A complete and comprehensive evaluation of a chip must be able to look at three levels:
The first layer to see the technical performance of products on the product, technology on the technical look, put a chip in accordance with the functional modules open to see the data to see the performance, such as using typical Geekbench measured CPU performance with GFXBench measured GPU performance, with Ann Bunny Test comprehensive performance, and further is to be able to use their own design to further interpretation of this type of typical representatives Anandtech, artillery, U Jun evaluation, other I do not think we have to read.
The second layer to see the innovation, only to reach the top of the players can talk about really invest in innovation, followed by the buttocks to run behind to be good.
The third layer to see the industry to see what the impact of the product on the industry level.
The first floor: unicorn 970 technical performance
Phone processor technically points to the overall is the number of modules: CPU, GPU, Modem, ISP, Audio, Video, the future is estimated to be an NPU. Evaluation of the technical level of these modules are used in our industry PPA (performance , Energy efficiency and area), but the area has nothing to do with the majority of consumers, this is not the key to the above several more truly comparable CPU and GPU, Modem is almost a specification, ISP is a part of the camera, Also more difficult, Audio is softer, Video is also about the same, NPU now almost only unicorn 970, let's put it behind.
CPU performance and energy efficiency
This picture we see to see that the CPU is not what the unicorn Xiaolong dispute, but the end of the explosion of Apple Andrews situation.Four chips are all 10nm process, why Android three are similar, and the apple outshine it? In fact Is also very simple, Apple uses a fully self-research architecture, while the other basic use of the ARM standard architecture. Expected to be sprayed here, said Xiaolong 835 obviously used self-developed Kryo280 architecture ah! Yes, nominally In fact?
We take a closer look at the Snapdragon 821, Snapdragon 835, and Geekbench single-core scoring for the three chips with the unicorn 970 ARM A73 (top) (see above), do you see that the Snapdragon two generations are obviously not like a mother , The opposite of the Snapdragon 830 and Kirin upside down how to look like, so we understand that no, the most reasonable guess is that Qualcomm has actually abandoned its own architecture and return to the ARM standard core. So why return? Of course, because Xiao Of course, the face is still very important, this thing directly or I am afraid to play face, it should be tried every means to retain the continued play Since the process is Kryo banner right bar. Since the process is 10nm, the nucleus is almost all ARM A73, then the snapdragon 835 and unicorn 970's CPU PPA there will be no big difference, unpredictable are the results, not the result There must be a problem with the test method, specifically not to ask, please self-examination. In fact the test of energy efficiency of the chip is actually very difficult for outsiders, because it is very difficult to extract the energy consumption data of each module , And Due to lack of knowledge of the details, often come to inaccurate conclusions.For example, the industry's most experienced hardware evaluation of the most Anandtech in the test CPU energy that CPUBurn program has a problem, did not consider the CPU instruction efficiency , So draw conclusions ARM A73 a lot worse than A72 energy efficiency (https://www.anandtech.com/show/11088/hisilicon-kirin-960-performance-and-power/4), if it is true that ARM do CPU people must collectively Jiang Xie, in fact, they are still alive.So the evaluation is a rigorous work, to use the brain to think about the results measured unreasonable, unreasonable to take a good look at their own method, or simply Speak out this issue, do not easily conclude.
So we give the unicorn 970 CPU to make a small conclusion, the unicorn 970 CPU and Xiaolong 835, Orion 8895 represents the highest level of Android camp, but with a huge gap between the A11 with the apple here spend a very short period of time about why Apple CPU performance so Strong, in fact, is also simple, Apple is equal to the A73, the two large nuclear into its nuclear, single-core performance of natural differences, but in fact the price is also very large, in fact, a CPU core of the Apple A11 is almost Top Kirin 970 two large nuclear area, with two add up so many transistors naturally better performance, but the power is not necessarily, so crazy in the performance of single-core upgrade, Apple also began to configure the small core.
GPU performance and energy efficiency
All along, the GPU is actually the users have been particularly concerned about, because playing games thing.In fact, more interesting is the heavy 3D game population in fact the proportion of the entire user base has been very low, that is, more than a year I do not know how The king put everyone engaged in pesticide high, even if the ratio is not high, but because this is probably the user experience directly mapped to the chip's dominant manifestation, so the various chip manufacturers are still very seriously. Here we simply to the phone GPU sent several major categories:
Apple is Imagination's PowerVR series Qualcomm is the self-developed Adreno series Samsung and Huawei ARM Mali series industry basically can see these three.First Imagination, considered the originator of mobile GPU, but the basic is Apple's maintenance, Apple dumped now. In fact, the business is not good with the apple is not good, almost to drink poison, because Apple control too strong, and gradually allow these manufacturers to limit the ability to do what you do on the line. As a result, it has become more difficult to actually support GPUs used by other vendors and we can see that MTK's performance in the past two years is exactly what the performance of the GPU in Imagination is. So after the apple was fully fed, The future upgrade will be more significant.
In the past two years, Qualcomm's Adreno can be described as prestigious, objectively speaking, Qualcomm's GPU is indeed good enough to rely on it to save Qualcomm through the Snapdragon 810, 821, two less successful products difficult period. Ask me where I am, for this series of good that may only be a framework, Qualcomm Adreno better adapt to the mobile computing demand for graphics operations.Is also estimated that Apple is also relying on Imagination is a game, determined to himself Dry it.
Take a look at Mali, ARM This company developed by the CPU, wanted to expand the business, the periphery also do it. In fact, the first two years of Mali 400 generation is a very good performance, a bit small flash look , But unfortunately did not go on, being a part of Qualcomm's Adreno thrown away.Samsung and Huawei in fact no choice, only on this boat, then what is the gap between the swollen? In the semiconductor area has a word called area for power Consumption, then put more nuclear Bai, like I beat you one to one, how many people called each, everyone would not have to spend a lot of effort, but also things to do, the only cost is the chip cost becomes high , But Apple, Samsung, Huawei are themselves for their own use, so their own commitment, consumers and no loss .We look at the performance of this generation more, half a catty, has been pulled flat.
However, although performance is close, it represents the same experience? In fact, this time efficiency is reflected in this issue, and energy efficiency is the same thing how much electricity consumption, in the industry will be accurate to Joule / frame this way .But the outside world No way to test, and Qualcomm's GPU at this point the advantage is obvious, and the more significant load more obvious.Effective means less heat, less heat means that the chip reaches the thermal protection for longer, triggering thermal protection The longer the time means lower frequency, down means Caton.In fact, in order to ensure the stability of the game frame rate, all mobile phone manufacturers and game manufacturers in the strategy will do some considerations, such as limiting the frame rate 30fps-40fps , This time the GPU workload will not be high, fever is relatively small, able to maintain, but when you open the 60fps continuous running will obviously feel fever, and then there will be occasional Caton, so for heavy 3D games and The words of the performance is Xiaolong 835> unicorn 970> Orion 8895.
Just mentioned, energy efficiency is more dominant in the case of heavy load, it is the biggest difference when running the benchmark, running heavy 3D games second, and run the largest proportion of low-load travel time, this situation The gap between the next will be significantly reduced.I always wanted to say that we do the chip design chip, who will be running for a long time full of design, are running at 70% or lower performance point and design, Highest performance point has always been designed for short-time peak, measure that point how much significance? In addition it is actually very difficult to test the calibration, because only the battery-side current data, but this data contains the whole power consumption , To exclude other components accurately is not possible .If you are really interested can also refer to the U-Jun game test, though not so accurate, but still can see. A while ago a monarch unicorn 970 evaluation, the conclusion is unicorn 970 energy efficiency compared to the Snapdragon 835 a great gap. Some people like this is like being stabbed to the same point G that excited ah, all groups of good pass, I see Just feel Funny, what do you want to explain? Well done bad are foreigners thing, circle HiSport Spirits people can do whatever you can but also the owner of it, you plug in the product so what's good excited.
In fact, the GPU is not finished, the above is talking about graphics capabilities, in fact, GPU parallel data features can do more things, each GPU addition to graphics, have designed the support for Computing, embodied in the OpenCL-based applications On the scene, in fact, many of the interface effects are achieved using OpenCL (industry standard) or RenderScript (Google), while subsequent GPU-based neural network operations are also based on OpenCL. Because this part of the performance is not so dominant The layman is not very concerned, but this is an integral part of GPU capabilities, in fact Ann Bunny and Geekbench have tested, to Geekbench, for example, we look at the level of unicorn 970 performance. You can see the performance level of GPU Computing Is: Orion 8895 ≈ unicorn 970> Snapdragon 835. Thus, Qualcomm's Adreno GPU is biased towards drawing, while Mali considers Computing a little more.
Write this to summarize the unicorn 970 GPU, the theoretical performance of the blueprint for the Snapdragon 835 and Orion 8895, but the energy efficiency gap is still obvious.Review the unicorn 950, 960 to the present 970 situation, we can see that HiSilicon to improve GPU performance and The determination of energy efficiency is still very large (below)
The second floor: Kirin 970's technological innovation
Kirin 970 generation on the AI innovation I am afraid we all see, this time and Apple stand side by side in the forefront of the industry.In fact, the innovation is not the AI itself, but added a NPU into. Do not underestimate the hardware, chip It is prudent to add an independent new module in the design, because first the chip is really expensive, and each additional area is the cost. If the added cost is two US dollars ($ 0.2), according to the shipment 50 million to calculate that is the profit of 10 million dollars ah. The second new hardware needs to be integrated into a real system, the entire software architecture of the system is a lot of work, which is the cost. So chip companies will be So cautious in this regard, only to be sure to be widely used to join. So why Apple and Huawei did so Qualcomm did not? In fact, very simple, because Qualcomm is the chip, and how confident that their customers can use AI to do, in case it can not afford this cost is not a plus.And Apple and Huawei are for their own use, of course, plus Canadian NPU can not be used, of course, can be easier under Kinds of decisions innovation. So the future, we wait and see, for the same reason, Apple, Huawei and Samsung BIG3 this will certainly be more aggressive in innovation, Qualcomm will be relatively conservative.
We come back to see Kirin 970 in the AI field of innovation what is the concept, what extent, what is the value.
Now the basic method of AI field is not something new, there are two or three decades ago, the problem has been computational performance constraints. You think, ah, an algorithm you want to run a few days there P use. The problem first of all Learning side is the rapid development of NVIDIA spent a lot of time to alleviate a lot of time, Google did TPU2 out also greatly reduced the threshold of learning side.So good, it will certainly pour a lot of AI algorithms and models come out, In fact, the AI inference performance of mobile phones has become the biggest bottleneck in the popularity of AI applications, so Apple and Huawei added NPU ah, I have some Do not understand such a clear trend how Qualcomm will be so conservative, but you are the boss of the year ah, now how to be conservative in this way.
Then Kirin 970 AI performance compared to Apple's A11 is what? According to its press conference released data, the following reference to the most valuable, because it shows in which networks are compared, the overall view is significantly higher than However, according to Apple's information, currently its Neural Engine is only used for built-in Face ID applications, and is not open to developers, Core ML can only access to GPU-based neural network acceleration.Therefore, to judge for AI applications Of developers, regardless of energy efficiency or performance is certainly better Kirin 970. From this point of view Haisi is still very Niubi.
Some people say that although there is no hardware unit Snapdragon 835, but there are DSP or GPU, can not be considered without AI capabilities. I would like to say is that this is a typical conservative, that without any cost can barely support very good. With this DSP or GPU certainly run dry hardware logic, ah, not to mention these are shared resources, an application when you run, can be assigned to AI how much?
Well, to sum up the level of innovation, it is clear that unicorn 970 this step forward a stride, became the fact that the phone AI chip leader, this throw thrown Xiaolong 835 and Orion 8895 several streets.
The third floor: Kirin 970 contribution to the industry
In fact, we can have a word for each generation of Kirin chip evaluation:
Kirin 950: The first one among the ranks of the top mobile phone SoC processor unicorn chip Kirin 960: bucket core Kirin 970: the industry's first real sense of the phone AI processor from the above few words you can feel the Kirin 970 industry Meaning, I'm further for everyone from several dimensions:
1, in the absence of a clear idea of Android, Haisi boldly made this innovation, but also for the development of a set of development kits, which is a very thought-provoking action, but also the first in the field of Android dare to cross This first step player
2, to help the Cambrian from a laboratory results into a product of things, the entire Chinese AI chip area brought a large number of experienced personnel and achievements
3, to unlock a group of long-term application developers who are unable to fall due to the limitation of the AI performance of the mobile phone, it is expected to bring a wave of AI applications to the users, and this wave of developers and experiences will first appear in China
4, substantial acceleration of the entire cell phone chip circle on the AI judgment, more investment will enter, more business will be tapped out
These can be said from our practitioners point of view can have very great significance, and Haisuisuu is the Whampoa Military Academy of China semiconductor industry, dear rice noodles, you have high hopes for the pine cones are not out of Hass ?
If you let me write a comment for Kirin 970, I would write: Kirin 970 is a brave attempt of Huawei, try to excel, try to take this leading step.Industry chip industry will remember it, AI industry will remember it More importantly, it is not done by others, but it is made by ordinary people around me, which deeply inspired us!
Appendix 1: Kirin 970 overall industry level comparison (1-5 points)
Appendix 2: The main reference data sources
Geekbench: https://browser.geekbench.com/v4/cpu/singlecore
GFXBench: https://gfxbench.com/result.jsp
Original turn knew almost columnist: whoknows