Dodgy Huawei Chips Nearly Sunk DeepSeek's Next-Gen R2 Model

DeepSeek's development of its next-gen R2 AI model was severely delayed after months of failed training attempts on Huawei's Ascend chips, which suffered from unstable hardware, slow interconnects, and immature software. The Register reports: Following the industry rattling launch of DeepSeek R1 earlier this year, the Chinese AI darling faced pressure from government authorities to train the model's successor on Huawei's homegrown silicon, three unnamed sources have told the Financial Times. But after months of work and the help of an entire team of Huawei engineers, unstable chips, glacial interconnects, and immature software proved insurmountable for DeepSeek, which was apparently unable to complete a single successful training run. The failure, along with challenges with data labeling, ultimately delayed the release of DeepSeek R2 as the company started anew, using Nvidia's H20 GPUs instead. The company has reportedly relegated Huawei's Ascend accelerators to inference duty.

DeepSeek has options

[williamyf]

Aside from Huawei and Cambricon, Deepseek also has Innosilicon, Moore Threads and Liusang as alternatives.

I guess that, for the R3 model, they will conduct small scale training trials to see if any of those work for them, to stay within chinese chips for training.

And it would not surprise me that they can achieve sway in to telling the "winner" what features they need emphasized in a future roadmap.

JM2C
YMMV

Re:

[Anonymous Coward]

Huawei has money and huge engineering capacity. They are coming from a long way behind in this area but they will almost certainly catch up, so doubtful that forcing it will be required in future, trump and Biden have virtually guaranteed the success of chip industry in China.

Re: DeepSeek has options

[drinkypoo]

So China would have just continued buying everyone else's chips forever and ever amen without these policies?

Re:

[larryjoe]

None of the Chinese AI chip makers have a workable solution (at least for training). How can I be so sure? Well, if they had a solution that sort of worked, then they would have dumped that product on a world market that is desperately yearning for an Nvidia alternative. That solution doesn't even have to be that good in terms of performance or power. It just has to work ... and be cheaper. The Chinese are good with subsidies and dumping to ensure a cheap price. However, these AI chips are complicated

Re:

[sg_oneill]

Yeah, right now the only plausible alternatives to the big nvidia monster chips seems to be the higher end of apples M3 & M4s , and even they are really only suitable for inference with light training at most. (The key here is memory. Apple's M range chips can use main memory as fast GPU memory allowing them to load in the really big models, at least the models with lots of ram do). Though i've heard AMD are making waves on this front.

There really is an opportunity here for Intel and/or AMD to get in on

Unsurprising

[DrMrLordX]

Given that the first DeepSeek relied on foreign hardware, it should come as no surprise that v2 is similarly dependent.

Wait what ????

[Bradac_55]

The Chinese agents here on /. keeps screeching about how far ahead China is in everything ... which is it?

Re:

[noshellswill]

I'll risk my (positive) karma again today. This post --- obviously from a paid Chinese agent --- deserves a much higher rating.   It ought to be read  by every American nationalist on Slash-Dot; and most of the globalists. The blatant arrogance and smooth deceptive delivery in his/her post  provides a measure of the dangers WE face from the chi.comz . The post is Chinese "take-out" and "washed-laundry" at the highest level.

Deets?

[bill_mcgonigle]

Does anybody have a detailed article or video about the challenges they faced?

This is seriously interesting material for microarchitecture nerds.

Re:

[Tablizer]

If Xi tells you, he'll have to kill you.

At least they tried

[allo]

Everyone is like "Without Nvidia we're nothing" and just buys overpriced cards. The Chinese at least try to build alternatives. And you bet that each good Nvidia card had a hundred bad prototypes. Nvidia just doesn't have so much pressure to test if one can train a model with a card that still has problems as the Chinese when it becomes harder to get Nvidia hardware. In the end it's only a matter of time, matrix multiplication in hardware is no witchcraft. And if there should be a new architecture (e.g. Bitnet that can work with binary operations and bitshifts) everyone starts at 0 again.