Big data and machine learning has already proven itself to be enormously useful for business decision making. However, CPU intensive activities such as big data mining, machine learning, artificial intelligence and software analytics is still being held back from reaching its true potential. Even with virtually unlimited computing power available through the cloud, getting fast results from massive databases has remained a challenge. The CPU model simply isn’t built for this type of use case. Fortunately, GPU computing has come to the rescue.
The challenges of a slow compute time
In the era of artificial intelligence (AI), data isn’t just big, it is massive. AI may be working with zettabytes of data and managing input as varied as IoT sensors, streaming video, natural language processing, machine learning, and connected vehicles. Slow run times in the arena of advanced analytics constrain the questions that can be asked, and creating iterative queries can be tiresome for data scientists. The ability of analysts to leverage their full creativity is also impaired because there simply aren’t enough resources and time for experimentation. From a technical standpoint, current workarounds to make data easier to manage are insufficient. Sampling misses the mark and pre-aggregation struggles at scale.
According to Alex Sabatier, Global Account Manager for Enterprise Sales at NVIDIA, adding GPU computing resources to the CPU mix offers much greater computing ability and speed. Analysts can ask better questions and go beyond the basics to dive deep into the data. With enough resources to fully explore all the data, the results are more accurate. Even the outliers and long-tail events are included for examination to see what insights they might deliver. It’s also easier to achieve significant ROI without the hidden costs of scaling out CPU infrastructure.
GPU computing is the right match for big data performance
GPUs are already well known in the gaming world where ultra-fast, graphic-intensive rendering is essential for a satisfactory user experience. These high power computing components are designed specifically to handle mathematically intensive tasks. According to Sabatier, there were more opportunities to explore outside of gaming. “We have been slowly moving into the enterprise world. There is a processor we invented for real time graphic processing that works within massively parallel architectures. We realized the architecture is suitable for AI and deep learning. So, we modified the processors and made other changes to make them work within the enterprise ecosystem.”
In fact, NVIDIA began working with clients across a variety of industries to accelerate deep learning, visualization, databases, and even entire data centers. Examples of potential use cases where faster analytics with GPU technology made sense included identifying optimal drilling and extraction sites for oil & gas companies or allowing pharmaceutical firms to reproduce and anticipate a drug’s effects on a molecule.
What about the existing infrastructure enterprises already have in place? Fortunately, Sabatier revealed that GPU is not a replacement for CPU architecture. Rather, it is a powerful accelerator for existing infrastructure. Within the context of an application, there might be a small percentage of code that hogs most of the compute resources. That code would be routed to the GPU while the remainder of the sequential code would be handled by the traditional CPU.
The accessibility of GPU computing
NVIDIA is taking steps to help it go as wide as possible, setting standards in the industry. The company has created a parallel computing platform and programming model called CUDA and made it open to the community. This open source platform is proving to be quite popular. “Now there are already 400 applications that use this language.” With contributions from the open source community, GPU development is certainly set to take off. This area of innovation may well provide the missing puzzle piece for organizations that can’t get enough value out of their data at the current speed of business.
According to Sabatier, “People are struggling to manage their data and query it. They know there are insights, but exploiting it takes a lot of time. Business owners need to extract and act on data insights as fast as possible. All traditional companies that move into Big Data are trying to find ways to accelerate analytics. With GPU, there is often a 50-100x acceleration factor.” It’s hard to argue with those metrics, and the future will definitely see more enterprises and data centers adding GPUs to the mix.
Is a GPU processing system a good fit for your project?
The things you need to know about GPU processing
The ins and outs of 3D graphics and GPU processing