Artificial Intelligence thread

[tphuang]

Please, Log in or Register to view URLs content!

Inspur has created 元脑 R1 推理服务器 inference machine for DeepSeek full version. Provides 1128 GB HBM3e storage, enough for FP8 precision inference. Video memory bandwidth of machine can reach 4.8TB/s
GPU P2P bandwidth reaches 900 GB/s
Single machine can support 20-30 concurrent users on latest inference framework.

It seems to me that they are probably using Nvidia H100 GPUs under the hood but don't want to mention it.

 

[GulfLander]

"Europe will invest an additional 50 billion euros ($51.6 billion) to bolster the bloc's artificial intelligence ambition, European Commission Ursula von der Leyen said on Tuesday (February 11).

It will come on top of the European AI Champions Initiative, that has already pledged 150 billion euros from providers, investors and industry, von der Leyen told the Paris AI Summit.

Von der Leyen said investments will focus on industrial and mission-critical technologies.

Europe is not late to the AI race, she said, describing Europe's brand of AI as driving innovation and adaptation, focusing on adopting complex applications, and being diverse and inclusive."

 

[tphuang]

"Europe will invest an additional 50 billion euros ($51.6 billion) to bolster the bloc's artificial intelligence ambition, European Commission Ursula von der Leyen said on Tuesday (February 11).

It will come on top of the European AI Champions Initiative, that has already pledged 150 billion euros from providers, investors and industry, von der Leyen told the Paris AI Summit.

Von der Leyen said investments will focus on industrial and mission-critical technologies.

Europe is not late to the AI race, she said, describing Europe's brand of AI as driving innovation and adaptation, focusing on adopting complex applications, and being diverse and inclusive."

this is great. every country should be building their own AI industry if they want to have a place in the future global economy.

I'm skeptical they will be successful in having a SOTA foundational model but at a minimum they can have their own ecosystem built around open/closed source models.

 

[tphuang]

https://twitter.com/i/web/status/1889643199787872438

Cerebras is apparently crushed by the demand for inference chips on DeepSeek R1

 

[european_guy]

this is great.

this is a joke.

Before EU can seriously attempt at AI, or at any other high tech, EU citizens have first to fix some deep issues with their politicians...VdL is one of the issues. btw for you to know, she and her commission are not elected by citizens, but by opaque deals among parties...and (foreign) lobbies

https://twitter.com/i/web/status/1889643199787872438

Cerebras is apparently crushed by the demand for inference chips on DeepSeek R1

It seems Cerberas runs only on the distilled models, not the full-blooded one

Please, Log in or Register to view URLs content!

 

[luminary]

https://twitter.com/i/web/status/1889643199787872438

Cerebras is apparently crushed by the demand for inference chips on DeepSeek R1

For them demand is "really high" because they don't have any capacity.

I would not be bullish on Cerebras. They've been around for a long enough time and their team and product is still meh. Additionally, there are reasons they don't put actual metrics out on mlperf. The thing with Perplexity was a demo no actual traffic.

 

[tokenanalyst]

Please, Log in or Register to view URLs content!

 

[tphuang]

this is a joke.

Before EU can seriously attempt at AI, or at any other high tech, EU citizens have first to fix some deep issues with their politicians...VdL is one of the issues. btw for you to know, she and her commission are not elected by citizens, but by opaque deals among parties...and (foreign) lobbies



It seems Cerberas runs only on the distilled models, not the full-blooded one

Please, Log in or Register to view URLs content!

you can guess the reason why certain American inference chips are not running the full blood one, but are running the Llama 70B version.

Although, I was surprised by this yesterday

https://twitter.com/i/web/status/1889347665072173171

Groq founder is seriously anti-China, but I guess demand for Qwen 32B version is too high

 

[Hyper]

For them demand is "really high" because they don't have any capacity.

I would not be bullish on Cerebras. They've been around for a long enough time and their team and product is still meh. Additionally, there are reasons they don't put actual metrics out on mlperf. The thing with Perplexity was a demo no actual traffic.

Running the full model is difficult though. Also Cerebras has a big memory bottleneck. Huge sram but no hbm.

 

[tokenanalyst]

Please, Log in or Register to view URLs content!