Exynos 9810 performance measured: Is it really “Android Light”?

Preface

After the release of S9/S9+, it received a lot of attention. Exynos9810 was also called “Android Light” by many KOLs; just Anandtech conducted a rough test on it, just for two days, it was a rare translation. article.

This review is from foreign media Anandtech, the original link: ” Exynos 9810 Hands On-Awkward First Results 」.

Speaking of it, although I used the 9810 version of S9+ in the first time, I did not do some detailed evaluation on it; I hurriedly returned it. (Because I am not satisfied)

Not much nonsense, let’s start the evaluation section directly below.

Note: This article has a second interpretation of the translator / the author is only shallow, if there is any omission, welcome to point out.

text

After we published some articles about the 9810 and M3 speculations, I promised to do a performance test on the 9810 on the S9. At the press conference, I conducted some tests on one of the demos to thoroughly understand the Exynos9810. performance.

Earlier this year, the Samsung LSI department threw an astonishing news — claiming that the single-threaded performance of the 9810 would double. Of course, this caused a lot of controversy and questioning; we also discussed the new Exynos at the beginning of this year. Structural features of M3 (Translator’s Note: This article about M3 also has a translation on the knowledge: ” Samsung Exynos M3 detailed: 6 decoding width, 50% IPC upgrade , Original link: ” The Samsung Exynos M3 – 6-wide Decode With 50%+ IPC Increase 」) I found that this is indeed something that can be done, not just marketing exaggeration. The new Samsung CPU core is actually a ‘fat core’, which is also the first to follow Apple’s footsteps in the Android camp, toward the “super strong” The characteristics of single-core performance are moving into SoC. Therefore, in the Android camp, this is a very interesting and theoretically very powerful processor.

We can also be sure that the 9810 GPU – Mali G72 MP18 only runs at 572MHz – a very conservative frequency, which is not what we expected. The previous generation of 9810, the 8895, the GPU is larger and also running At a similar frequency – 546MHz. Therefore, the 9810’s GPU performance improvement is lower than we expected, and I think that 600-700MHz should be the appropriate frequency for the 9810.

Storage delay

Through the delay test, we also determined the CPU’s cache configuration: 64KB L1D cache, twice as high as the previous generation 32KB; while M3 has 512KB L2 cache, and shared 4MB L3 cache.

Interestingly, the four A55s are not in a DynamIQ cluster with the big core, but they form a cluster separately. At the same time, Samsung does not seem to be planning to equip the A55 cluster with L2 cache, and instead it is 512KB. L3 DSU Cache. 9810’s cache latency control is very good – it is the best in the Android camp so far. We speculate that Samsung should use a new generation of interconnects and memory controllers.

Geekbench test – dominate the Android camp

In our tests, we confirmed the authenticity of those leaked scores; the Exynos9810 did achieve a huge performance boost, far beyond the 845, which is close to the A10/A11 level. Compared with the previous generation 8895, Floating point capability has increased by 114% at 2.7 GHz compared to the previous generation, surpassing Samsung’s expected return of doubling.

It is very obvious that the high single-core performance brought by ‘fat core’ makes the 9810 separate from other SoCs in the Android camp.

PCMark and Web Testing

When I finally tested the system performance, I found some suspicious performance data. I am not going to discuss the scores of each test in detail, because the graphics they finally draw are very similar.

When testing the S9+ with 9810, I found some very serious problems. The score of 9810 can’t even distinguish myself from the previous generation of 8895, let alone the 845 engineering machine we tested earlier this month. I checked The frequency, found that the big core is indeed running at the highest frequency of 2.7GHz. The only possible explanation is that the DVFS configuration and scheduler are very conservative at present, and did not give much load to the big core.

In the system, I found that Samsung is currently using a new scheduler called “eHMP”. I am not sure if this is based on EAS. (Translator’s Note: Energy-Aware Scheduling, the following schedutil is one of the signs of the EAS kernel) , but the system does use schedutil as a frequency regulator.

A Samsung spokesperson confirmed to us that since this is optimized for MWC firmware, the above tests have not been optimized. I am a bit hard to believe that they limit their performance so much, not because of the scheduler configuration. Indiscriminately set up. At the same time, I also got the news that in order to balance the performance gap with the Snapdragon 845, Samsung has imposed some performance limitations on the 9810 – that is, the score I get now may be meaningless. I hope this is just a mistake, In the official firmware of S9, this restriction can be removed, releasing the full potential of 9810.

Graphics performance

On the GPU side, the 9810 has reduced the behavior of the two GPU cores compared to the previous generation, and the slight increase in frequency does not seem to make up for its loss. Assuming that the G71 core and the G72 core have the same performance per core at the same frequency, then The 9810’s GPU is more like a slightly lower frequency on the basis of the MP20. So the performance improvement of the 9810 on the GPU should be entirely due to the improved performance of the G72, the improvement of power efficiency and the internal improvement of the SoC.

In the Manhattan 3.1 test, the 9810 only improved the poor 7% compared to the previous generation, far behind the 845 Adreno 630.

In T-Rex, the 9810 is up 18% from the previous generation. Samsung officially claims that the GPU should increase by 20%. This is the benchmark. Here, the 9810 GPU is closer to the 845 Adreno 630.

(Translator’s Note: Compared to T-Rex / Tyrannosaurus, Manhattan 3.1 graphics test pressure is greater.)

Power consumption

Due to the incompatibility of some tools, I can’t accurately measure the power consumption of the device. So I can only roughly infer the power consumption based on the current readings in the system.

On the CPU side, the power of a single core operating at 3.1 GHz is 3.1 W; the power of a dual-core 2.3 GHz is fluctuating from 3.1 W to 3.5 W, and the power of the four cores at 1.8 GHz seems to remain at this level.

In the next few days, I need some time to do SPEC tests to make some more accurate graph curves and inferences. At present, Exynos M3 has advantages and disadvantages. Obviously, M3 will consume power at high frequencies. Greatly upgraded to an unacceptable level, so usually M3 will run at a lower frequency, and the current scheduling also confirms this idea.

On the GPU side, the power consumption fluctuated between 4.5 and 5.2W when running the Manhattan 3.1 test, which was improved compared to the 8895, but it was still at a disadvantage compared with the 845.

to sum up

The results of this test have caused us a lot of doubts to be explained. Next we will further test 9810 further. And, I am eager to see today is not satisfactory. (Translator’s Note: Unsatisfactory refers to the PC Mark and other test results lag significantly behind the 845 phenomenon) The test results will not be reproduced in the officially sold models, otherwise the 9810 may become an unpopular SoC.

Full text

Expand reading:

Want to know 845?
CITREA: 骁龙 845 performance measured: qualified successor

Want to know 970?
Wednesday Technology: Huawei HiSili Kirin 970 Performance Analysis: Is the GPU still burning? NPU is really so god?

Want to learn about Exynos M3?

SoC two things: Samsung Exynos M3 Detailed: 6 decoding width, 50% IPC upgrade