Recently, Hangzhou Acceleration Cloud Information Technology Co., Ltd. (referred to as Acceleration Cloud), an industry-leading manufacturer of heterogeneous acceleration and business offloading solutions, is committed to providing an overall solution for heterogeneous computing acceleration. "Accelerating New Technology, Driving Smart Future" Science and Technology Summit and New Product Launch Conference in big cities. At the press conference, Acceleration Cloud released four major innovative products and three solutions, and invited Intel and Cytech experts to share solutions such as industrial Ethernet, FPGA-based OpenCL and gene acceleration.
In the era of big data, deep learning is the main driver of artificial intelligence. According to research firm Gartner, by 2018, 80% of data scientists will have deep learning in their toolkit. The success of machine learning and artificial intelligence projects requires more than just data and algorithms, but also a fusion of skills, infrastructure and business acceptance.
Wu Gang, Founder and CEO of Accelerator Cloud
The four major products of Acceleration Cloud this time are: two series of hardware acceleration products (SC-OPS, SC-VPX) ——
SC-OPS is the world's first Intel Stratix 10 FPGA accelerator card launched by Accelerator Cloud. It adopts Intel's latest 14nm process Stratix10GX2800 FPGA device, which can be widely used in data centers, cloud computing, machine vision, deep learning, high performance computing, simulation, Finance and other fields.
SC-VPX is the VPX blade acceleration platform with the highest computing density in the world. It adopts Intel Stratix 10 GX2800 device and is compatible with GX1650. It constructs an advanced, flexible and efficient signal processing and deep learning architecture in the industry. , rapid construction of product prototypes and algorithm development and application in deep learning related fields.
Two IP libraries (FDNN, FBLAS) -
The deep learning acceleration library FDNN is the first FPGA acceleration library in China that supports general-purpose convolutional neural networks. Based on RTL-level code, it can provide high performance and flexible configuration features. The high-performance computing acceleration library FBLAS is the industry's highest-performance RTL-level math acceleration library.
Three major solutions (deep learning solutions, high-performance computing and digital signal processing solutions, edge computing solutions) ——
Deep Learning Acceleration Solution : Acceleration Cloud has launched a complete set of FPGA-based deep learning acceleration solutions, including SC-OPM/SC-OPF/SC-OPS acceleration cards and FDNN acceleration libraries, to meet customer requirements for high-performance, flexible acceleration of deep learning . In order to facilitate customers to use high-level language development, Acceleration Cloud provides a complete OpenCL heterogeneous development environment based on FPGA to quickly implement user-defined deep learning acceleration solutions. At the same time, Acceleration Cloud also provides fast deep neural network customized acceleration services.
Digital signal processing solution : In response to the requirements of digital signal processing systems such as radar and communication, combined with Intel's latest 14nm process Stratix10 FPGA series, Accelerator Cloud provides a complete set of solutions combining hardware and software to achieve high-performance matrix Operations (matrix multiplication, transpose, inversion, QR decomposition) and ultra-fast FFT (Fourier transform). In order to facilitate customers to use high-level language development, Acceleration Cloud provides a complete OpenCL heterogeneous development environment based on FPGA to quickly implement user-defined signal processing acceleration solutions.
Edge computing solution : Accelerated cloud intelligent industrial control solution adopts high-performance Intel Arria10 GX660 device, with modular design, strong real-time characteristics and high-performance algorithm IP acceleration, and a complete OpenCL heterogeneous development environment, which can realize a new generation of high-performance edge Computing gateway, used in various industrial environments.
Why FPGA?
According to Wu Gang, founder and CEO of Acceleration Cloud , GPUs are particularly suitable for deep learning training. Deep learning is divided into two processes: training and online deployment (inference). It is usually necessary to collect enough images, build a model for training after labeling, and then deploy it. But in fact, the market and demand for training are not large, and GPUs have natural disadvantages, not only high latency problems, but also increased operating costs.
In contrast, FPGAs have absolute advantages in terms of operating costs, and are also superior to GPUs in terms of power consumption and latency. Wu Gang also said that the smart network card of the data center is also based on FPGA, because the GPU has no other interface except the display interface. After all, the GPU is specially created for display. But the FPGA is different. Each Pin can be used as an I/O, and the I/O can be used as a serial port or SPI. It can be used for various interfaces and supports customization, which is very useful for industrial environments.
The advantages of intelligent industrial control
Wu Gang pointed out that there are relatively big problems in the current industrial control industry. First of all, except for PLC, the rest of the industrial control is called an industrial computer. It is based on a PC machine used in an industrial environment, so it is called an industrial computer, which is actually a computer. But the processor used is 2 generations behind the notebooks that everyone uses now. After all, reliability and stability must be considered.
The industrial control environment requires the addition of cameras and sensors to the machines on all production lines, because the things that need to be processed are very complex. In the past, workers were also equipped with semi-automation, but now it is required to be automated or even intelligent. However, the X86-based industrial computer can handle 1-channel high-definition camera, but it is obviously unable to handle multiple channels.
In the future, the requirements for motion control will become higher and higher. For example, the precision and synchronization requirements of maglev trains must be precise, and the core is digital signal processing. FPGA can handle real-time and accuracy requirements very well. Therefore, Accelerator Cloud has made a new set of intelligent industrial control based on FPGA to replace the traditional old industrial control and solve the problems that these industrial controls could not solve in the past.
Finally, Wu Gang said: "Artificial intelligence has entered our lives, but there are still bottlenecks in the future development, which requires breakthroughs in hardware technology and algorithms. Heterogeneous computing is the future trend of computing architecture, and FPGA is the perfect solution to realize heterogeneous computing. Choice. Heterogeneous computing acceleration platform solutions that accelerate cloud innovation, with high performance, high efficiency, low latency, programmability and remote reconfigurability, are very suitable for the needs of elastic business on the cloud. We hope to be able to pass Our technology helps more enterprises implement deep learning and gain a head start in the era of big data."