iFlytek: Huawei's Ascend 910B's capabilities are basically comparable to Nvidia's A100, and they are working together to create a new base for my country's general artificial intelligence

According to news from this site on October 22, in the third quarter of this year, iFlytek achieved a net profit of 25.79 million yuan, a year-on-year decrease of 81.86%; the net profit in the first three quarters was 99.36 million yuan, a year-on-year decrease of 76.36%.

Jiang Tao, vice president of iFlytek, revealed at the Q3 performance briefing that iFlytek has launched a special research project with Huawei Shengteng in early 2023, and jointly developed a high-performance operator library with Huawei to jointly create a new model of China's general artificial intelligence. The base allows domestic large-scale models to be built on independently innovative software and hardware.

He pointed out that the current capabilities of Huawei’s Ascend 910B are basically comparable to Nvidia’s A100. At the upcoming iFlytek 1024 Global Developer Festival, iFlytek and Huawei will make further joint announcements on the artificial intelligence computing power base.

# He also mentioned that the company has been committed to achieving algorithm improvement and acceleration in engineering technology. After being included in the U.S. Entity List in 2019, the company was once again subject to U.S. sanctions on 28 leading Chinese companies and institutions in the fields of artificial intelligence, high-performance chips, and supercomputers, including iFlytek, on October 7, 2022. sanctions.

This site checked public information and found that HiSilicon Ascend 910 was released in 2019, and also launched the accompanying new generation AI open source computing framework MindSpore, and MindSpore has also been open sourced in 2020.

Currently, the Huawei Ascend Community has disclosed three models of Atlas 300T products, corresponding to Ascend 910A, 910B, and 910 Pro B, with a maximum power consumption of 300W. The AI ​​computing power of the first two is 256 TFLOPS, while the 910 Pro B Up to 280 TFLOPS (FP16).

##For comparison, NVIDIA A100 was released in 2020. It uses double-precision Tensor Core and is based on the Ampere architecture. Its power consumption reaches 400W, FP32 floating point performance is 19.5TFLOPS, and FP16 Tensor Core performance can reach 312TFLOPS.

According to Huawei’s official specifications, Ascend 910 Pro B is about 18% slower than A100.

Speaking of which, I would also like to mention the A800 chip. This chip is a model specially launched to solve the semiconductor export regulations of the U.S. Department of Commerce last year. The computing power and other parameters remain completely unchanged, but the transmission rate is reduced from 600GB per second to 400GB. Therefore, the U.S. semiconductor exports released this week The new rules close this loophole.

According to feedback from AI practitioners on Zhihu, even though Ascend 910B still has many minor problems, single card performance lags behind A800, and Arm ecosystem is lacking (should refer to CANN vs. CUDA), but

With the ban on Nvidia's advanced products, domestic manufacturers will be forced to choose Ascend. I believe that Ascend products will be more complete

, and domestic manufacturers can also complete the replacement by stacking capacity and increasing the scale of computing power clusters, at least There is not much overall difference in the field of large model training.

It is worth mentioning that the PyTorch Foundation officially announced on Wednesday that Huawei has joined the foundation as a Premier member. This is also China's first and global The tenth highest-level member of the PyTorch Foundation.

In addition, the latest 2.1 version of PyTorch has simultaneously supported Ascend NPU, and has updated a more complete third-party device access mechanism under the promotion of Huawei. Based on this feature, third-party AI computing devices can connect to the PyTorch framework without modifying the original framework code. Ascend also provides an officially certified reference implementation of Torch NPU, which can guide third-party devices to easily connect.

Based on the new version, users can directly enjoy the native PyTorch development experience on Ascend NPU, and obtain models and applications that run efficiently on Ascend computing devices.

Related reading:

"Huawei and iFlytek launch AI storage base joint innovation project"

"iFlytek Liu Qingfeng: Huawei GPU can benchmark NVIDIA A100, General large model will benchmark against GPT-4 in the first half of next year"

Advertising statement: The external jump links contained in the article (including but not limited to hyperlinks, QR codes, passwords, etc.) are used to deliver more Information, saving selection time, the results are for reference only, all articles on this site contain this statement.

The above is the detailed content of iFlytek: Huawei's Ascend 910B's capabilities are basically comparable to Nvidia's A100, and they are working together to create a new base for my country's general artificial intelligence. For more information, please follow other related articles on the PHP Chinese website!