Scaling Kubernetes to 7,500 nodes

OpenAI Blog2021-01-25 16:000 views

We’ve scaled Kubernetes clusters to 7,500 nodes, producing a scalable infrastructure for large models like GPT-3, CLIP, and DALL·E, but also for rapid small-scale iterative research such as Scaling Laws for Neural Language Models.

Source：https://openai.com/index/scaling-kubernetes-to-7500-nodes

My favorites

PreviousUnderstanding the capabilities, limitations, and societal impact of large language models NextDALL·E: Creating images from text

Do you let your coding agent read production errors?

I've been wiring a coding agent into the error tracker I'm building, and I keep getting stuck at the same spot: the agen

0dev.to Discuss

10 Technical Interview Questions That Test How You Think, Not Just How You Code

Solving a coding problem is one thing. Explaining why your solution makes sense is another. You might write a working so

0dev.to Discuss

What's the one client-admin task you still refuse to automate, and why?

I've automated most of the boring client stuff by now. Invoice reminders, meeting notes into a summary, first drafts of

0dev.to Discuss

A New HttpClient on Every Call Exhausts TCP Ports. Reuse One Client, or Use IHttpClientFactory

A small .NET service can look healthy in a quiet test. Then it can fail on a busy morning because it opens a new network

0dev.to C#

When One Demonstration Steers The Robot

If robots adapt at run time by copying a single retrieved episode, quality control has to move from dataset averages dow

0dev.to Machine Learning

Fifty-Four Approval Cards Are Not Safer. They Are Push-Bombing Yourself.

TL;DR: Approval is a finite attention budget, not a wall that gets thicker with every prompt. The number of prompts is a

0dev.to Machine Learning

Scaling Kubernetes to 7,500 nodes

Related posts

Do you let your coding agent read production errors?

10 Technical Interview Questions That Test How You Think, Not Just How You Code

What's the one client-admin task you still refuse to automate, and why?

A New HttpClient on Every Call Exhausts TCP Ports. Reuse One Client, or Use IHttpClientFactory

When One Demonstration Steers The Robot

Fifty-Four Approval Cards Are Not Safer. They Are Push-Bombing Yourself.