"Recently, Xilinx, a global FPGA chip giant, held a developer conference in Beijing and announced the launch of versal, the industry's first adaptive computing acceleration platform, and alveo, a data center and AI accelerator card.
Versal ACAP integrates scalar processing engine, adaptive hardware engine and intelligent engine as well as cutting-edge memory and interface technology. Hardware and software can be programmed and optimized by software developers, data scientists and hardware developers, and can provide powerful heterogeneous acceleration function for all applications.
Versal product portfolio is based on TSMC 7Nm FinFET process. It is the first platform to combine software programmability with hardware acceleration and flexible adaptability in specific fields. The product portfolio includes versal prime, versal premium and HBM series, which can provide industry-leading performance, connectivity, bandwidth and integration functions for the most demanding applications.
The product portfolio also includes AI core series (AI core), AI edge series (AI edge) and AI RF series. It integrates a new hardware module AI engine specially designed to meet the new requirements of low delay AI inference in various applications. At the same time, it supports advanced DSP implementation schemes to meet the requirements of wireless, radar and other applications. It is closely combined with versal adaptive hardware engine to support overall application acceleration of software and hardware, so as to ensure the highest performance and efficiency.
The first in the portfolio is the versal prime series, which provides a wide range of applicability across multiple markets and is optimized for connectivity and online acceleration of various workloads. At the same time, versal AI core series is launched. Compared with the industry-leading GPU, AI inference performance is expected to be improved by 8 times.
In addition to versal's product portfolio, Xilinx also launched data center and AI acceleration cards alveo U200 and u250 based on Xilinx ultrascale + FPGA.
In terms of machine learning, the real-time inference throughput of alveo u250 is 20 times higher than that of high-end CPU. Compared with accelerators with fixed functions such as high-end GPU, alveo u250 can improve the performance of low latency applications less than 2 milliseconds by more than 4 times. In addition, alveo accelerator card can reduce the delay by three times compared with GPU, providing significant performance advantages when running real-time inference applications. Some applications such as database search can be fundamentally accelerated, and the performance is more than 90 times higher than that of CPU.
"The launch of alveo accelerator card has further promoted the transformation of Xilinx to a platform company, enabling the growing application partner ecosystem to accelerate innovation faster than ever before." Manish mutHal, vice president of Xilinx's data center, said, "we are glad that customers are highly interested in alveo accelerator, and we are also glad to cooperate with our application ecosystem to jointly launch various producible solutions using alveo to customers."
It is reported that alveo has been supported by the ecosystem of partners and OEM manufacturers. The applications developed by 14 ecosystem partners such as bigstream, blacklynx and ctaccel can be immediately deployed. The key applications developed and certified include AI / ml, video transcoding, data analysis, financial risk modeling, security and genomics. In addition, top OEM manufacturers such as Dell, EMC, Fujitsu Limited and IBM are also cooperating with Xilinx to certify multiple server SKUs using alveo accelerator cards.
A universal and flexible platform is the future
Adaptive computing acceleration platform (ACAP) is a new generation of chip architecture launched by Xilinx in March this year. It was developed by Xilinx over four years, 1500 engineers and more than US $1 billion.
Xilinx CEO Victor Peng said that at present, it is in the early stage of the AI revolution. Many technology giants are building AI chips, and different industries are looking for different solutions. ACAP will not replace all GPUs and ASICs, but in the future, only a few industries may have fixed chips (such as ASICs) that can play a role, and a general and flexible platform is required for most scenarios in the future.
For Chinese start-ups to make AI special chips, Victor Peng believes that their choice is not for technical reasons. First of all, he said that he did not assert that startups could not do ASIC, but that AI startups should focus on innovative algorithms and architectures.
"How many startups are successful because of ASIC?" Victor Peng said bluntly, "if you can do better than Intel, NVIDIA and Xilinx, but more start-ups should focus on specific fields and applications rather than developing chips from scratch, because most enterprises are doing it.
As for the current hot AI application market in China, Victor Peng believes that the AI innovation speed in the Chinese market is very exciting. At present, in China, the emergence of new technologies and the launch of new products are very fast. The flexibility of Xilinx products can not only enable Chinese AI users to innovate faster, but also need not be limited to the speed of chip iteration (1-2 years).
It is worth mentioning that for Xilinx, which just completed the acquisition of Shenjian technology, a domestic AI start-up, Victor Peng said "this is a good way out". He pointed out that since Shenjian technology is mainly an engineering department, after the acquisition, the company will report to Salil raje, vice president of Xilinx's software and IP products department, and Xilinx will take over and further expand Shenjian's market and other business departments. Xilinx will help Shenjian technology serve global customers, use its technology, and conduct joint marketing in business.
Victor Peng revealed that the main reason why Xilinx values deep learning is their achievements in network optimization, DNN and some architecture and practical technologies. After being acquired by Xilinx, Shenjian technology will focus on the field of FPGA and no longer develop its own AI special chip. He hopes that the team and Xilinx can complete the integration as soon as possible. It is expected that Shenjian technology will bring substantial revenue in fiscal year 2020.
Original address: https://www.eeboard.com/news/versal/
Search aiban.com and pay more attention. You can master the latest development board, intelligent hardware, open source hardware, activities and other information every day. Recommend attention!
[wechat scanning can be followed directly]“
Our other product:
