Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

A few weeks ago, OpenAI's models started refusing normal work for my business. So I built 3 servers with 12 AMD R9700 GPUs and moved the workload in house. $32.2k total. Here's the whole story in 2 minutes. #LocalAI

37,898 Aufrufe • vor 20 Tagen •via X (Twitter)

16 Kommentare

Profilbild von Nick Reese
Nick Reesevor 20 Tagen

Plan A was more Mac Studios. We already had six, but keeping up would have taken at least 41 more. That's $280k. Plan B was building our own. On our real workload, six Macs did 12 jobs/min. The 12 R9700s do 175. About 15x the work for about the same money

Profilbild von Nick Reese
Nick Reesevor 20 Tagen

Each server runs on a regular 15-amp home circuit. The GPUs do the AI work, so the rest of each server runs our CI/CD. I get the argument that "the cloud is cheaper." But owning my business is the real goal, and nobody can change these models on me.

Profilbild von Brian
Brianvor 15 Tagen

This is the way in my opinion. I don't know how CTOs and that right now are even able to do a budget with all the changes in cloud models that is out of their control and things can change over night with 0 notice.

Profilbild von Shawn Herrin 🔫
Shawn Herrin 🔫vor 15 Tagen

Nice, love to see it.

Profilbild von Nick Reese
Nick Reesevor 15 Tagen

I love the handle @WeLikeShooting

Profilbild von Alex Garcia
Alex Garciavor 15 Tagen

This is pretty awesome. Im curious about the harness, did you/your company did it or was it a third party? I see you are hitting 53 tok/s, whereas AMD publishes 51. You a have a pretty solid setup right there.

Profilbild von Nick Reese
Nick Reesevor 15 Tagen

Working on a video about the harness and how to build a good one as I'm rewriting production again for like the 10th time haha. What do you want to know. What should I include in the video?

Profilbild von Josh Builds
Josh Buildsvor 20 Tagen

respect, that's a serious level of commitment 💪

Profilbild von Mike Koker
Mike Kokervor 15 Tagen

Each server is solo? 4 cards per? What specs?

Profilbild von Tim Gilmour
Tim Gilmourvor 15 Tagen

love it. i've got a similar system, though only 2 R9700s (went with a Ryzen vs Epyc/Threadripper).

Profilbild von Alberto Álvarez
Alberto Álvarezvor 20 Tagen

Have you tried comparing the results of your model vs those of OpenAI and Anthropic? I guess what I want to ask is: are you not sacrificing a lot of accuracy?

Profilbild von Nick Reese
Nick Reesevor 20 Tagen

Well my work is getting refused by OpenAI and candidly Qwen 3.8 27b on our own workflows collects about 2 more data points than OpenAI Luna given the exact same prompts.

Profilbild von Just Tools
Just Toolsvor 15 Tagen

@AAGresearch You’re the one managing this 🦑 or you have a team of devs 👨🏼‍💻👩‍💻working for you

Profilbild von Nick Reese
Nick Reesevor 15 Tagen

@AAGresearch It is all me.

Profilbild von Reneil
Reneilvor 15 Tagen

this is the way 🕳🐇

Profilbild von armorpiercinggluten
armorpiercingglutenvor 15 Tagen

Everyone is gonna want to set this up for their businesses. At least I do. And maybe with water cooling so I can run a jacuzzi for free.

Ähnliche Videos