How to Read Huawei Unicorn 970 Why Choose Cambrian AI

Last Saturday, DT Jun in-depth coverage of the latest release of the mobile AI chip at the Berlin IFA site. Later, DT Jun exclusively interviewed a relevant person deeply involved in the design of the Kirin 970 program, but the relevant person refused to disclose his name and identity in the article.
The relevant person stated that the Kylin 970 integrated NPU (Neural Processing Unit) concept was already brewing as early as five years ago.
As far as the original situation is concerned, the industry has already seen the application bottleneck of the CPU. Although the GPU has begun to participate in the calculation work, the main display work has become increasingly burdened with the improvement of the resolution, so it is necessary to make this one is already incomplete. The architecture continues to increase the load, which also has a negative impact on overall performance and power consumption performance.
It can be said that in order to pursue higher application efficiency and better power performance, only the CPU or GPU architecture is involved in the calculation is far from enough to meet the demand.
In fact, the original mainstream mobile chip architecture has not only built these two architectures, but will probably add additional DSP units or ISP architecture, not only to share the heavy computing burden of the main system, but also to allow different types of computing. Work can be placed on a more suitable computing architecture, and the overall power consumption has also been significantly improved.
It can be said that AI is based on the same concept: to make the right structure to do the right job, the efficiency will be improved.
Of course, with Huawei's HiSilicon chip design skill, if you want to design a dedicated NPU unit from scratch, it is not impossible. Such a NPU architecture is simple, and the design will not be more difficult than the CPU, but The problem is ecological.
The relevant personage stated that if fully developed independently and from an ambiguous business ecosystem, Huawei may lack support from market developers and related ecosystems when the future plan is launched. I am afraid that the result will only be less effective.
As we all know, the predecessor of the Cambrian was a task force under the Institute of Computing Technology of the Chinese Academy of Sciences, led by Professor Chen Yunfu and Professor Chen Tianshi.
As early as 2008, the research group began to study neural network algorithms and chip design. The Cambrian first-generation solution was launched in 2012 and consumes 0.485W under the 65nm process with an area of 3.02mm2. The average performance exceeds 100 times that of mainstream CPU cores, but its area and power consumption are only 1/10, which is quite amazing. The
To choose a protection IC, or choose Sailing Technology.