Cozy Compute

Local ai - Quad Rtx Pro 6000 - Home Labbing

[Cozy Compute @cozy_compute]

For those of you wanting to get into local ai, it's still not too late. Start with a gpu first and get two 3090's for $2,500. Why 2? So you can learn to operate multigpu setups and get a feel for what different quants run like.

[Cozy Compute @cozy_compute]

@loktar00 @Tech2Wild Have you seen these? Do you think these are worth getting or just stick to the bambu?

https://t.co/RjSa3RiHY1

[Maxim Lobovsky @MaxLobovsky]

Professional level resin 3d printer now available for $700.

One cool thing you can do with it - print threads down to m2.5 directly in any orientation. No insert or soldering iron needed.

[Cozy Compute @cozy_compute]

@net_termina Do you see a use case for 4 rtx pro 6000 rigs? Something like deep seek 4.1 coding on 2 cards while glm 5.3 flash runs as the judge on the other 2 cards all while sharing a pool of ram to hold cached experts?

[Cozy Compute @cozy_compute]

@DigitalNav Once it's set up the right way how often do experience breakage? How much maintenance is there over a year?

[Cozy Compute @cozy_compute]

@rexsalisbury Won't happen.

The boomers have taken over Apple

[Jukan ✈️OCP 2026 @jukan05]

The most fucking valuable slide at OCP: Rubin Ultra NVL576 uses CPO for scale-up.

High-volume production remains uncertain, but it looks like they’ll try to make it happen.

[Cozy Compute @cozy_compute]

@Blackwellboy @ViC305 @Blackfrost_AI @DJLougen He says they used a "counting to 200" trick to manipulate the results which is using a prompt like count to 200 which the model will be able to do very fast

[Cozy Compute @cozy_compute]

@Blackwellboy @ViC305 @Blackfrost_AI @DJLougen He's saying he tried a model recipe from Mia but it wasn't near as fast as advertised because it was benchmaxxed. Then he warns that you can't just trust what's out there because influencers will exagerate for clout.

[Cozy Compute @cozy_compute]

@wikiwayne The judge role such as selecting the best solution or judging code and anything that requires obscure niche knowledge. But I did just buy 2 more rtx pro 6000. So I'm hoping with four I can fulfill the judge role locally

[Cozy Compute @cozy_compute]

@Blackwellboy @ViC305 @Blackfrost_AI @DJLougen

[Cozy Compute @cozy_compute]

@Blackwellboy @ViC305 @Blackfrost_AI There is an interesting nugget in the article this post links to. Guess who the author of the recipe is?

[ふみのり｜ AIと同化して生きたい @fumi_KPFstudio]

この記事を深く読んでしまった。すごい面白かった。
この人はRTX6000とかのグラボ系からdgx sparkも、実用目的で実践されているので参考になった。ローカルLLMの本当の価値は量だけではなく用途も制限が無くなる事だと思う。本当に自由になる。

ローカルLLM系の記事は割と、LLMの実用とは別の研究側のベクトル(レシピ作成や非検閲化とか)が多い傾向があったので、個人的には貴重な体験情報が知れた感じで嬉しい。

https://t.co/wl52uhUcW3

[Cozy Compute @cozy_compute]

@firstadopter @ChipsandCheese9 The motherboard that aims to bankrupt me.

Look at all those mrdimm slots!

[Cozy Compute @cozy_compute]

@juliarturc It's because of Dario. He gives off really bad vibes and the public just doesn't trust him. Contrast that with Meta. Zuck also gives off creepy autistic vibes but he successfully platformed Alexander Wang who is a perfect shepherd for what they are doing.

[Cozy Compute @cozy_compute]

@gabu3d_pl @coldniko @PeasantSmith Glm 5.3 flash. Very impressed with this model and can't wait till the next release

[Cozy Compute @cozy_compute]

@mudler_it It's peer pressure. People are afraid to say anything because they fear being ostracized. That's why I respect those that call it out.

[Melvin @MelvinInvests]

Samsung just showed HBM5 right here in the United States.

At the OCP Global Summit in San Jose, Samsung put HBM5 and custom HBM on display. HBM4E was on the floor too. The industry is still ramping HBM4, and Samsung is already showing the generation after next.

They also demoed CXL memory pooling with $AMD that lets AI systems share BIGGER pools of memory instead of each chip being stuck with only what sits next to it. Their pitch is speed, bigger memory pools, and keeping AI data closer to the chips. That is exactly where the bottleneck is in AI right now.

Custom HBM is the part I'm watching, it means memory built for a specific customer's chip and that makes Samsung much harder to swap out.

Samsung doesn't just want to sell memory anymore but wants a seat at the table where AI systems get designed and this comes just one day after Korean memory stocks got crushed on China fears.

[Cozy Compute @cozy_compute]

@MichaelGannotti Wishing you the best ❤️

[Cozy Compute @cozy_compute]

@EspenJD @TheZachMueller I'm going through the same process myself and have learned a lot from both of you. It's these little details that aren't available anywhere else but have a huge impact on time and money that brought me onto Twitter. Appreciate you guys sharing