Sportmaster implemented the Nvidia DGX-2 complexes for solving of tasks in the field of artificial intelligence

2020: Implementation of a HSS of Nvidia DGX-2

On January 20, 2020 the Softline company announced project completion which is designed to help the international network of sports shops "Sportmaster" to accelerate work of models of artificial intelligence. Thanks to implementation of the hardware and software systems NVIDIA DGX-2 intended for the solution of the most difficult tasks in the field of artificial intelligence, the retailer expects to improve process of demand forecasting and replenishment of trade inventories, to increase efficiency of marketing programs and personnel management of retail chain stores.

Competition gain, fight for the client, increase in efficiency, cost reduction, sharp increase in amount of data and complexity of processes stimulate business to use of artificial intelligence technologies. The Sportmaster company actively develops models of artificial intelligence for demand forecasting, optimization of management of commodity transactions, increase in conversion of sales, increase in efficiency of programs of loyalty.

Quickly to create, unroll and train models of artificial intelligence, Sportmaster company the solution capable to provide high performance in problems of machine learning and deep learning was required.

Experts of Softline suggested to use the hardware and software system NVIDIA DGX-2. This system developed by NVIDIA is intended for training of neural networks, solving of tasks of mathematical statistics and machine learning.

"Having experience in project implementation based on AI technologies, Softline suggested the customer to use the NVIDIA DGX-2 systems for acceleration of development of models of machine learning. Possibilities of the DGX platform allow to train neural networks several times quicker, than functionality of normal processors. Having implemented hardware-software the NVIDIA DGX-2 complexes, Sportmaster will solve relevant business challenges. First, acceleration of development of forecast models will allow to avoid deficit of goods on counters and accumulations of illiquid products in warehouses. Secondly, data analysis using methods of machine learning will help the retailer to increase efficiency of a marketing activity. For January, 2020 are actively developed the loyalty program, the quality of targeted advertizing is brought to higher level, marketing efforts prepare", 'Anton Yudin, the manager on sale of solutions of Softline company noted'

"The complexity of business processes constantly increases. Sharply the amount of data and quantity of the factors influencing a forecast indicator and subject to classification increases. Time necessary for the person for acceptance of optimal solutions is reduced. As an example — daily formation of the order of goods in retail chain stores from thousands of shops and tens of thousands of goods items. Or, say, attraction of necessary number of employees for work in shop in hours of peak load for each shop of network. Or preparation of marketing promo-companies with optimal parameters of segments (clients, goods, shops). In the conditions of restriction of resources and time of people is not able to solve so major problems of optimization effectively. The systems of artificial intelligence for which effective work specialized high-performance platforms, with acceleration of calculations based on GPU are required are capable to assist it in decision making. The team of experts on data analysis Sportmaster conducted researches on modeling of difficult business processes of the company, saved up considerable competences of development of production models of artificial intelligence. Jointly with specialists of Softline analyzed the market of GPU platforms and became interested in the hardware and software system NVIDIA DGX-2. Prepared requirements to a pilot project, created the working group of experts in data analysis Sportmaster, agreed with NVIDIA company about delivery to our data center of "younger brother" — the NVIDIA DGX Station servers. For the maximum use of opportunities of DGX it was necessary to rewrite scripts for data preparation and learning process, to hold testing of the selected fighting models of machine learning. In comparison with 10 a nodal cluster of BigData, results were very optimistic. The key role was played by GPU compatibility of specific algorithms in specific frameworks of artificial intelligence. Where there was no support of GPU (i.e. actually classical CPU worked for NVIDIA DGX), there was an expected sag of performance, in comparison with BigData. Where there was a support of GPU (especially with full implementation of the multi-threaded GPU mode), there was a sharp gain of performance. For example, on a gradient busting in Catboost the 30-fold gain, and on XGBOOST in H2O – 20-fold was recorded. Taking into account quite good results of testing and also active development of the RAPIDS project supported by NVIDIA (where more and more algorithms are transferred to GPU), the decision on purchase and implementation of the server NVIDIA DGX-2 complex was made. Time to develop requirements to the industrial platform on the basis of which experts of Softline company created the specification on the equipment came and in the shortest possible time executed delivery of two NVIDIA DGX-2 complexes to our data center. Further migration of fighting models of artificial intelligence on the NVIDIA DGX-2 production-platform confirmed successful results of pilot testing. Moreover, it was succeeded to receive an additional gain of performance at the expense of more powerful iron of NVIDIA DGX-2. Thus, in comparison with old architecture, "window of opportunities" for accumulation of functionality appeared. For example, transition with weekly to daily technology of complete training of models, significant increase in quantity of factors and possibilities of "echeloning" of models, use of new resource-intensive algorithms for quality improvement of the forecast and results of classification. Summing up the results, it is possible to consider project deliverables successful. Further, we are going to increase the rates of cooperation with NVIDIA and Softline companies, to contribute to the active development of a software part (open source the RAPIDS project)", 'Alexey Tyurenkov, the head of the department of data warehouses, business intelligence and researches of data of penal correction system of Sportmaster company noted'