News Huawei's new AI CloudMatrix cluster beats Nvidia's GB200 by brute force, uses 4X the power

[Admin]

Huawei unveils AI CloudMatrix 384 system based on 384 Ascend 910C processors that can beat Nvidia's GB200 NVL72 in performance, but at the cost of efficiency.

Huawei's new AI CloudMatrix cluster beats Nvidia's GB200 by brute force, uses 4X the power : Read more

 

[hotaru251]

China throws egregious power at its AI problem

The Intel answer I see xD
I mean it works short term i guess.

 

[jp7189]

Datacenter space costs a lot too, nvl72 can be deployed as a single rack vs 16 racks for this system.

Also, how do 384 accelerators use 6,912 800G network connections? 18 links per accelerator!?

 

[Alvar "Miles" Udell]

The Intel answer I see xD
I mean it works short term i guess.

AMD released their share of horribly inefficient products as well to attempt to compete against Nvidia and Intel. The HD 2900XT and the insane AMD QuadFX system come to mind. Those didn't work.

 

[hotaru251]

AMD released their share of horribly inefficient products as well to attempt to compete against Nvidia and Intel. The HD 2900XT and the insane AMD QuadFX system come to mind. Those didn't work.

yeah, but they figured out how to gain performance w/o rebounding to "more power!"
also 2900XT was likely not full amd as that came out 4 or so months after amd bought ati (and development of a gpu takes a long time)

 

[zsydeepsky]

if you actually look into the data, then you will see Huawei solution's W/TFlops (energy efficiency) is 2.3 times of Nvidia's cutting-edge solution.

just quoting Nvidia's own numbers: GB200 provides 25X better energy efficiency compares to Nvidia's last gen H100

NVIDIA GB200 NVL72

Discover the powerful GB200 NVL72 GPU, engineered for AI workloads and next-gen data centers.

www.nvidia.com

even if Nvidia was providing that 25X number by using FP4 to compare with FP16 performance, thus reducing the 25X to ~6X, it's still safe to say that Huawei provided a solution that's better than Nvidia's last-gen solution.

which puts Huawei only 1-2 years behind Nvidia, as an AI solution provider.

I was already quite bullish on Huawei, but this still exceeds my best expectations. I was only expecting they provide something similar to H100 at this point.

 

[phead128]

China throws egregious power at its AI problem

As inference compute trends towards electricity prices, it will be more of a matter of cooling and generating supply of cheap electricity. China's industrial electricity prices are already a fraction of US industrial electricity prices (see the Bitcoin GPUs set up alongside hydroelectric dams, and thorium nuclear reactors coming online).