Source: Hacker News
Article note: The intersection of not just two but three dumb sets of incentives?
The latest cycle of ML/AI BS is all low-precision linear algebra, and everyone wants to ride the hype train before winter arrives again.
LINPACK, the stalwart old ruler for the international HPC dick-measuring contest is mostly higher-precision linear algebra, and as always, once a metric becomes a target, you only get what you measure.
Packing more compute engines you can't possibly feed with any realistic workload is lower-risk than trying memory architectures, and especially easier to do without a bunch of collaboration between players, so it's what gets done.
Comments 