The system features two independent head nodes, 256 compute nodes and 60 InfiniBand and Ethernet switches, all tightly coupled into a single, powerful computing resource.
The initial A-Class peak performance of 420 teraflops per single system can be scaled up to 54 petaflops by combining a maximum of 128 systems. The first released compute module features four custom system boards with a single Intel Xeon E5-2600 v2 microprocessor and NVIDIA Tesla K40 accelerator. A set-up with one processor and one GPU accelerator provides a balanced configuration to solve various computational problems, while optimal implementation of NVIDIA GPU Direct technology helps to achieve the maximum optimal computational efficiency.
The A-Class offers a comprehensive network environment, providing two independent Ethernet networks for system management and two independent InfiniBand networks, one used for MPI traffic and the other for data access. Popular 3D- and 4D-torus, flattened butterfly, and hypercube, among other advanced topologies, are available to reach optimal interconnect performance and efficiency.
The A-Class management, compute and switching modules are directly cooled by hot water, with input temperature as high as 45°C and output of over 50°C. Printed circuit boards of nodes and switches are attached to a customer radiator plate with spill-proof water connectors.
The use of hot water as a coolant significantly reduces system operational noise and enables the system's year-round 'free cooling' mode for cost savings on compressors and chillers. To improve TCO during winter, customers can reuse the hot outlet water to heat offices.
A host of resilient hardware and advanced software features enhance A-Class reliability. A management failover function is implemented via two independent hot-swappable management modules with separate networks. A-Class management modules monitor system components, control programme load states and configuration of installed software. Independent power inputs and redundant power supplies ensure the entire system's continued operation, critical for petascale computations.
The Moscow State University is running A-Class field tests and recently submitted system performance details for inclusion in the TOP500 list of the world's most powerful supercomputers. A single A-Class system was ranked 130th in June 2014 edition of the list.
"The A-Class is an impressive HPC system", stated Professor Vladimir Voevodin, Deputy Director at the Research Computing Center, M.V. Lomonosov Moscow State University (MSU). "The newest solution from T-Platforms gives us more choices as we prepare for the new multi-petaflops supercomputer deployment."
"At the earliest A-Class design stages we listened carefully to the input from MSU, the largest customer and primary research partner of T-Platforms", stated Mr. Seva Opanasenko, T-Platforms' CEO. "And to address current and future HPC challenges we worked very closely with suppliers to implement the right mix of performance and system efficiency."
The A-Class system is going through testing and certification stages and is now available for pre-orders.