spainter_vfx - stock.adobe.com
GPU implementation is about more than deep learning
Simulmedia is using GPU technology to power reporting tools, while eyeing future deep learning applications, helping to justify the cost of the hardware while building experience.
When you consider a typical GPU implementation, you probably think of some advanced AI application. But that's not the only place businesses are putting the chips to work.
"[GPUs] are obviously applicable for Google and Facebook and companies doing AI. But for startups like ours that have to justify capital spend in today's business value, we still want that speed," said Kyle Hubert, CTO at Simulmedia Inc.
The New York-based advertising technology company is using GPUs to make fairly traditional processes, like data reporting and business intelligence dashboards, work faster. Using a platform from MapD, Simulmedia has built a reporting and data querying tool that lets sales staff and others in the organization visualize how certain television ads are performing and answer any client inquiries as they come in.
Using GPUs for more than deep learning
GPU technology is getting lots of attention today, primarily due to how businesses are using it. The chips power the training underlying some of the most advanced AI use cases, like image recognition, natural language translation and self-driving cars. But, of course, they were originally built to power video game graphics. Their main appeal is speedy processing power. And while that may be crucial for enabling neural networks to churn through millions of training examples, there are also other use cases in which the speed that comes from a GPU implementation is beneficial.
Simulmedia, founded in 2008, helps clients better target advertising on television networks. Initially, the team used spreadsheets to track metrics on how clients' advertisements performed. But the data was too large -- Simulmedia uses a combination of Nielsen and Experian data sets to target ads and assess effectiveness -- and the visualization options were too limited.
Reports had to be built by the operations team, and there was little capability to do ad hoc queries. The MapD tool enables sales and product management teams to view data visualization reports and to do their own queries using a graphical interface or through SQL code.
Business focus pays off in GPU experience
Kyle HubertCTO, Simulmedia
Some benefits of a GPU implementation focused on a standard business process go beyond simply speeding up that process. Hubert said it also prepares the business to implement the chips in a more pervasive way and prepares for a more AI-driven future.
He said the process of predicting which ads will perform best during particular time slots and on certain networks is heavy on data science. Simulmedia is looking at adding deep learning to its targeting, and these models will train on GPUs. Hubert said starting with GPUs in a standard business application has helped the team build a solid foundation on which to build out more GPU capability.
"There's a lot of implicit knowledge that's required to get GPUs up and running," he said.
Aside from building institutional knowledge around how GPUs work, starting by applying the chips to more traditional use cases also helps to justify the cost, which can be substantial.
"They're costly when you say, 'I want a bunch of GPUs, and I don't know what kind of results I'm going to get,'" Hubert said. "That's a lot of capital investment when you don't know your returns. When you do a dual-track approach, you can say, 'I can get these GPUs, set them up for business users now, and I have a concrete ability to get immediate gratification. Then, I can carve out some of that to be future-looking.'"