HP has introduced a three-GPU blade server that plugs into a modular chassis to provide up to one teraflop of computing per rack unit.
The ProLiant SL390s server is a "completely new architecture" with a form factor explicitly designed for the new SL6500 Scalable System chassis, Mark Potter, HP's general manager for industry standard servers, told ZDNet UK on Tuesday. It is HP's first server to have been wholly designed around GPU (graphics processing unit) functionality, according to the company.
The 2U SL390 server is designed for the SL6500 chassis. Photo credit: HP
GPUs are adept at performing floating-point calculations, which makes them capable of improving graphics performance. They are also useful for cheap high-performance computing (HPC) and for emerging business needs such as cloud computing and web serving.
The server comes in two configurations — a 1U half-width server without GPUs, and a 2U half-width server that can be fitted with up to three GPUs. The latter configuration is "tailored for HPC workloads that need incredible floating-point performance", such as weather simulation and modelling work, Potter said.
According to HP, one SL390s with three GPUs can in practice handle a teraflop's worth of calculations. Its theoretical maximum is around 1.6 teraflops. At a launch event in Barcelona, the company demonstrated one 2U SL390s delivering a teraflop while running the Linpack benchmark, which is used to compile the Top500 supercomputing list.
The SL390s servers have been designed explicitly for the SL6500 chassis, which places the server trays, disks and networking equipment at the front so they are accessible from the cold aisle. The SL6500 is an update of HP's SL6000 line, launched in June 2009, and both the chassis and the server are in HP's Converged Infrastructure portfolio.
Each server can hold up to two six-core Intel Xeon 5600 processors. As for communications, they support two 1Gb Ethernet ports, one 10Gb Ethernet port and the option of an on-board InfiniBand port. All of these are on the server nodes themselves, which helps maximise the transmission of the data generated when performing general-purpose computing on GPUs, according to HP.
There is an element of lost performance as the servers are scaled up, due to the limits on internal server components such as PCI Express and on the overall cluster's communications fabric. A network of 100 SL390s blades can handle 77 teraflops in practice, and this accurately reflects the "de-rating" that occurs as scale increases, according to HP.