Artificial Intelligence thread

[tphuang]

You deleted my post in which I was explaining why that has to be a monthly pay. I hope you are not thinking I was the one who assumed 9k to 13k USD as a yearly salary. I wish you had given a reason for the deletion to clarify.

And this complaint from you sounds quite patronizing. Personally I don't think it a big deal for some people mixing up because it is mostly a cultural difference. We are all here to share and to learn. If you want to make a thread your own stage, make it clear for us.

I don't really care you were later explaining that this monthly pay, because you made this an issue in the first place. And we had a page full of posts that were just not on topic.

And that is just not cool. I saw this article 2 days ago and didn't post it because I knew it would have caused this thread to go in the wrong direction.

And yes, I am going to enforce this thread to a higher standard. If you don't want to go with that, you can feel free to never post on this thread again.

 

[SanWenYu]

I don't really care you were later explaining that this monthly pay, because you made this an issue in the first place. And we had a page full of posts that were just not on topic.

And that is just not cool. I saw this article 2 days ago and didn't post it because I knew it would have caused this thread to go in the wrong direction.

And yes, I am going to enforce this thread to a higher standard. If you don't want to go with that, you can feel free to never post on this thread again.

Fine you are the mod you won.

 

[tamsen_ikard]

how powerful is huawei ascend 910c compared to nvidia blackwell?

 

[tphuang]

how powerful is huawei ascend 910c compared to nvidia blackwell?

I think I posted the link earlier where hyperbolic CEO said 910C's inference is at 60% of H100 at the moment with further optimization closing this gap further. I would say there are probably some rationale for this.
One is that the nominal computer is smaller (my guess is 80% of H100)
The processor and memory chips are likely inferior (Using Kunpeng vs Intel CPU & HBM2E vs HBM3)
Lastly, probably the one they can really work on, is just the low level optimizations where CUDA is super well tuned on PyTorch and there is plenty of work left for CUNN (apparently the language for Ascend chips)

And this is just me talking about inference. The gap in training is larger. Again, these are all things they can improve on, especially with low level optimizations and such.

 

[tphuang]

So for my work recently, we did some benchmark/testing of DeepSeek V3 vs GPT-4o, because that's just more relevant to the customer base. We found the V3 to be quite competitive in how they answered. Basically, it's usable for the applications we are developing. GPT is going to lose business on API side of things longer term just because it is too expensive. The question is just how public consumes DeepSeek APIs. Are they going to go through chat.deepseek.com (by the way, it uses same library as OpenAI for python) or one of the public open source servers like Huggingface, Sambanova and Hyperbolic or are people going to get their own hardware.

I'm personally a big fan of owning your own hardware. Then, you are not beholden to any server related issues.

 

[interestedseal]

Please, Log in or Register to view URLs content!

Official Huawei Cloud WeChat account confirms: Ascend-based full-size R1/V3 inference

 

[Bellum_Romanum]

Here we go..

we have some details.

"Based on a 14-month pay structure, the annual salary for this position can reach up to 1.54 million RMB (approx. $212,520) before tax.

The minimum salary is 300,000 RMB (approx. $41,400) for a client-side R&D engineer. Interns can earn 500 RMB (approximately $69 USD) per day, with non-Beijing interns receiving a housing allowance if they relocate to Beijing, and opportunities for full-time conversion are available.

Additionally, some platforms indicate that Deep AGI large model interns can earn between 500 to 1,000 RMB (approx. $69 to $138) per day."

@caudaceus @gadgetcool5

Before those folks open your mouth/fingers to be typing some dumb takes, we ought to exercise some COMMON SENSICAL question/logic in our heads if what we think believe to be true (mostly guided by our ignorance of contemporary China) then we should avoid making sweeping generalizations and assumptions that's going to come back and make us dumber than we already are. With today's availability and advancement in search engines and gen AI one would have thought to utilize those platforms before flapping your metaphorical mouthes blathering such utter erroneous nonsense.

Are we this idiotic and ignorant?

 

[broadsword]

"Scientists from the two countries have developed a breakthrough algorithm using information from reverse engineering video card accelerators. The algorithm allows gaming GPUs to be used for scientific computing[...].
The innovation was achieved by specialists from Shenzhen MSU-BIT University, co-founded by Lomonosov Moscow State University and Beijing Institute of Technology[...]It also means that Russia and China need to buy less NVIDIA GPUs[...]
These advancements highlight Russia’s technological expertise, allowing collaboration with China and potential future partnerships with India." Sputnikintl

https://twitter.com/i/web/status/1885425968854241661

I want that algorithm so I don't have to buy the 5090. I wish it could be made available to the world and then Nasaq would suffer another meltdown.

 

[SanWenYu]

I want that algorithm so I don't have to buy the 5090. I wish it could be made available to the world and then Nasaq would suffer another meltdown.

What are you going to use that algorithm for? It is quite specific to only certain scientific applications:

the new algorithm enhances the computational efficiency of peridynamics (PD), a cutting edge, non-local theory that solves difficult physical issues such as cracks, damage and fractures.

It requires CUDA so the impact will not be that big for Nvidia unless the algorithm can be ported to other architectures.

To address these challenges, Yang Yang, an associate professor, leveraged

Please, Log in or Register to view URLs content!

CUDA programming

Please, Log in or Register to view URLs content!

to create the PD-General framework. By making an in-depth analysis of the chip’s unique structure, her team optimised algorithm design and memory management that led to a remarkable performance boost. Their research was published in the Chinese Journal of Computational Mechanics on January 8.

More details of the said algorithm: News on China's scientific and technological development.

 

[OptimusLion]

Shenzhen Tencent Computer Systems Co., Ltd. announced today that the DeepSeek-R1 large model supports one-click deployment to Tencent Cloud "HAI", and developers can access and call it in just 3 minutes.

Please, Log in or Register to view URLs content!