NBA players are improving performance with AI on Azure AI infrastructure

With its electrifying pace, world-class athletes, and compelling on-court storylines and rivalries, it’s no wonder that professional basketball boasts a wildly diverse fan base and is one of the world’s most watched sports. In fact, the 2023–24 National Basketball Association (NBA) Regular Season—across ABC/ESPN/TNT/NBA TV—was consumed for a total of 762.69 million hours, making it the most-consumed NBA Regular Season in five years.

"We know that high levels of player skill and team performance thrill audiences and increase fan engagement,” says Charlie Rohlf, Vice President, Stats Technology Product Development at NBA. The NBA Stats Team is responsible for capturing all data for every game, for each player and every play. The team then shares stats and insights with coaches, broadcasters, and fans as games are in play. The Stats Team also shares larger reports on levels of play with teams so they can assess and improve both performance and strategy. “Technology can help us highlight and deliver the best data—and that ultimately helps the NBA community connect with the game in new ways," says Rohlf.

Caroline McKee, Senior Software Engineer, Machine Learning at NBA, agrees and says she has been a fan of the NBA since she was a kid and played basketball too. “Fans always want players to be performing their best, and with data we can strategically help teams do that, which always benefits the fans.” McKee says the NBA Stats Team wanted to invest in an advanced analytics solution with a specific focus on helping to optimize player performance, from training regimens to in-game strategies.

But with 30 NBA teams, more than 500 NBA players, and each team playing 82 games per season, not including playoffs, there is an enormous amount of player data to collect and analyze. The NBA Stats Team wanted to explore how AI could help them process that data and deliver analyses not just on shots made and rebounds, but on players’ specific body movements and things like speed, dunk height, number of passes and dribbles, and even injury risk—in real time.

Delivering more details and insights

Historically, player statistics were only recorded manually by an official basketball statistician. Even today, the official box score stats only include a few hundred events. But with the new computer vision–based player tracking system, the NBA is now gathering millions of data points for every game. To handle the stepwise increase in data, all coming in with subsecond latency, the NBA turned to Microsoft Azure for storing and processing this data.

“When we started with Azure, much of our existing infrastructure was on-premises with legacy-style virtual machines in the back office,” says Rohlf. “But now we are able to run a very successful hybrid environment.” Rohlf adds that his team was already using .NET Framework and Microsoft SQL Server, along with other Microsoft products, and that “now our legacy systems connect seamlessly with new infrastructure on Azure. We get the power of Azure to handle the size of the player tracking data and incredible synergy connecting the new and legacy systems.”

But with Azure infrastructure and AI, the NBA Stats Team knew it could do more to enhance the performance. “That means delivering more details and insights into those details,” says Rohlf. The NBA tracks data on 29 specific points on each player’s body simultaneously with the aim of capturing, processing, and delivering detailed player insights live, in the midst of the action.

Once the player data is pulled from Hawk-Eye, the NBA’s camera data provider, the Stats Team runs a series of microservices based on Azure Kubernetes Service (AKS). AKS helps orchestrate and manage various AI models, ensuring scalability and reliability, which is critical since the data is coming through at 60 hertz. “Sixty times a second we get data from all 29 points on each individual player,” Rohlf says. This averages out to be about 15 to 16 gigabytes of raw data a game, not including RGB video signals—sometimes more if the game goes into overtime. 

McKee says that the team has a “whole microsystem set up,” built mostly on Kubernetes pods with metadata stored in Azure Cosmos DB for scalability, speed, and efficiency in high-speed data retrieval. “It is a tremendous amount of data, and we store all of it in Azure Cosmos DB.” With its elasticity and dynamic scaling, Cosmos DB can efficiently handle variable workloads and contribute to cost-savings.” In terms of delivering data to consumers via the NBA API and social media, McKee adds that it’s all run on a Redpanda cluster, which runs on a series of Azure virtual machines.

The information is pulled into a real-time pipeline of 3D models where, McKee says, the NBA Stats Team can analyze everything. “From the height of a dunk to who’s more effective in the paint. We can pinpoint 29 distinct points on a player’s body. We see how a player moves. We see which way they’re facing on specific plays,” McKee says. “That kind of detailed data is extremely valuable to the teams, and for fans.”