With the rise of competitive mobile games, smartphones have put forward more demanding requirements for 3D graphic rendering/computing capabilities. At the same time, 4K ultra-high-definition recording, video editing on the mobile phone, AR and VR and other entertainment application experiences are also indispensable to the participation of GPUs (as well as UNITS and ISPs). In other words, today, when CPU performance is basically sufficient, "nuclear war" from the GPU level is inevitable. So, what are the GPUs that deserve our attention in the field of smartphones?
Let's start with the cooperation between Samsung and AMD
In early June 2019, a blockbuster news broke out in the mobile phone circle - AMD and Samsung jointly announced a multi-year strategic partnership, Samsung will be authorized by AMD Radeon graphics API and focus on enhancing mobile devices (only in market segments where AMD does not compete, such as smartphones and tablets) to innovate and innovate advanced graphics technologies and solutions that are crucial. There are not many details officially revealed, and it can only be confirmed that AMD will license the "highly scalable RDNA graphics architecture" to Samsung for the time being.
So, what is RDNA?
In simple terms, RDNA (Radeon DNA) is AMD's latest GPU architecture in the graphics field, it is the successor of the longest-lived GPU architecture "GCN" (2011-2019) in history, which has been innovated in CU compute units, caches, pipelines, etc., with better performance (to meet the load requirements of modern games), energy efficiency (fully optimize power consumption and bandwidth utilization), functions (to strengthen the relevant ecosystem) and scalability (eat-all mobile, desktop and cloud). Arguably, RDNA is a fully redesigned architecture that ushers in a new era of the fifth major architecture in amD graphics history.
So, will Samsung's future Exynos mobile platform integrate RDNA architecture GPUs directly?
The answer, of course, is no. AMD has long lacked experience in the ARM ecosystem (AMD has been involved in ARM servers, but soon abandoned), and the core requirements of mobile phone SoC integrated GPU are low power consumption and high energy efficiency, AMD's existing GPU technology obviously does not meet the basic requirements of mobile devices.
In fact, Samsung has been carrying out self-research (independent research and development) of CPUs and GPUs in recent years, and has already had a successful "Mongoose" (Mongoose, such as Exynos M3/M4) core based on ARM instruction set in the field of CPU architecture, and has made a lot of reputation through mobile platforms such as Exynos 9810 and Exynos 9820.
In the GPU field, it is reported that Samsung has carried out its own "S-GPU" project as early as 2012, and this cooperation between Samsung and AMD only requires the integration of part of amD Radeon's graphic IP into the GPU, and it is impossible to copy the entire RDNA architecture. At the same time, the cooperation between the two parties also involves the relevant patent licensing to avoid possible legal disputes in the future, after all, MediaTek is a good lesson for the past.
At the 2015 MWC show, it was reported that MediaTek had reached a cooperation with AMD in the field of mobile SoC graphics computing, and this news was not officially confirmed. In early 2019, AMD took MediaTek to court, saying that MediaTek's smart devices infringed multiple patents related to its APUs and GPUs. You know, MediaTek's SoCs integrate GPUs under ARM or Imagination.
With the injection of AMD graphics IP, Samsung can devote more energy to the polishing of the self-developed CPU core "mongoose", and AMD's popularity in the PC field can also help Samsung compete with Qualcomm, Huawei and MediaTek to differentiate . I'm integrating an AMD GPU! At the same time, when the Samsung Exynos CPU is paired with the GPU that integrates the AMD RDNA gene, it may also have an impact on an industry, which is the game handheld represented by the Nintendo Switch. Can Samsung +AMD replace the exclusive power of NVIDIA Tegra X1 and its successor chips in the handheld field? Let's wait and see.
Qualcomm and AMD have to tell the story
The Adreno GPU integrated by Qualcomm Snapdragon mobile platform should be the strongest voice in the field of Android mobile phones, and it is also the only existence that can compete with the integrated GPU of Apple's A series chips in the same period. But did you ever know that as Qualcomm's unique "self-developed" GPU, Adreno was originally the "bloodline" of AMD?
The predecessor of the Adreno was actually the ATI Imageon series of low-power GPUs born in 2002, which was originally used as a Zodiac handheld for pre-installed PalmOS systems. After ATI was acquired by AMD, AMD packaged the related mobile device assets and dumped them to Qualcomm in 2008, so the latter acquired AMD(ATI)'s vector drawing and 3D drawing technology and related intellectual property rights, and developed the Adreno GPU that we are familiar with.
If AMD had not sold the original ATI Imageon that year, and Intel had not sold XScale (with the full license of StrongARM and ARM architecture) and handheld device chip business that year, the two X86 chip giants may have achieved greater success in the era of mobile Internet.
How Adreno GPUs judge strength or weakness
In the field of Android mobile phones, the Qualcomm Snapdragon mobile platform occupies most of the country, and Adreno is also the GPU we are familiar with and least familiar with. Say it is familiar because every time Qualcomm releases a new Snapdragon mobile platform, the media will always emphasize that it integrates the Adreno xxx model GPU, how much it has improved over the previous generation, and the ears are grinding out of the cocoon; saying that it is not familiar is because the Adreno GPU model is too many and too messy, so that it is difficult for us to judge the relative performance level between the same generation of Adreno GPUs from the name.
This was particularly evident during the Adreno 5 Series GPU era (Table 1).
For example, the Adreno 506 (Snapdragon 625) is only more than the Adreno 505 (Snapdragon 430) +1 in terms of model numbers, but the performance of the former is almost 2 times higher than that of the latter. The Adreno 508 (Snapdragon 630) is better than the Adreno 506+2, but its performance is only 30% better, and it is completely irregular.
In order to understand the above problems, we need to understand the key parameters that affect the performance of Adreno GPUs. Leaving aside the core architecture, rendering methods, and various graphics interfaces, the performance of Adreno GPUs is mainly limited by the process process, the number of ALU (computational units, also known as arithmetic logic units), and the frequency of GPUs.
We can understand the Adreno GPU-integrated ALUs compute unit as the "stream processor" in the PC graphics card, and the greater the number of architectures, the stronger the performance. Unfortunately, Qualcomm has never disclosed the number of ALUs of Adreno GPU, and the relevant data in this article and on the Internet are detected by users themselves, so they are only for your reference.
For GPUs, the more advanced the process process, the higher the frequency at the same heat generation and power consumption levels, and the higher the gain effect on performance. The Adreno 506 has a huge performance improvement over the Adreno 505, thanks to the increase in the number of ALUs and the frequency increase brought about by process upgrades. Similarly, Adreno 512 is influenced by the above factors compared to Adreno 508 and Adreno 530 compared to Adreno 512.
Today, Adreno GPUs have fully evolved into the Adreno 6 Series era, and their biggest feature is support for full graphics interfaces such as OpenCL 2.0FP, OpenGL ES3.2, DX12, and Vulkan1.1.
Unfortunately, the author has not been able to find accurate information on the number and frequency of their specific ALUs, so in this article, we will only compare the theoretical performance of the Adreno 6 series GPU through the two professional test software of 3DMark and GFXBench (Table 2).
We use the Snapdragon 660 integrated Adreno 512 as a benchmark (100%), which gives a more intuitive understanding of the differences between the Adreno 6 Series GPUs. Among them, Adreno 610 and Adreno 612, although the number sequence is higher, the overall performance is actually almost the same as the Adreno 512. The slight increase in performance between Adreno 615 and Adreno 618 is behind the difference between the different COMBINATIONs of ALUs units and GPU frequencies.
As Qualcomm's flagship in 2018, the reason why the Snapdragon 845 (Adreno 630) has the 3D performance of Qualcomm's latest launch of the Snapdragon 730 (Adreno 618) is because the Adreno 630 integrates up to 512 ALUs units, and the Snapdragon 730's ALUs should be less than 256 (sorry no specific parameters were found).
It should be noted that from the second half of 2018, Qualcomm's latest Snapdragon SoC mobile platform has strengthened the AI computing capabilities, and they will use some of the ALUs units in the GPU, combined with the HVX units integrated in Hexagon, together with the CPU to form Qualcomm's unique Snapdragon AI Engine artificial intelligence platform.
In other words, in the latest Snapdragon mobile platform, more ALUs units can not be exchanged for equivalent performance gains. Taking the Snapdragon 855 as an example, Qualcomm has revealed that this chip (compared to the Snapdragon 845) has increased the arithmetic logic unit by 50%, that is, the ALUs unit has increased from 512 to 768, but from the actual performance difference between the Adreno 640 and the Adreno 630, the increased 50% of the ALUs has only been exchanged for about 20% of the performance gain, and obviously many of them have been used by Qualcomm for AI operations. According to testing and analysis by some professionals, half (256) of the 512 ALUs of the Adreno 630 are used for heterogeneous AI operations to compensate for the fact that the Snapdragon mobile platform does not have independent NPU units.
In short, the "most capable" in the Snapdragon mobile platform is nothing more than the Snapdragon 845 and Snapdragon 855, and there is also a big gap between the GPU of the Snapdragon 7 Series and the Snapdragon 6 Series. Therefore, if you want to buy a mobile phone that can run the vast majority of games at full frame, it is still necessary to look for the Snapdragon 8 Series. As for the Snapdragon 6 series, when the Snapdragon 710 has dropped to the price of 1,000 yuan, unless you don't play the game much, you should stay away.
Apple and Imagination's feud
In the field of smart phones, the only chip designers who can build their own GPUs are Qualcomm and Apple, Qualcomm's Adreno GPU is derived from AMD, and the GPU of Apple devices is inseparable from the support of Imagination Technologies.
Imagination was also one of the early entrants to the GPU market, just to avoid the edge of Nvidia and ATI, the company focused more on the polishing of ultra-low-power mobile GPUs, the PowerVR MBX GPU introduced in 2001 was integrated on Intel, Freescale, Texas Instruments and Samsung's SoCs, and made the first generation of iPhones, Nokia N95 and Dell Axim X50v (PDA Handheld) classic devices.
Imagination's fifth-generation PowerVR SGX series should be the most well-known GPU (such as SGX530, SGX535), Apple A4, Samsung Exynos and Deyi OMAP3 series SoC are its customers, sony PSV handheld integrated is also a quad-core PowerVR SGX544MP4. Subsequently, PowerVR gradually became Apple's "exclusive" GPU, and other brands (such as MediaTek, Quanzhi, Rockchip and Jingchen, etc.) even if they occasionally cooperated, the integrated PowerVR GPUs were mostly low-end models, and the treatment was far from the "gold lord" Apple. However, it is precisely because of the too much dependence on Apple that it has laid the groundwork for the decline of Imagination.
The Apple A10 processor should be regarded as the last cooperation between Apple and Imagination, and its integrated PowerVR GT7600 GPU can even lead the Adreno 540 (Snapdragon 835) in the same period, making the iPhone 7 series the most powerful smartphone of the year (in fact, each generation of iPhones is basically the strongest in the same period).
Unfortunately, starting from the A11 chip, Apple officially abandoned Imagination and switched to a self-developed GPU, the chip's GPU is configured with 6 shader cores, each 2 cores share a Text unit, similar to the design and PowerVR GPU, so many players have said that Apple GPU is stepping on Imagination - after the news of Apple's self-developed GPU, Imagination quickly fell into an operational crisis. Today, only a very small number of SoCs such as MediaTek Helio P60 (PowerVR GM9446), P35 and P22 (PowerVR GE8320) have chosen to integrate PowerVR GPUs in cooperation with Imagination, and the only hegemon of the iOS ecosystem has fallen to "other" in the GPU field.
Imagination's latest GPU solutions include PowerVR 9XEP/9XMP/9XTP (positioning is raised in turn), 9XEP and 9XMP using the older Rogue architecture, and the main competitor is ARM's Mali-G72. The 9XTP is built by the latest Furian architecture, each ALU increased to 40 pipelines, supporting 4K 120FPS, HDR, Vulkan API, Android neural computing API and other advanced technologies, which can theoretically be used to compete with ARM Mali-G76/G77.
ARM's "pro-son" Mali GPU
Next, it is ARM's turn to "kiss the son", the Mali series GPU. As a partner with the Cortex A series CPU architecture, it is also the existence of the Android mobile phone circle market share second only to the Existence of Qualcomm Adreno GPU, HiSilicon Kirin, Samsung Exynos, MediaTek Helio, most of the non-Qualcomm SoCs we are familiar with have directly chosen to marry with Mali GPU.
CFan has written in "The Biggest Regret of Kirin 990! What's so good about the ARM Cortex-A77 architecture? The article interprets the features of ARM's latest Cortex-A77 CPU and Mali-G77 GPU, so this article will not be repeated, but simply introduce the special features of Mali GPU.
Compared with Adreno GPUs, Mali pays more attention to the number of GPU cores, and packages more GPU cores into a larger GPU matrix, which is the only way for Mali GPU to pull up performance.
The Mali-G76 can lay out up to 20 compute cores
For example, in the Mali-G72 era, the Kirin 970 integrated 12 cores of Mali-G72MP12, and the Samsung Exynos 9810 was even more powerful, and a brain was stuffed into 18 cores (Mali-G72MP18). However, the more GPU cores, the heat generation and power consumption will also become larger, and the choice of SoC manufacturers at this time is usually to reduce the GPU frequency and turn off some cores to achieve the purpose of controlling power consumption.
In other words, there is a critical value for the number of Mali-GPU cores, beyond which the GPU frequency must be reduced, which ultimately leads to an increase in the number of GPU cores and the corresponding performance increase disproportionately (1+1<2).
Therefore, how to allocate the limited chip area to the Cortex-A CPU and Mali GPU, and reasonably control the size and frequency of the Mali GPU, and play the best energy efficiency ratio, this ARM is handed over to the SoC vendors for the final exam.
In short, Qualcomm Adreno and ARM Mali are the main forces in the current Android mobile phone field, and the specifications and performance of Adreno GPUs depend on the model of the Snapdragon mobile platform, which is relatively easy to compare. ARM Mali GPUs involve Mali models (such as Mali-G52, Mali-G72, Mali-G76), the number of cores (suffix MPx), and the CPU architecture and main frequency they go with, whether it is comprehensive, CPU or GPU performance requires us to spend more energy comparing.
When Samsung brings amD authorized IP self-developed GPUs, Imagination gets more SoC chip manufacturers support, think about it, right? So, if you want to stay up to date with the latest mobile GPUs and recommendations for buying, stay tuned for the latest CFan coverage from the same period.