How GPU-Accelerated Databases Are Helping Advance Cognitive Computing

Artificial intelligence has been around in different forms for decades, but only now is it truly coming of age. AI experiments have made it out of the lab and are now being deployed into the real world by leading technology companies. For example, Amazon has smarter buying recommendations, Facebook has automatic facial recognition and Netflix offers movie recommendations. These things are possible because more horsepower is now available. Graphics processing units are helping provide much of the new power required to run these data-intensive apps. This eWEEK slide show offers some expert industry data points on the topic from Amit Vij, CEO and co-founder of database maker Kinetica.

Opportunities abound for AI, such as helping provide a more complete customer experience (CX) for shoppers and enabling buying recommendations for online retailers, judging the risk of various trade decisions at a bank, accelerating drug development in life sciences or providing smarter planning for logistics companies.

Data scientists typically rely on costly, complex and specialized tools and hardware. They frequently need to copy large volumes of data into these specialized environments to build and train their models. Once the AI models are working in the lab, it can be difficult to make that functionality available to business users. Finally, it’s difficult to find, train and retain the people with the right skills to develop, deploy and manage AI models.

GPU acceleration, which has been developed and optimized for years in the videogame world, immediately brings a 10X to 100X performance boost to business analytics. With streaming data and increasingly large datasets, many business users suffer from slow performance and complexity of their analytics solutions. GPU acceleration brings huge performance improvements and simplifies IT infrastructure for faster time to value.

Data scientists can develop AI models on live data via user-defined functions. A GPU database is especially fast at the matrix and vector operations common with AI workloads.  User-defined functions can call out to third-party machine learning/deep learning libraries such as BIDMach, Caffe and  TensorFlow.

Data scientists save time and improve results because they no longer need to extract data from separate systems if using a GPU database. They can build, train and deploy models using the same hardware and data as is being used for business analytics.

AI algorithms can be made available to business users in much less time. Using a tool such as Kinetica, for example, business tools can call UDFs directly through a REST API. This makes it possible to make custom algorithms available to business tools through easy-to-use, point-and-click interfaces.

The CPU has become the bottleneck for many large analytics systems, resulting in server sprawl for handling complex analytical workloads. GPU-accelerated databases improve efficiency and customers have been able to reduce their hardware footprint to one-fourth or one-tenth the size of conventional analytics systems.

Hardware costs can be connected to business results. The benefits of investment in hardware for AI can be seen quicker when that hardware is made available to the business. GPUs are gaining acceptance as the next big thing for the data center and are now available on the cloud platforms such as AWS, Azure and Google Cloud.