also sets a new bar in power efficiency -- which has rapidly become the most important factor in supercomputer design."

To ensure the maximum level of performance, each Tesla GPU card is routed through the dedicated Intel 5520 North Bridge chip and has a dedicated single port QDR Infiniband chip, providing necessary bandwidth for intra- and inter-node traffic.

To provide further flexibility, the TB2 chassis design allows customers to mix and match Tesla-based TL blades and pure Intel Xeon blades to match the processor ratios required by different applications.

"We designed this solution for the high end segment of the supercomputer market," said Vsevolod Opanasenko, CEO of T-Platforms. "Working with NVIDIA, we have architected the world's most dense blade system while achieving a performance to power consumption ratio twice that of the top system now listed in the Green500. The TB2-TL will support petascale operations today while giving our customers the infrastructure necessary for the eventual transition to exascale."

"Moscow State University (MSU) is at the forefront of academic research and we are interested in deploying potential GPU-based solutions to identify applications that allow HPC users to increase their productivity and lower computational costs," said Vladimir Voevidin, deputy director of the Research Computing Center at MSU.

The TB2-TL will be available from T-Platforms for select customers in the fourth quarter of 2010 with general availability in the first quarter of 2011. The T-Platforms TB2-TL is an ideal solution for the most demanding applications in segments such as oil and gas, bioinformatics, financial services, CAE, and climate solutions, requiring the versatility and scalability of heterogeneous systems with unparalleled performance.

T-Platforms will be showcasing the new Tesla GPU-based HPC solutions at Booth 29 during the GPU Technology Conference, September 20-23, 2010, at the San Jose Convention Center.