Tmall Double 11 2018 set many new records in the world of technology, one of which was that Alibaba made personalized recommendations for users on the whole platform on November 11
Together, the recommended images would circle the earth 70 times.
When you buy buy buy buy buy on Tmall/Handtao, the pictures will be transcoded in different formats or resolutions, which requires the background system to have strong computing power to guarantee the transcoding requirements of several times as much as normal. FPGA is the best energy ratio scheme for programmable chip processing this kind of parallel computing. The once cold “Wang Xetang Qiyan” is enabling more and more technological innovation through the value addition of cloud computing.
At this year’s XDF 2018 Conference, ali Cloud engineers unveiled the mysterious veil of Ali Cloud FPGA platform “Sainty sky” to the industry for the first time, and this new engine of cloud computing acceleration showed its edge.
The biggest characteristic of FPGA is that it is directly based on hardware programming and has the characteristics of high throughput and low latency. It is easy to build a data parallel channel and complete pipeline parallel at the same time. Compared with ASIC, it also has the programmability and flexibility of software, which makes it perfect for AI and other scenarios that have rigid requirements for large computing power and are still in the rapid iteration period.
FPGA capability model
However, the traditional FPGA offline development process is very complicated. IP providers or solution integrators need to develop and maintain their own hardware, which is difficult to promote and costly. For customers, development is not around the hardware link, trial and error costs, long cycle.
FPGA+ cloud is the best solution to solve these problems.
Ali Cloud engineer Zhang Zhenxiang introduced,
“Sainty” platform integrates multiple links of the traditional FPGA industry chain: original chip manufacturers, hardware, IP suppliers, etc., all of which are uniformly packaged into platform resources, so that customers can focus on the realization of computing power, while IP suppliers can simplify the delivery process, greatly optimize the delivery cycle and operating costs.
Since August 2017, Saintymarine has gone through three generations of product iterations. The latest example of F3 series launched can provide more than 10 million logic units and up to 47 TeraMACs of DSP computing power.
Taking the timing database as an example, the processing efficiency of the saintian platform is more than 30 times higher than that of the CPU: the measured peak performance of the single-channel FPGA implementation is 2.97GB/s, compared to the measured performance of the single-core CPU compression of about 90M/s. In the image recognition scenario, the computing power provided by SAINty Marine F3 is equivalent to about 7 CPU servers, greatly reducing Capex cost. Compared with CPU servers,
.
Moreover, this capability can be extended, as developers dealing with high-performance computing and machine learning often need multiple chips to compute together. The SAINty Marine provides various high-speed interfaces, implements the 1/2/4-chip interconnection topology, and supports dynamic configuration. Taking in-card dual-chip communication as an example,
, simultaneously supports a variety of lightweight transport protocols,
.
Inside Alibaba, Sainty Marine is already battle-hardened,
Take Double 11 as an example. When customers browse commodities through Taobao/Tmall APP or web, pictures need to be presented in different formats or resolutions. The parallel computing capability of shuntian platform,
Sainty Sun has also performed well in the genetic arena, shortening the 100-gbp genome analysis time to two hours, faster than in the past
Unit cost is also greatly reduced. Hologene refers to the analysis of the entire genome sequence of a species cell from the first DNA to the last DNA, which can identify almost any type of mutation on the genome.
In the eyes of traditional software engineers, FPGA chip development threshold is very high, on the one hand, need to understand HDL hardware programming, on the other hand, need to know the underlying hardware and interface.
At the XDF conference, Aliyun engineer Pan Yue introduced the solution of sainty Platform.
First, “Sainty Marine” added support for advanced system languages, including C, C++, OpenCL, etc. This means that developers can work smoothly with programming languages and workflows they are familiar with, and ensures that the original CPU or GPU-based development can be easily ported to FPGA devices in the cloud. At the same time, Ali Cloud also provides a large number of reference design and development packages to help developers quickly establish the system.
Interface aspect, “sainty Marine” provides unified standard bus interface, including: DDR4/5, PCIe, DMA, etc., developers do not need to work, only need to care about the internal logic implementation
.
In addition, Ali Cloud also supports a variety of mainstream FPGA Shell+Role implementation, can be adapted to cross-platform multi-boot scheme, so that users can easily complete cross-platform migration.
As the most important part of the FPGA ecosystem, IP vendors are very concerned about multi-tenant environments on the cloud. You need to ensure that IP products are not compromised and that the IP purchased by users can only be used by them.
To this end, Ali Cloud adopts IP encryption, network isolation, IP ownership and use rights separation and other methods to ensure that users’ source code (or IP source code), distribution packages, deployment environment are completely invisible to the third party in the whole process. At the same time, it also provides health monitoring ability, timely perception of underlying faults and real-time alarm.
Dr. Haile Yu, CEO of CTAccel, one of the earliest IP providers of Sainty Marine, said,
Security is our top concern. Ali Cloud’s security isolation is in place, which is why we choose Ali Cloud first for deployment.
Previously Ali Cloud’s Slogan
Has been deeply rooted in the people, actively preaching inclusive computing. Since the introduction of heterogeneous computing dependencies, developers can use more powerful GPU/FPGA instances at less cost on Aliyun, cutting development costs in deep learning by more than half.
In order to enable more developers to enjoy the benefits of FPGA, Ali Cloud launched the FPGA elastic cloud server and established the FPGA mirror market, so that IP products with innovative ideas and mature solutions stand out through the platform, expand the market scale, accelerate the development progress, and improve the design stability.
Zhang Zhenxiang said,
Sainty is a powerful tool for Ali Cloud to enable artificial intelligence industry. We welcome more developers to join ali Cloud FPGA ecology and promote technology to be implemented in the industry.