AI models can now run 50% faster on existing GPUs for real-time ads

A new software technique makes large AI models run much faster on common graphics cards. This means companies can use AI for things like real-time online advertising without delays, serving more requests with the same hardware.

Running large AI models in real-time for applications like online advertising was limited by how fast graphics cards could process requests. Each small calculation on the card added a tiny delay, which added up. This new technique cuts those delays by half in some cases. It means companies can deploy more responsive AI services, or handle a lot more users with their current hardware.