We had a great trip to Barcelona for FabCon and while I was really excited about the new tools for creating ontologies from your business data (because that's what makes metadata into a living model of how your business works), it was the economics of how Fabric is using 'old' GPUs for accelerating SQL queries without making you change the way you run your data warehouse and lakehouse that caught my editor's attention. Microsoft definitely isn't the only company trying this out, but they hands down have the easiest and most mainstream option for adopting this.
The research behind this was also fascinating to dig into, even though I couldn't fit much detail into the piece - and CoddSpeed is just a fabulous pun.
GPU-accelerated databases aren’t new but the way Microsoft is doing them is – and it puts them a step ahead of the competition
I've always been dubious of the idea that an AI bubble popping would a lot of stranded hardware. the neoclouds COULD use their older GPUs for lots of things that aren't frontier models, but I don't think they're nimble enough. hyperscalers already busy doing just that to speed up database queries
GPU acceleration isn't new; nor are GPU-accelerated databases but they've never been mainstream @andrewbrust.info notes. you need a second SQL engine or switch to an engine most people haven't heard of. Microsoft putting it right into the Fabric platform as a switch speeds up existing SQL queries
The really clever bit: you don't need the latest, most expensive GPUs to speed up your database so it can keep up with all the extra queries agents are throwing at it. you get faster queries and Microsoft gets more like out of older GPUs, @tinot.bsky.social tells me.
Not only are older GPUs not getting scrapped faster (a suggestion that was doing the rounds last year); so many teams at Microsoft who aren't working on frontier models have ideas to use them that there's internal competition, which may gate availability of database acceleration. Definitely demand
GPU acceleration
capex
hardware lifecycles
database engines
research
Fabric Data Warehouse
IBM
Starburst
Spark
Presto
Trino