Dell Launches GPU Expansion Chassis For HPC

Dell has boosted its high-performance computing credentials with an expansion chassis capable of holding up to 16 NVidia graphics cards.

The new chassis can hold up to 16 external devices on a PCIe bus, and will connect to up to eight of Dell’s PowerEdge servers. So far, the PowerEdge C410x holds nVidia Tesla M1060 or M2050 graphics processor units (GPUs), but network cards or memory could be allowed in future.

17 Teraflops in a 3U chassis

“This solution stack provides 25 percent higher density than those of our competitors,” explained Paul Brook, solutions, Dell EMEA. To get higher performance graphics on high performance computers (HPC or supercomputers), users have had to either get higher end servers or increase the density, he said. The C410 gives HPC users better graphics without a major upgrade.

At first the chassis will work with the PowerEdge C6100 family, but in future all PowerEdge servers will be able to use it, said Brook.

This product comes from Dell’s Data Centre Solutions (DCS) division, founded in 2007, which develops customised solutions for cloud operators and large data centres, with customers including Facebook, Ask.com and Microsoft, and takes pride in offering a service which is not “on-size-fits-all”.

PowerEdge servers are in use for NASA’s climate research, and other specialist users. The server range uses  both AMD and Intel processors.

The chassis was developed to meet the demands of specific large companies in oil and gas exploration, but has been made available to others, and could be useful for other energy companies, bio-medical science organisations or finance companies which have a need for simulation processes.

With 16 cards installed, the C410x will provide a processing capacity of 16.5 teraflops in a compact 3U chassis. Dell claims that an example configuration, with two PowerEdge C6100 servers connected to the chassis, would give the equivalent of eight dual-socket servers. This would have 2.5 times greater capacity the equivalent of a 100-server system developed in the University of Cambridge in 2006.

The chassis can be managed with existing tools, including a tool for capacity planning. There is no price quoted as yet.

Based on reporting by Pierre Mangin, Silicon.fr.