News China claims to have developed the world's first AI-designed processor — LLM turned performance requests into CPU architecture

[Admin]

China’s leading scientific institution has taken the wraps of QiMeng, an AI-powered system designed to accelerate chip design.

China claims to have developed the world's first AI-designed processor — LLM turned performance requests into CPU architecture : Read more

 

[bit_user]

QiMeng-CPU-v2, which is claimed to rival an Arm Cortex A53

The A53 is a 64-bit, 2-way, in-order pipelined core with split code & data caches, vector floating point, branch prediction, and branch target prefetching. If it has all of that, I'd be impressed. It also has page-table support, which is a key detail that distinguishes it from microcontrollers. I wonder if their core is microcontroller-tier or something capable of running a full-blown OS, like Linux.

Given that the die shot shows only one core, I'm guessing it also lacks SMP support. That'd be another huge functional gap, if so, since it requires cache coherency support and atomics.

 

[Pierce2623]

The A53 is a 64-bit, 2-way, in-order pipelined core with split code & data caches, vector floating point, branch prediction, and branch target prefetching. If it has all of that, I'd be impressed. It also has page-table support, which is a key detail that distinguishes it from microcontrollers. I wonder if their core is microcontroller-tier or something capable of running a full-blown OS, like Linux.

Given that the die shot shows only one core, I'm guessing it also lacks SMP support. That'd be another huge functional gap, if so, since it requires cache coherency support and atomics.

Why you gotta be realistic and ruin the hype?

 

[Ktbpylon]

The fact that anyone keeps reading after the words "China claims" is baffling to me.

Sincere question...are any of their technological creations ever tested or verified outside of China?

 

[heffeque]

The fact that anyone keeps reading after the words "China claims" is baffling to me.

Sincere question...are any of their technological creations ever tested or verified outside of China?

Quite a few actually... there's plenty of BYD cars around the world, not only inside China (BYD already sells more cars around the world than Ford). Their 1 MW charger + battery tech does work as they say it works.

CATL batteries are used by multiple car brands around the world, and are far superior, safer and better built than LG batteries (3rd largest battery manufacturer, after CATL and BYD).

Toyota is joining Huawei and Xiaomi to improve their tech.
Volkswagen is seeking Oppo's help to get more advanced 5G in their cars.
Europe's main 5G backbone used to be Huawei (before sanctions kicked in).

You must be one of those that think that China is still how it was in the 90s.

 

[nookoool]

The fact that anyone keeps reading after the words "China claims" is baffling to me.

Sincere question...are any of their technological creations ever tested or verified outside of China?

Because you are reading this in english which means it is 99% just a copy n paste from some other article by a guy sitting in some western country that can not speak / read chinese natively

 

[bit_user]

Quite a few actually... there's plenty of BYD cars around the world, not only inside China (BYD already sells more cars around the world than Ford). Their 1 MW charger + battery tech does work as they say it works.

CATL batteries are used by multiple car brands around the world, and are far superior, safer and better built than LG batteries (3rd largest battery manufacturer, after CATL and BYD).

Toyota is joining Huawei and Xiaomi to improve their tech.
Volkswagen is seeking Oppo's help to get more advanced 5G in their cars.
Europe's main 5G backbone used to be Huawei (before sanctions kicked in).

You must be one of those that think that China is still how it was in the 90s.

5G, batteries, and electric cars are all areas they've invested in for more than a decade. Those have nothing to do with this "AI-generated CPU". So, I agree with @Ktbpylon that we need to see some hard data on this thing.

In contrast to your examples, we have only to look at how far short Moore Threads' S80 fell of its promises. That is also a much more relevant example than yours.

 

[bit_user]

Because you are reading this in english which means it is 99% just a copy n paste from some other article by a guy sitting in some western country that can not speak / read chinese natively

Google translate is pretty good, in my experience - it will even translate text in images! If they published any meaningful details about this thing, I'll bet they would be clear enough, when reading an automatically-translated version of it. At least, that was my experience, last time I used it to read about a Chinese CPU.

 

[jp7189]

AI is trained on existing data and cortex core plans are readily available. The fact that this can regurgitate something as complex as an A53 core is a testament to the training and inferencing techniques, but i don't believe this says anything about the ability to innovate something novel.

 

[bit_user]

AI is trained on existing data and cortex core plans are readily available.

They're readily available, if you're a business and have the money to by an ARM Technology License. However, any CPU designed by an AI trained on that IP could potentially be considered a derivative work. I think ARM wouldn't knowingly give someone a Technology License, if they knew their IP would be used in this way.

The fact that this can regurgitate something as complex as an A53 core

We don't know that it actually did. What they said was that it's as fast as one. That could mean a lot of things, but it's almost certainly not as complex as one.

 

[jmills1981]

However, any CPU designed by an AI trained on that IP could potentially be considered a derivative work.

Meanwhile PROPRIETARY code designed by AI trained off gpl licensed code gets a pass because of the "eew gpl eew opensource" false narrative.

 

[bit_user]

Meanwhile PROPRIETARY code designed by AI trained off gpl licensed code gets a pass because of the "eew gpl eew opensource" false narrative.

Fair point, but I'm not a lawyer, so I've already said about as much as I should on this subject.

 

[jp7189]

They're readily available, if you're a business and have the money to by an ARM Technology License. However, any CPU designed by an AI trained on that IP could potentially be considered a derivative work. I think ARM wouldn't knowingly give someone a Technology License, if they knew their IP would be used in this way.

I made two assumptions 1. ARM China has those designs (which is think is safe to say) and 2. ARM China doesn't give a darn what ARM HQ says as evidenced by their public fights.

 

[SirStephenH]

Everyone's focused on chat bots, but science and engineering are where AI is going to be truly disruptive.

The problem with stuff like this, at least for now, is that the AI could be ripping off numerous patents so humans would have to go through any results to weed out the violations. Even if the AI model was trained on data that didn't include patented information, the AI could come up with the same solutions as existing patents, which would still technically be a violation.

 

[bit_user]

The problem with stuff like this, at least for now, is that the AI could be ripping off numerous patents so humans would have to go through any results to weed out the violations.

I'm sure patent trolls are already busily training AI bots to search for infringements, much like the rights owners to music are trying to scan all the works of major artists (e.g. Ed Sheeran, Taylor Swift) for copyright infringements.

 

[J9000]

You must be one of those that think that China is still how it was in the 90s.

China consistently makes questionable and outlandish claims of major technological and scientific breakthroughs on an almost daily basis. In fact, you’d think China had solved every problem known to man by this point, based on the amount of propaganda floating around on the internet. It’s amazing what you can achieve when you steal everybody else’s intellectual property.

 

[thekingofALLmonkz]

Google translate is pretty good, in my experience - it will even translate text in images! If they published any meaningful details about this thing, I'll bet they would be clear enough, when reading an automatically-translated version of it. At least, that was my experience, last time I used it to read about a Chinese CPU.

Google translate is absolute trash at translating Chinese. GPT is unironically the best I've seen so far.

 

[bit_user]

Google translate is absolute trash at translating Chinese. GPT is unironically the best I've seen so far.

The couple times I've used it, it worked well enough for me to figure out what all the numbers meant. I was most impressed that it handled text in images, such as graphs.

Thanks for the suggestion!

 

[bit_user]

China consistently makes questionable and outlandish claims of major technological and scientific breakthroughs

To be fair, they're not the first or only ones to do that. We see Western companies sometimes make outlandish claims or promises. I also wouldn't paint all Chinese companies or researchers with the same brush.

I think the basic rule that applies in any case is that extraordinary claims deserve the strongst evidence.

 

[bit_user]

I made two assumptions 1. ARM China has those designs (which is think is safe to say) and 2. ARM China doesn't give a darn what ARM HQ says as evidenced by their public fights.

Even if they had access to the HDL for a core like the A53, conventional AI models need to see many examples of a thing, in order to learn how to replicate it. You can't just show them one or even a few examples and expect them to know in the same way a human might.

Another possibility is that it was designed using a technique like "vibe coding", where the humans took the role of guiding the high-level architecture and prompted it to flesh out all of the details. I didn't get that impression, based on the article, so I think what it made is probably very minimalistic and just done as a proof-of-concept.

 

[J9000]

To be fair, they're not the first or only ones to do that. We see Western companies sometimes make outlandish claims or promises. I also wouldn't paint all Chinese companies or researchers with the same brush.

I think the basic rule that applies in any case is that extraordinary claims deserve the strongst evidence.

Not trying to paint anyone "with the same brush." Read the South China Morning Post. China literally has a breakthrough every day.

 

[redgarl]

Typical CCP propaganda... in the meantime, accelerators don't exist...

 

[redgarl]

Quite a few actually... there's plenty of BYD cars around the world, not only inside China (BYD already sells more cars around the world than Ford). Their 1 MW charger + battery tech does work as they say it works.

CATL batteries are used by multiple car brands around the world, and are far superior, safer and better built than LG batteries (3rd largest battery manufacturer, after CATL and BYD).

Toyota is joining Huawei and Xiaomi to improve their tech.
Volkswagen is seeking Oppo's help to get more advanced 5G in their cars.
Europe's main 5G backbone used to be Huawei (before sanctions kicked in).

You must be one of those that think that China is still how it was in the 90s.

You cannot compare battery technologies and semiconductors.

We really love to blend them together, but they have nothing in common beside the requirements for chemical engineering for both.

In the meantime, all the processor designs for AI are American. Even their ARM counterpart are developed by American companies. Huawei is the only exception and we are far from being able to verify their claims, but I give them the benefit of the doubt because they are a huge company backed by the CCP with many fields of leadership.

 

[redgarl]

Not trying to paint anyone "with the same brush." Read the South China Morning Post. China literally has a breakthrough every day.

But they are behind 5-10 years...

 

[jp7189]

Even if they had access to the HDL for a core like the A53, conventional AI models need to see many examples of a thing, in order to learn how to replicate it. You can't just show them one or even a few examples and expect them to know in the same way a human might.

Another possibility is that it was designed using a technique like "vibe coding", where the humans took the role of guiding the high-level architecture and prompted it to flesh out all of the details. I didn't get that impression, based on the article, so I think what it made is probably very minimalistic and just done as a proof-of-concept.

There's more than one circuit in an A53; there are many, many example of how circuits are put together.

Nuance aside, I was trying to use that as an example to make my orginal point that I don't believe this approach is enough to make something new. It's only a way to spit out variations of something that's already existed.