I think this chip has a huge impact on Alibaba and plays a very critical role.
On September 25, Pingtou Ge took out the first AI chip Hanguang 800. Ali, who loves fancy names, did not miss this opportunity - "Hanguang" is one of the three ancient divine swords, and is closely related to the embedded CPU "Xuantie" released 2 months ago and the system chip platform "Xuantie" released 1 month ago. "No Sword" echoes this, adding another member to the martial arts weapons camp.
According to Zhang Jianfeng, in the industry standard ResNet-50 test, the inference performance of Hanguang 800 reached 78563 IPS, which is 4 times higher than the performance of the best AI chip in the industry; the energy efficiency ratio is 500 IPS/W , 3.3 times that of the second place, and is “the world’s highest-performance AI inference chip.”
Hanguang is not the only chip that entered the media's field of vision in September this year. In the same month, Huawei released the AI ??chip Ascend 910, which is also known as "the most powerful computing power in the world." There are two "strongest" in one month, but they are not the same thing. "The concept of chips is as big as the Internet. There are as many chips as there are applications on the Internet. Many chips are cheaper than vegetables." Zhang Jianfeng said.
But in fact, Shengteng 910 and Hanguang 800 are more similar than different. Both are AI chips, both deployed in the cloud, but the former is a "training" chip and the latter is a "reasoning" chip, which belong to the two stages of deep learning. The training chip focuses on absolute computing power, while the inference chip pays more attention to comprehensive indicators such as unit energy consumption computing power, latency, and cost.
Cloud computing manufacturers’ independent core manufacturing is a new trend that has only emerged in the past two years. Behind this is the rising cost of the entire industry. This month, Hou Jinlong, president of Huawei Cloud & AI Products and Services, said that the server cost ratio of Huawei Cloud data center has exceeded 60%, and it will be difficult to make a profit in the future by buying servers from outside parties to provide cloud.
Intel, Nvidia, AMD and other companies have long monopolized the chip manufacturing of cloud servers. According to data from DRAMeXchange, more than 90% of the world's servers are currently using Intel's x86 architecture.
"Alibaba is a newcomer in the chip field. Xuantie and Hanguang 800 are the first step in Pingtou's long march. We still have a long way to go." Zhang Jianfeng said.
Huawei’s thinking is different from Alibaba’s. Although it does not sell chips separately, unlike Alibaba's idea of ??"selling chips with the cloud", Huawei's approach is to sell "chips with servers".
For Alibaba and Huawei, the development cost of self-developed chips is high and the cycle is long. Putting it into their own products first is a quick way to ensure that the research and development costs will not be wasted. At least two Jiayun Factory’s own needs can consume most of the chips. In addition, Huawei's model has long-term significance for winning over hardware manufacturers and building Huawei's ecosystem.