GPU computing and energy efficiency will be key to a host of new and enhanced high-performance computing offerings at the Supercomputing show starting Nov. 14.
And according to John Lee, vice president of advanced technology solutions for Appro, his company won't be the only one at the show in Portland, Ore., talking about the increasing role of graphic processing units in mainstream computing.
"GPUs, from our perspective, are really making a third processor a real valuable option for HPC," Lee said in an interview, referring to CPUs from Intel and Advanced Micro Devices and GPUs from Nvidia. "Last year, everyone was talking about [Intel's] 'Nehalem' [architecture]. This year [the focus will be on] the new 'Fermi.'"
Nvidia has been a key driver in pushing GPUs into general-purpose computing environments, particularly for parallel computing workloads. Now the GPU maker is readying Fermi, its much-anticipated next-generation CUDA architecture.
At the show, Nvidia will demonstrate new Tesla products based on the Fermi architecture, according to its Website.
Appro will be showing off offerings based on its Supercomputing Cluster Architecture, and will feature systems running Intel Xeon chips, AMD Opteron processors and Tesla GPUs.
Appro also will give attendees a look at an upcoming server powered by AMD's Opteron 6100 processor-code-named Magny-Cours-which is due out in the first quarter 2010.
One of the systems Appro will showcase will be its HyperPower GPU performance clusters, a hybrid system running both Nvidia Tesla GPUs and Xeon processors from Intel. The system supports up to 304 CPU and 18,240 GPU cores in a 42U (73.5-inch) standard rack.
Appro also will tout its Xtreme-X1 supercomputer that will include the vendor's ACE (Appro Cluster Engine) management software. ACE is designed to automatically discover and manage all the server and networking hardware. The software comes with such tools as a workload manager, resource management resources, remote control and advanced power management.
New additional features include fully stateless compute nodes, dynamic instant provisioning, revision control and support for full installation of Red Hat or CentOS distribution. The diskless booting of the compute servers enables ACE to boot 10 to 10,000 compute nodes quickly and simultaneously.
The Xtreme-X architecture groups high-performance servers into an integrated "Scalable Unit," which can be managed and provisioned as a single supercomputer.
The enhanced system also features the new MPI 4.0 for improved I/O performance. Appro worked with Intel on making some of the improvements, Lee said.
The system, which is based on Intel's quad-core Xeon 5500 Series "Nehalem EP" chips, gives users interconnect fabric options of single-rail InfiniBand Network for price/performance environments or dual-rail for businesses concerned with performance and availability.
Appro also will show off its HyperGreen integrated clusters that offer up to 80 GreenBlade nodes, with 640 processing cores per 42U rack. The cost-competitive GreenBlades, which are designed to consume less energy than traditional blades, are powered by AMD's six-core "Istanbul" Opteron chip.
Also at the show, Appro will demonstrate a technology sample of its Appro 1143H server, which will be based on the Magny-Cours Opteron G34 socket system. In addition, the company will showcase a new 1U (1.75-inch) system powered by Intel's Xeon 5500 Series chips with two independent, dual-socket nodes in a 1U chassis. The system will include high memory capacity and flexible interconnect capabilities to fit into a wide range of HPC environments.
Appro also will unveil its next-generation GPU blade system that will run on Nvidia's Tesla products.