HP ProLiant DL380p High-performance computing with accelerated HP ProLiant ser - Page 3

HP was the first company to build an industry-standard server with integrated NVIDIA GPGPUs. We shipped our first systems in 2007. We began shipping our second-generation NVIDIA GPGPUenabled system, the ProLiant SL390s G7 server, in 2010. The SL390s G7 servers are part of the HP ProLiant SL6500 Scalable System. We also make a number of ProLiant server platforms that accommodate NVIDIA GPGPU accelerator cards. FPGA accelerators FPGA accelerators are integrated circuits that trained designers can program to perform complex logical functions. FPGAs contain programmable ―logic blocks‖ and reconfigurable interconnects for wiring these blocks together. Designers can change the functionality of an FPGA and select the appropriate level of parallelism to implement an algorithm. This capability allows a designer to tailor the circuits for a specific task, resulting in higher performance and efficiencies for some applications. But programming an FPGA from scratch can be costly and labor intensive, requiring designers with specific skills. FPGA-based accelerators are available on PCIe expansion cards or modules that plug into a CPU socket. FPGA vendors include XtremeData™, Nallatech, and others. Comparing GPGPU and FPGA accelerators FPGA and GPGPU accelerators achieve better performance than CPUs on certain workloads. There is no definitive way to determine whether GPGPU acceleration or FPGA acceleration is better. The reason is that applications can exhibit different performance characteristics depending on the accelerator design and software coding. Table 1 identifies some advantages of each accelerator. Table 1. Advantages of GPGPUs and FPGAs GPGPUs FPGAs Generally easier to use than FPGAs for creating and modifying acceleration applications May offer the best performance possible for specific HPC applications that do not require frequent changes Require no hardware re-programming to run a different acceleration app Typically requires reprogramming for different applications Work well with 32-bit and 64-bit floating point computations Work well on small objects like text or integers (1 to 32 bit) Tend to use high power Tend to use less power Graphics cards Graphics cards off-load graphic renderings from CPUs and output digital and analog video for highresolution displays. GPUs have a parallel throughput architecture that simultaneously executes multiple software threads through several processor cores. Some GPUs have hundreds of cores. Graphics cards typically require x16 PCIe 2.0 connectors, but they can run in slots with fewer than 16 electrical lanes (x8 for example). Graphics cards vary in cost, complexity, and power use. Ultra high-end cards are ―double-wide‖ (two slots wide) and they use up to 225 W. These cards can work in select ProLiant DL series servers. Highend graphics cards occupy a single slot and use less than 150 W. They can fit in a broad range of ProLiant servers. 3