News Huawei Ascend AI 910D processor designed to take on Nvidia's Blackwell and Rubin GPUs

[Admin]

Huawei's Ascend 910D AI accelerator is expected to match Nvidia's H100 performance on the GPU level, with full systems based on the Ascend 910D set to compete against Nvidia's Blackwell and Rubin based pods.

Huawei Ascend AI 910D processor designed to take on Nvidia's Blackwell and Rubin GPUs : Read more

 

[freedomnow]

This is why politicians should not make decisions on things they know nothing about. The only thing they are doing is making China pollute the earth more. Nvidia parts are more efficient per watt which allows them to train models with less power. This does not stop China from training models they can simply create larger more power hungry data centers.

This reminds me of stupid environmentalists who will block the US production of oil to force us to import it from the other side of the world. They think they are making a difference but in reality they are contributing more pollution to the Earth.

Critical thinking skills have depleted in this world and they just follow the leader like mindless zombies.

 

[Pierce2623]

This is why politicians should not make decisions on things they know nothing about. The only thing they are doing is making China pollute the earth more. Nvidia parts are more efficient per watt which allows them to train models with less power. This does not stop China from training models they can simply create larger more power hungry data centers.

This reminds me of stupid environmentalists who will block the US production of oil to force us to import it from the other side of the world. They think they are making a difference but in reality they are contributing more pollution to the Earth.

Critical thinking skills have depleted in this world and they just follow the leader like mindless zombies.

The only problem with what you’re saying is the Huawei chip is so uncompetitive that all the big Chinese companies are hoarding previous generation Nvidia stock. If Blackwell is platinum and Hopper is gold, Huawei Ascend is a cheap copper alloy.

 

[nookoool]

The only problem with what you’re saying is the Huawei chip is so uncompetitive that all the big Chinese companies are hoarding previous generation Nvidia stock. If Blackwell is platinum and Hopper is gold, Huawei Ascend is a cheap copper alloy.

These Chinese tech company have a decade of code built off CUDA. Their access to nvdia can be ban at any moment and Huawei has limited production rates, of course they will hoard while the hoarding is possible

 

[Pierce2623]

These Chinese tech company have a decade of code built off CUDA. Their access to nvdia can be ban at any moment and Huawei has limited production rates, of course they will hoard while the hoarding is possible

Oh Huawei is claiming they can run CUDA code. They’re not affected by any US ruling on the Zluda CUDA translation layer. They’re hoarding Nvidia because the purchase cost and operating cost on the Huawei makes them flat out uncompetitive even if they can run 4x the number of chips in a single cluster, and that’s just to be competitive with Hopper. Compared to Blackwell, they’ll be miles off. Blackwell is Nvidia’s first architecture ever built specifically for AI performance from the ground up

 

[regs01]

H100 only delivers 990 TFLOPS, just slightly higher of 910C. 2000 TFLOPS is B100.

 

[Pierce2623]

H100 only delivers 990 TFLOPS, just slightly higher of 910C. 2000 TFLOPS is B100.

I’m not talking about raw FLOP numbers. I’m talking about actual use cases and well benchmarked AI performance. If the 910c was truly competitive with a H100 at half the price(They’re actually well under half the price), then why are Chinese companies hoarding even weaker H20 chips? Also why does their new cluster need 4x the number 910c chips to reach “2x Hopper performance” ? You’re also neglecting the fact that running the Huawei chips is over twice as expensive in electricity for a given level of performance. At the power levels new data centers are pushing, that’s a MASSIVE expense.