New Trick Makes AI Learn Up to 48% Faster on Mixed Chips
Faster, cheaper AI training usually means cheaper AI tools and quicker updates for you.
Modern AI gets smart in two stages. First it reads huge amounts of text, then it practices — trying tasks, getting feedback, and adjusting. That second stage, called reinforcement learning (teaching AI by trial and error), has become the expensive part. Because these practice sessions have outgrown any single data center, companies now stitch together chips from several locations. Some chips are fast and pricey, others slower but cheap. That's like running a kitchen with a fancy oven and a microwave — keeping both useful at the same time is hard.
The core problem is memory speed. Chips can only pull data so fast, so if you run many practice sessions at once, the chip stays busy but each individual session crawls. It's like a supermarket: opening more checkout lanes serves more shoppers per hour, but each shopper waits longer. AI teams want both — lots of work done and quick finishes — and those goals fight each other.
CadenceRL sidesteps the fight by reshaping the work instead of micromanaging each job. It chops long practice sessions into shorter ones so chips stay busy. When older tasks pile up and need to finish, it sends the remaining long ones to the fastest chips. And it prepares data ahead of time, before deciding which chip gets it. The result, on mixed hardware: up to 48% more work completed per second, and the slowest 5% of tasks finishing up to 64% faster — with no one manually assigning jobs.
Two honest caveats. This is a research paper, not a shipping product, and the numbers come from the authors' own test setups, so real-world gains may differ. Still, the direction matters. AI companies spend staggering sums on chips and electricity, so doing nearly half again as much work on the same hardware is like getting a chunk of your equipment free. Over time, efficiency like this tends to show up as lower prices, faster new features, and less energy burned. The broader playbook — match each job to the chip best suited for it — is one nearly every AI lab is now chasing.
- CadenceRL speeds up the practice-and-improve phase of AI training by sending the right jobs to the right chips.
- On mixed hardware it did up to 48% more work per second and cut worst-case wait times by up to 64%.
- It's a research paper, not a product, but efficiency gains like this often lead to cheaper AI tools later.
Why It Matters
Cheaper, faster AI training could mean lower prices and quicker new features in the AI tools you already use.