3 ms·
Success! Meanwhile, all 7500 nodes are, computationally, replaced by a 96 core, $10k server, in a dude's basement. With power to spare.
by b112 4y ago
Success! Meanwhile, all 7500 nodes are, computationally, replaced by a 96 core, $10k server, in a dude's basement.
With power to spare.
- intelVISA 4y agoBut I thought Good System Design involved reserializing the same data multiple times across the cloud(tm) and had a dedicated SRE and infra team - it's cheaper than one sys admin!
- jmillikin 4y agoYou'd generally want each of those 7500 machines be a full-sized server. No point running Kubernetes on tiny VMs, since its purpose is to provide bin-packed scheduling in a datacenter.
- electroly 4y agoThis isn't some dipshit enterprise running LOB software. This is OpenAI. These are all giant multi-GPU nodes getting slammed all day with machine learning jobs.
- mardifoufs 4y agoYeah, I'm sure openai could've trained gpt4 on a 10k$ machine.