Huawei Shows Off 384-Chip AI Computing System That Rival Nvidia's Top Product (msn.com) 25
Long-time Slashdot reader hackingbear writes: China's Huawei Technologies showed off an AI computing system on Saturday that can rival Nvidia's most advanced offering, even though the company faces U.S. export restrictions. The CloudMatrix 384 system made its first public debut at the World Artificial Intelligence Conference (WAIC), a three-day event in Shanghai where companies showcase their latest AI innovations, drawing a large crowd to the company's booth. The CloudMatrix 384 incorporates 384 of Huawei's latest 910C chips, optically connected through an all-to-all topology, and outperforms Nvidia's GB200 NVL72 on some metrics, which uses 72 B200 chips, according to SemiAnalysis. A full CloudMatrix system can now deliver 300 PFLOPs of dense BF16 compute, almost double that of the GB200 NVL72. With more than 3.6x aggregate memory capacity and 2.1x more memory bandwidth, Huawei and China "now have AI system capabilities that can beat Nvidia's," according to a report by SemiAnalysis.
The trade-off is that it takes 4.1x the power of a GB200 NVL72, with 2.5x worse power per FLOP, 1.9x worse power per TB/s memory bandwidth, and 1.2x worse power per TB HBM memory capacity, but SemiAnalysis noted that China has no power constraints only chip constraints. Nvidia had announced DGX H100 NVL256 "Ranger" Platform [with 256 GPUs], SemiAnalysis writes, but "decided to not bring it to production due to it being prohibitively expensive, power hungry, and unreliable due to all the optical transceivers required and the two tiers of network. The CloudMatrix Pod requires an incredible 6,912 400G LPO transceivers for networking, the vast majority of which are for the scaleup network."
Also at this event, Chinese e-commerce giant Alibaba released a new flagship open-source reasoning model Qwen3-235B-A22B-Thinking-2507 which has "already topped key industry benchmarks, outperforming powerful proprietary systems from rivals like Google and OpenAI," according to industry reports. On the AIME25 benchmark, a test designed to evaluate sophisticated, multi-step problem-solving skills, Qwen3-Thinking-2507 achieved a remarkable score of 92.3. This places it ahead of some of the most powerful proprietary models, notably surpassing Google's Gemini-2.5 Pro, while Qwen3-Thinking secured a top score of 74.1 at LiveCodeBench, comfortably ahead of both Gemini-2.5 Pro and OpenAI's o4-mini, demonstrating its practical utility for developers and engineering teams.
The trade-off is that it takes 4.1x the power of a GB200 NVL72, with 2.5x worse power per FLOP, 1.9x worse power per TB/s memory bandwidth, and 1.2x worse power per TB HBM memory capacity, but SemiAnalysis noted that China has no power constraints only chip constraints. Nvidia had announced DGX H100 NVL256 "Ranger" Platform [with 256 GPUs], SemiAnalysis writes, but "decided to not bring it to production due to it being prohibitively expensive, power hungry, and unreliable due to all the optical transceivers required and the two tiers of network. The CloudMatrix Pod requires an incredible 6,912 400G LPO transceivers for networking, the vast majority of which are for the scaleup network."
Also at this event, Chinese e-commerce giant Alibaba released a new flagship open-source reasoning model Qwen3-235B-A22B-Thinking-2507 which has "already topped key industry benchmarks, outperforming powerful proprietary systems from rivals like Google and OpenAI," according to industry reports. On the AIME25 benchmark, a test designed to evaluate sophisticated, multi-step problem-solving skills, Qwen3-Thinking-2507 achieved a remarkable score of 92.3. This places it ahead of some of the most powerful proprietary models, notably surpassing Google's Gemini-2.5 Pro, while Qwen3-Thinking secured a top score of 74.1 at LiveCodeBench, comfortably ahead of both Gemini-2.5 Pro and OpenAI's o4-mini, demonstrating its practical utility for developers and engineering teams.
There are always power constraints (Score:2)
SemiAnalysis noted that China has no power constraints only chip constraints
Of course there are constraints. But they can address them with other constraints, for example they could permit high power use only during times of high solar production. Their central control makes it feasible. Or they could just tell other users to f off when they want to do some training.
Re: (Score:2, Insightful)
The actual news isn't really the power use here. Moving the attention there is nothing but cope.
The news is the fact that China is successfully speedrunning their chip industry, and making a mockery of our sanctions and trade wars on the side.
Did any of us even yesterday believe that they would take on Nvidia any time soon? Yet here we are. Rest assured they are working on power use, too. But what are we doing at the same? Circle jerk and photo ops.
Re: (Score:1)
They've been doing it with electric cars too. The fastest production EV to lap the Nurburgring is now Chinese. https://www.topgear.com/car-ne... [topgear.com]
Re: (Score:2)
BYD has a charging system that can fully charge a vehicle in around 10 minutes, too. Unfortunately in spite of all the lip-flapping about "free trade" we're not allowed to buy them here.
Re: (Score:3)
Re: (Score:2)
Re: (Score:2)
That's because you don't know anything about the electrical industry, their electrical grid is the envy of power engineers the world over. It helps that they essentially started from a 'green fields' situation, rather than having a century of installed cruft to rip and replace. More important is that they allow engineering projects to be run by engineers rather than thundering herds of useless MBAs who wouldn't know which end of the screwdriver to hold onto.
Power is hardly an issue for China (Score:2)
The Chinese have a lot of very cheap solar power in desert areas and hydroelectric power in the Southwest. I don't think they think much about power.
Re: (Score:2)
Power is always a problem. No matter how much they produce, hardware can still consume it.
They could run 2.5x as many NV machines.
Imagine a Beowulf cluster of these ! (Score:3)
EOM.
Re: (Score:2)
Can you make that into a car analogy?
Predictable (Score:2)
China was perfectly willing to "drink the poisoned wine" of being dependent on American technology. But we decided they couldn't have it any more. Predictably, they are now creating their own replacements.
Our problem is we have an incompetent ruling elite that has run "the most powerful country on the planet" into the ground. And that didn't start in 2025 or 2017. They have been patting themselves on the back through failure after failure. They cover their tracks by managing perceptions rather than learn
Re: (Score:2)
China was perfectly willing to "drink the poisoned wine" of being dependent on American technology. But we decided they couldn't have it any more. Predictably, they are now creating their own replacements.
Our problem is we have an incompetent ruling elite that has run "the most powerful country on the planet" into the ground. And that didn't start in 2025 or 2017. They have been patting themselves on the back through failure after failure. They cover their tracks by managing perceptions rather than learn from failure. They deny the consequences and blame everyone but themselves for the failure.
This is with wisdom of hindsight.
When Huawei was banned, the expected result was that it would slowly fade into irrelevance. When the whole suite of bans came to Chinese phones and technologies and they disappeared from US, the expectorated result was the same as what happened with USSR and that the commercial sector for advanced technologies would dry up and would turn to the west to supply them and it would be a win for Apple and Samsung.
Re: (Score:2)
This is with wisdom of hindsight.
Perhaps, but I doubt our rulers will learn from it. Because the folks at Harvard and Yale don't learn wisdom from hindsight. They are smart people who can figure out why they were right to begin with. They'll blame Trump.
Re: (Score:2)
This is with wisdom of hindsight.
Many people, including myself, know perfectly well what would happen when Washington, with then president Clinton, sold out the US industrial base to China in the 1990s. So fuck you and your "hindsight."
I guess (Score:2)
It's kind of worthy of achievement. But like, it's 5x the number of chips, for only ~2x performance, 3.6x memory capacity, and 2.1x memory bandwidth, while consuming 4.1x the amount of power and taking up at least 4x as much physical space.
Also, though I can't be sure, this type of setup almost begs to suffer from diminishing returns, so I doubt it can be scaled up further
Re: (Score:2)
To be fair, getting in the ball park is already pretty good. These are numbers that were not easily possible 5 years ago, I would think. So even if you call it "a couple years behind", it's still pretty good.
Re: (Score:2)
The thing is, our oh-so-wise masters decided that if we banned our high tech chips that China would bend over and agree to be reamed because there was no other possible source in the universe. Instead they proceeded to design and build their own chips, with the result that within a couple of years Huawei's phones were using domestically created chips that in many respects exceeded Apple's M1. So instead our owners decreed that if no one sold them lithography systems they would have to give up their goals
But.... (Score:2)