They aren't the minimal hardware configurations, but the ones in their respective class that they had access to. Besides that, I'd imagine the batch nature of

They aren't the minimal hardware configurations, but the ones in their respective class that they had access to. Besides that, I'd imagine the batch nature of training to do better with paging the weights due to being able to bundle more compute to each parameter chunk access. And/or more sensitivity to speed for interactive inference causing the 8xA100 configuration even though half as much memory on each would suffice.