Hacker News Reader

Comment by nk8620 | Hacker News Reader

> Moreover, for non-AI compute workloads, we offer only a single server type, equipped with one CPU and the same amount of DRAM (previously 64GB, now 256GB).

I'm reading something like this for the first time. Is this common across industry or only typical to Meta? In contrast, we use multiple instances for sub-components of our ML training pipeline.

View on HN

At least in my experience in a somewhat bespoke industry on the non-cloud side, things have been moving in that direction due to most of the reasons they identify (it also ties nicely into the global "Datacenter as a Computer" concept they hit on).

People in my industry have a tendency to over-optimize (technical folks love nothing more than to tinker), and in doing so create very unique deployments per group that require a significant amount of infra and operational support, and drastically slow down the rate of progress/change (not to mention the cost). When you peel back all the requirements it turns out we really only need three unique deployment options. Makes it all significantly less cumbersome.

One reason for using a single CPU is that the number of cores per CPU has become very high and only using a single CPU avoids a lot of complexity around Non-Uniform Memory Architecture where the OS should allocate data to RAM that is "closest" to the CPU it is being used by for best performance.

Not to my knowledge, no. If it's a virtualized server, maybe this works, but to my knowledge most companies try to maximize their hardware usage through a variety of slices/sizes based on underlying resources. I am not a datacenter expert however.

This wasn’t true for the server side compiler jobs I was doing. Those were multithreaded af.

Reply on news.ycombinator.com