Socionext big ideas of AI server systems and big data analytics leveraging little cores SoCs
Press release [Wednesday 25 October 2017]
Artificial Intelligence (AI) and Internet of things (IoT) are seeing an explosion of growth and applications. It is changing our life every day. The flood of data that is generated and collected every second and the computations required to allow smarter algorithms to manage that data more cost-effectively driving AI development and cloud solutions at a rapid pace. With such advances come new technology challenges and enable the disruptive innovations across data centers and edge computing nodes. The ear of Big Data is coming.
As companies of all sizes begin their digital transformation journey, the amount of data we're now generating is astonishing. Accessing electronics files, videos, digitized records and other digital format data from smart devices around the globe, it is reaching Petabyte based IP traffic in the data centers. Furthermore, the immense unstructured data makes the exponential growth of computer server systems. Data analytics infrastructures need to become faster, smarter and more robust than ever before.
Socionext Inc., one of major Japanese IC design houses with unique strength to designing imaging, networking, computing system-on-chip (SoC) products and solutions, maintain a high level of visibility in semiconductor industry. Facing these strong needs of new servers to address such a key role of Big Data and analytics solutions, Socionext set a very different approach to develop new creative ideas of server system.
In a recent interview, Shuichi Yamane, SynQuacer Sever project leader of Socionext, talked about the evolutions of server systems in the Big Data era. Yamane has long time involvement and well-recognized expertise in ARM based SoC projects. Through his years of participation in the Silicon Valley and witness first-hand server industry development, he has unparalleled experience with various types of SoC chips design experience and market trends.
Facing the impact of Petabyte based data volume growth, the enterprise is migrating to the cloud services and gaining the benefits of cost reduction, high efficiency and massive control upgradation. According to IDC report, there will be one-third of data stored in the cloud servers by 2020. This trend will drive the Big Data revolution to build the next generation of server systems and services to satisfy immense needs ahead of competition.
However, the penetration rate of ARM cored servers are still low in the market, Yamane added. Most of the ARM cored CPU chip makers use Cortex-A57 or Cortex-A72 to build the server CPU chips. But Socionext enters server SoC market with a very different product idea. Large number of little cores on a single chip will have almost the same performance of big core CPU by leveraging parallelization strategies. Socionext builds its new SynQuacer SC2A11 server SoC featuring 24 ARM Cortex-A53 cores, which are the 64-bit, v8-based, small die-sized low power cores.
Data centers aiming the solutions to fit Petabyte data analytics needs
The first thinking behind the idea to use little cores to replace big cores is energy efficiency goals. The higher the performance, the higher the power consumption and heat generated of a high clock rate and big core CPU. From one typical performance calculation, Yamane highlights, six SC2A11 chips are equivalent to the performance of Xeon E-type middle-end server CPU. Each SC2A11 SoC consumes 5 Watts of power. And six SC2A11 SoCs will spend 30 Watts power. Comparing with 100 Watts power consumption of Xeon E-type server CPU, the Socionext server SoC solutions will have two thirds of total power saving.
The second, to achieve more computing performance of multiple CPU server systems, Socionext builds up hyper scale parallel computing servers by inventing the new technology called Direct Data Transaction or DDT. This is a core concept of SynQuacer server systems. DDT is a proprietary data transaction protocol for handling Petabyte data throughput between multiple CPU chips in the server system. In the general server systems, the technology of 40Gbps Ethernet technology will have an effective data bandwidth around 4-Gbps because the protocol overhead of TCP/IP handshaking, Yamane explained. Socionext provides a real 10-Gbps bandwidth by leveraging DDT strength. This is a massive breakthrough of data handling capability in server systems.
Energy saving and high speed data throughput with proprietary DDT will provide a very massive strength when Socionext decided to enter the server systems market sector. At this moment, Socionext has verified the operation of sixty-four SC2A11 chips working in parallel. With this innovative scalability, the SynQuacer SC2A11 achieves the optimum balance between performance and power consumption, in a wide range of applications from the gateway, the edge, to the cloud.
Aiming AI server and indexing server for processing big data demands
Socionext is aiming two major high demand applications. The first is the indexing server. Search engine relies on indexes of most of the files to improve content searching performance. As the data growing in a very fast path, the index server needs the solution to handle Big Data. And Socionext hyper scale parallel computing servers are very suitable for this application.
The second target is AI applications. AI counts on various neural network frameworks to develop smart algorithms. Through the Deep Neural Network (DNN), the system will be trained to learn to solve the problems while massive data processing. Today, the data centers use neural network frameworks such as TensorFlow to deal with metadata for hyperparameter optimization. Leveraging the Multi-node hyper scale parallel computing structure, the SynQuacer server systems save tens of time to perform the AI training. It massively reduces total cost of ownership (TCO) and increases performance of overall systems.
In the data center, engineering teams use TensorFlow and Spark together to train and apply deep learning models. Through Spark and a cluster of servers to improve deep learning pipelines with TensorFlow, there are more use cases to find the best set of hyperparameters for neural network training and leading to a great reduction in training time and lower error rate. This called Hyperparameter tuning, which means the process to pick the right parameters leads to high performance.
When applying a trained neural network models at scale to the edge nodes, considering lots of battery powered devices equipped in the client sides, it is essential to consider a very low power consumption design to meet the requirements. In the recent demonstration cases, the third party developers present facial recognition solutions by implemented a unique design to use three or four SC2A11 chips for edge server or micro server and integrated with Socionext Hybrid Codec technology and dedicated chipset. This design approach not only cuts down the training time but also improves accuracy and gives a better performance to leverage the efforts of data centers. The configurable SynQuacer server systems to use different number of SC2A11 chips give a very good flexibility for fitting many demanding applications of AI development.
SynQuacer server systems will be launched in the first quarter of 2018
Yamane talks about advanced development of SynQuacer SoC planning. Socionext is thinking to integrate more little GPU cores through the same idea of hyper scale parallel processing technology. This will provide new solutions to fit strong needs for future edge servers having a combination of CPU and GPU together along with sufficient memory makes it a perfect testbed for deep learning. Socionext is pushing the optimum balance design for taking consideration of power consumption and computing performance.
One of the most prominent trends that have recently emerged in order to deal with Big Data is the growing importance of AI on Hadoop. The developer is looking for servers with High Performance Computing technologies fitting the requirements of AI infrastructure capable of quickly incorporating large amounts of data from a wide range of sources. For this reason, Socionext is planning to introduce DDT2 technology for preparation of dealing with the data volume of Zetabyte in the future.
Socionext is creating a new path for new AI server systems by introducing brand new server architecture with hyper scale parallel processing and high data bandwidth. Building this new server system, Socionext brings together the solution expertise of the Taiwan ecosystem partners to ensure the product quality to deliver the excellent performance. Meanwhile, Fujitsu, one of the key investors of Socionext, provides technical and market consultant to help to find the better market segment in Japan.
Socionext will begin mass production of SC2A11 SoC in October 2017. And the SynQuacer server systems will be released to market in the first quarter of 2018. For achieve this goal, Socionext has sent engineering teams to Taiwan and has close cooperation with Taiwan ecosystem partners to explore the potential business opportunities of AI server business. Big data and advanced analytics have swiftly moved to a set of capabilities that need to be deeply embedded across data handling and computing performance. Socionext takes a brave step to embrace this emerging market sector.
For more product information, please visit Socionext official web site at http://www.tw.socionext.com/. Linkage to Facebook page, please check https://www.facebook.com/socionext.tw
Mr. Shuichi Yamane, SynQuacer Server project leader of Socionext, talks about the development of SynQuacer Server systems showcasing Petabyte scale data processing strength.
DIGITIMES' editorial team was not involved in the creation or production of this content. Companies looking to contribute commercial news or press releases are welcome to contact us.