Alibaba has made another attempt to change its name. Can it still make chips in a down-to-earth manner?
[Copy link]
At noon, I read the news that Alibaba Group announced that it had acquired the only self-owned embedded CPU IP Core company in mainland China, Zhongtian Microsystems Co., Ltd. Suddenly, I felt that Jack Ma was quite responsible. However, what I want to talk about today is that Alibaba Damo Academy is developing a neural network chip, Ali-NPU. On April 19, according to media reports, Jiao Yang, a researcher at Alibaba Damo Academy, said that Alibaba Damo Academy is developing a neural network chip, Ali-NPU. According to the design, the performance of Alibaba's Ali-NPU will be 10 times that of the mainstream CPU and GPU architecture AI chips on the market, while the manufacturing cost and power consumption are only half, and the cost performance is more than 40 times. In the future, the capabilities of Ali-NPU will not only better meet the needs of AI inference calculations such as image and video analysis and machine learning, but also output computing power through Alibaba Cloud and empower all walks of life. Ali did not disclose more progress on the specific development and launch time. At this critical juncture, the news disclosed by Alibaba Damo Academy will inevitably make people feel that it is a hot topic, after all, there are no more details on the specific development and launch time. Leaving aside other things, I personally feel that the performance is 10 times that of mainstream CPUs and GPUs. In 2017, the YouTube LinusTechTips channel gave a review that showed the progress of AMD and NVIDIA graphics cards over the past 10 years through game frame rate tests. The above picture is the performance test of NVIDIA graphics card HeavenDX9. The performance gap is about 10 times. The above picture is the AMD graphics card Heaven performance test. The gap between Fury and HD3870 is 13 times (there is a small error here. The R9390X graphics card is of the same generation as FuryX. Both were released in 2015, not 2014) The results show that the performance improvement is about 10-20 times, of course there will be big differences in different games. As for the CPU, Intel's 8th-generation Core has a 40% performance improvement over the 7th-generation Core, and a 2-fold improvement over the PC performance 5 years ago, which is still the largest performance improvement of all generations of Core. So I think the 10 times performance proposed by Alibaba is questionable. If we take AI applications into consideration, it is more complicated because it involves different hardware architectures and different benchmarks. At the same time, most neural networks are compressed and optimized before deployment, which makes the evaluation unfair. However, when I later searched for some press releases about AI processors, I felt relieved. The Kirin 970 chip is equipped with a dedicated neural network processor (Neural Network Processing Unit, NPU). Its new mobile computing architecture will be built for artificial intelligence. It has been optimized in terms of user language understanding and real-time image recognition processing. It is 25 times faster than the CPU and consumes 50 times less energy. Horizon Robotics uses an innovative technology architecture, namely the data processing flow mode of the Attention Engine + Cognition Engine. Through this combined algorithm, the chip's computing speed can be increased by more than 10 times. Of course, there is nothing wrong with doing this for marketing purposes. After all, it is a matter of Guan Gong fighting Qin Qiong. There is no general comparison. When it comes to actual application scenarios, real battles are still needed. In fact, Alibaba has a broader layout in the field of AI chips. Previously, Alibaba has invested in Cambrian, Barefoot Networks, SenseTime, Kneron, ASR, C-SKY Micro and many other chip companies.
| 
提升卡
变色卡
千斤顶
京公网安备 11010802033920号
