SALT LAKE CITY, Utah, Nov. 16 — Inspur and Intel announced partnership on an R&D project — the Field Programmable Gate Array (FPGA) (a programmable logic microchip) accelerator card F10A at the 2016 International Supercomputing Conference (SC16). This is a high density and high performance FPGA card based on OpenCL.
Due to the easy programmability of OpenCL’s high-level development language, the F10A makes a qualitative leap in Software Productivity. In Traditional FPGA development, developer requirements are higher and the development period is longer – leading to applications in high-performance computing being limited. Based on OpenCL, F10A substantially shortens the development period by adopting software high-level language and programming models. According to data provided by Inspur, an engineer can independently transplant a GZIP application within one month with F10A using OpenCL – while it would take a period of three months to complete the transplantation using Verilog.
F10A also features high performance, high density, and high applicability. Based on Altera’s Arria 10 microchip, F10A has a single precision floating-point performance of 1.5 TFlops, while the peak power consumption is only 35W, and the performance per watt is 42 GFlops. At the same time, F10A is designed for high-density half-length and low-profile, and has flexible on-board storage allocation and maximum support of 32GB memory, which is 4 to 8 times the industry average. Additionally, F10A supports two SFP+ GE/10GE interfaces, which can transfer data directly from the internet to the board card to be processed without needing to go through the CPU, thus greatly reducing transmission delay.
The F10A supports data parallelism in DNRange mode and task parallelism in the Pipeline mode, fully supporting low-latency, high-intensity applications such as high-performance computing, deep learning, data acquisition, high-frequency transaction, network processing and signal processing, etc.
With the explosive growth of HPC and AI, the performance and power consumption ratio of the traditional processor chip is a bottleneck. The OpenCL FPGA lies between a dedicated chip and a universal chip with programmability, reflecting distinct hardware reconfiguration and software-defined features, and in dealing with specific applications it has even more efficiency. Previously, Inspur teamed with IFLYTEK and Altera to successfully apply the FPGA chip to the online intelligent voice recognition sphere. This was 2.871 times faster than the CPU performance, while the power consumption was equivalent to 15.7% of the CPU performance. Power consumption increased by 18 times.