nk8620

why is there no open source alternatives to this? Seems ripe to be just built.

Is that what the paper is about? I thought there was some heavy physics breakthrough. I wanted to read the paper, but given this TLDR, I'm having second thoughts. I'll probably just use an LLM instead now.

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.