Customers, partners, standards, products and tools! In the past two years from 2014 to 2016, Xilinx has worked step by step from the above four aspects and has made historic contributions to the application of FPGAs in ultra-large-scale data centers, especially The reconfigurable acceleration stack launched at SC 2106 this year at the Global Supercomputing Conference provides an acceleration engine for it to replace GPUs, CPUs and even similar FPGAs in data centers, machine learning and other fields.
Back in 2014, Xilinx was an unfamiliar name in the field of data centers. In less than three years, Xilinx has achieved countless milestone breakthroughs in this field, and the market has also won At the same time, it has become an important partner for many technological leaders to join hands.
From the customer point of view — 3 of the world’s seven largest hyperscale cloud service companies have adopted Xilinx FPGAs. Among them, Baidu announced in October this year that it has designed a Xilinx UltraScale™ FPGA pool to accelerate machine learning. infer.
Partnership — Qualcomm and IBM announced an accelerated strategic cooperation with Xilinx on data centers. The cooperation with IBM has built a storage network acceleration framework CAPI SNAP, which can help developers to use Xilinx FPGAs to more easily accelerate applications such as NoSQL.
Standards leadership — Xilinx has been leading the industry plan for the development of the smart cache coherent interconnect standard CCIX. In May 2016, Xilinx, AMD, ARM, Huawei, IBM, Mellanox and Qualcomm joined forces to form the CCIX alliance. In the past five months, the number of members of the plan has tripled from the beginning of its establishment.
Software-defined tools and products for data centers — SDAccel software-defined development environment for FPGA acceleration was released in 2014. In November 2016, Xilinx released the latest 16nm Virtex® UltraScale™+ FPGA product details using high-bandwidth memory and CCIX technology.
Let us review the relevant news released by Xilinx in promoting data center applications in the past two years:
-On November 18, 2014, Xilinx announced the launch of the SDAccel development environment for OpenCL, C and C++, which will increase the unit power consumption performance of the data center by 25 times.
SDAccel is the latest member of Xilinx's SDx™ series. It combines the industry's first architecture-optimized compiler, library, and development board that supports any combination of OpenCL, C and C++ cores. It is the first to achieve a completely similar CPU/ GPU development and runtime experience.
-On January 16, 2015, Xilinx SDAccel development environment for OpenCL, C and C++ successfully passed the Khronos conformance test.
Xilinx announced that its SDAccel™ development environment for OpenCL™, C and C++ has successfully passed the Khronos OpenCL 1.0 standard conformance test. The OpenCL standard provides software developers with a unified programming environment that enables them to write efficient and portable codes, which can easily accelerate the operation of a series of various algorithms on Xilinx FPGAs.
-On November 19, 2015, Xilinx and IBM announced a strategic cooperation to accelerate data center applications.
IBM and Xilinx jointly announced a multi-year strategic collaboration to use Xilinx FPGA accelerated workload processing technology on IBM POWER systems to create higher performance and higher energy efficiency data center applications. The teams of both parties have jointly developed open acceleration infrastructure, software and middleware to meet the needs of emerging applications such as machine learning, network function virtualization (NFV), genetic analysis, high-performance computing (HPC), and big data analysis.
-On October 8, 2015, Qualcomm and Xilinx teamed up to provide a heterogeneous computing solution with higher efficiency and performance to meet the technical requirements of the evolution of the next-generation cloud computing architecture.
The two parties announced a strategic technical cooperation to provide industry-leading heterogeneous computing solutions with new levels of efficiency and performance through FPGA-based dynamic workload acceleration on the Qualcomm Technologies server platform. This strategic cooperation covers multiple generations of products from the board level to highly integrated solutions. Target applications include computing acceleration, big data analysis, machine learning, storage and CloudRAN.
-On February 15, 2016, Xilinx released its data center ecosystem investment plan.
The new plan specifically targets emerging workload application solutions, such as machine learning, image and video processing, data analysis, storage database acceleration, and network acceleration, and is committed to further expanding cloud computing and NFV acceleration solutions. The plan is fully executed by "Xilinx Technology Ventures".
-On April 8, 2016, Xilinx joined hands with IBM to achieve FPGA acceleration through the SuperVessel OpenPOWER development cloud platform.
The Xilinx SDAccel™ development environment built into SuperVessel supports FPGA acceleration with C, C++ and OpenCL, and will provide support for the development of demanding applications including big data analysis and machine learning.
-On May 25, 2016, Xilinx announced that seven technology leaders have joined forces to jointly launch an open acceleration architecture for data centers and other markets.
AMD, ARM, Huawei, IBM, Mellanox, Qualcomm, and Xilinx have joined forces on the latest accelerator cache coherent interconnect (CCIX) specification to support multi-processor architectures and accelerators to seamlessly share data.
-On May 27, 2016, Xilinx announced the expansion of its 16nm UltraScale+ product roadmap, adding accelerated enhancement technologies for data centers.
Combining 16nm UltraScale+ programmable logic and high-bandwidth video memory (HBM) memory and new accelerator interconnection technology to meet the requirements of heterogeneous computing and support the newly announced accelerated cache coherent interconnection (CCIX) technology.
-On October 18, 2016, Xilinx announced that the membership of the CCIX alliance has tripled and the relevant specifications have been launched.
The number of CCIX alliance members has rapidly increased to three times the original number, and relevant specifications for alliance members have been released. As the founding members of the CCIX Alliance, AMD, ARM, Huawei, IBM, Mellanox, Qualcomm and Xilinx expressed their warm welcome to the new members. Through close collaboration, member companies have reached a consensus on the CCIX specification and standards. By using today’s existing ecosystem and improving it, it supports higher bandwidth, lower latency, and overall consistency, so as to meet the requirements of data centers. Interconnection requirements.
-On October 20, 2016, Xilinx announced that Baidu's data center uses Xilinx FPGAs to accelerate machine learning applications.
Chinese Internet search engine provider Baidu is using Xilinx FPGAs to accelerate machine learning applications in its China data center. The power consumption efficiency provided by Xilinx FPGAs allows accelerators to be deployed throughout the data center and can increase the unit power consumption performance ratio by 10-20 times. Baidu's optimized FPGA platform is specially crafted for machine learning applications such as image and language recognition. In addition, Baidu also plans to use these platforms to develop commercially viable self-driving cars.
-On November 14, 2016, Xilinx released details of a new 16nm Virtex UltraScale+ FPGA using high-bandwidth memory and CCIX technology. Four new devices revolutionized the memory bandwidth by 20 times, which can fully meet the needs of computationally intensive applications.
The new 16nm Virtex® UltraScale+™ FPGA with HBM and CCIX technology has the highest memory bandwidth. Compared with DDR4 DIMM, the memory bandwidth is increased by 20 times, and compared with competing memory technologies, the power consumption per bit is reduced by 4 times. These new devices are designed to meet the higher memory bandwidth required by computing-intensive applications such as machine learning, Ethernet interconnection, 8K video, and radar. They also provide CCIX IP to support the cache coherency acceleration of any CCIX processor. Meet the requirements of computing acceleration applications.
-On November 14, 2016, Nimbix, a leading HPC (High Performance Computing) cloud platform provider, announced that it has joined forces with Xilinx to expand the application of FPGA load acceleration in the cloud.
Nimbix announced that it has provided Xilinx SDAccel™ development environment on its Nimbix cloud platform, which can provide on-demand development, testing and deployment of FPGA-based accelerated workloads.
-On November 15, 2016, Xilinx reconfigurable acceleration stack provides the industry's highest computing efficiency, 40 times faster than CPU, and 2-6 times faster than other FPGAs
Xilinx launched a reconfigurable acceleration stack solution that can help the world's largest cloud service providers to quickly develop and deploy acceleration platforms. The FPGA-based Xilinx reconfigurable acceleration stack specifically designed for cloud-level applications, including libraries, framework integration, development boards and support for OpenStack. Through Xilinx FPGA, the reconfigurable acceleration stack solution provides the industry's highest computing efficiency: 40 times higher than x86 server CPU; 6 times higher than competing FPGA solutions.