Tian Haili @ CSDN 2020-10-23
With the release of Huawei's Mate 40 Pro mobile phone and its Kirin 9000 chip, Kirin 9000 also topped the AI Benchmark list. In the article " AI Benchmark v4 Top Ranking: Dimensity 1000+ vs Kirin 990 5G ", I saw that Kirin 990 5G is doing optimization and improvement, but now it is not 990 5G, but the newly released Kirin 9000. At that time, it may just be the name unchanged. , Leading to reveal some AI performance data spy that thought it was 990 5G.
Recently, I have been paying attention to the AI-Benchmark list. As expected, in conjunction with the release of the Mate 40 Pro mobile phone and its Kirin 9000 chip, AI-Benchmark was updated for the first time. Let's take a look at the latest soc list [2020-10-23]:
Kirin 9000 leads the list with 148k points (vs 92.3k of Dimensity 1000+).
A detailed analysis of individual data found that it was as analyzed in " AI Benchmark v4: Dimensity 1000+ vs Kirin 990 5G ", which did not disappoint everyone, and some data still surpassed:
- CPU: Because 9000 is also A77, not A76 of 990 5G, and it is more advanced than Dimensity 1000+. Therefore, the scores of the CPU's quantitative performance data (CPU-Q) and floating point performance data (CPU-F) are higher than 1000+, but it is slightly beyond, which is reasonable;
- INT8 NPU data: NNAPI 1.1 can surpass Dimensity 1000+ by so much (24398 vs 17124), which was unexpected in previous analysis. It seems that the NPU Da Vinci architecture has also made a qualitative leap in INT8 support. However, NNAPI1.2 is not high (14280 vs 18559), which may be caused by the low support for the newly added operator of NNAPI1.2.
- FP16 NPU data: Whether it is NNAPI1.1 or NNAPI1.2 data, as in the previous analysis, it completely surpasses (NNAPI1.1 score: 37064 vs 12970; NNAPI1.2 score: 73781 vs 25879).
If there are no major changes to floating point support in the future, MTK's Dimensity and Qualcomm's Snapdragon should not be able to surpass.
Of course, with the release of mobile phone data , it is no surprise that Huawei Mate 40 Pro is also far ahead at the top of the list.
I have been following the list and specific changes between " AI Benchmark v4: Dimensity 1000+ vs Kirin 990 5G " [2020-10-6] and this article [2020-10-23]. Let’s take a look at the middle (2020) -10-13) Data retained by the author:
Here the author found:
The 990 5G "non-public or prototype-implemented HiSilicon 990 5G" data is gone. At that time, it was suspected that it was for the release of the new Kirin 9000, and the intermediate data was cancelled. It was just that there was no real hammer at the time, and it was not easy to say. Sure enough, with the release of Kirin 9000, the data in " AI Benchmark v4: Dimensity 1000+ vs Kirin 990 5G " should be prepared for Kirin 9000, or maybe it is the intermediate performance data generated by Kirin 9000.
Recall the time point:
- 2020-10-06 Published " AI Benchmark v4 Top Ranking: Dimensity 1000+ vs Kirin 990 5G " and analyzed the intermediate modified version of "990 5G";
- 2020-10-13 Concerned that the intermediate modified version of "990 5G" disappeared;
- Published this article on 2020-10-23 " AI Benchmark v4: Kirin 9000 tops the list " With the release of Kirin 9000, 9000 topped the list [Note: Kirin 9000 was released on the night of 2020-10-22, and it is also seen in China It’s the morning of the 23rd].
Anyway, congratulations to Huawei HiSilicon. Of course, this will cause the next round of an arms race in computing power. However, I hope that the performance indicators of the entire AI ecosystem will eventually be reflected in the user experience in actual applications.