When will our smartphones get faster? We are looking for an answer.
Advertising campaign Vivo, focusing on the speed of the smartphone
It often happens that on the box with a new smartphone there is a promise that is not fulfilled in practice. The biggest disappointment is the lack of understanding of what the manufacturer means by talking about a 'fast' smartphone. What does a 'fast' smartphone mean? For a simple user, this is a lightning-fast reaction of a smartphone to touching the screen, whether it is launching an application or scrolling through a site. But the manufacturer, it seems, relies on the results of synthetic tests and its own conscience. A mismatch between expectations and reality often occurs when a new chipset is introduced to the market, which, in fact, turns out to be not so good. Bad solutions from Qualcomm, Samsung and HiSilicon come to mind. Fighting for supremacy in the big competition for the buyer's wallet, hardware manufacturers stick their spurs deep into the sides of developers, and chipsets are born, about which you want to say 'made from under the stick' or 'hastily'. The loudest such case that immediately comes to mind happened with the Snapdragon 810 chipset, which worked great for the first few minutes, and then went into overheating and deep denial of the need for work. In the case of top-end Samsung chipsets, complaints were formed not only by those who lacked performance in complex games, but also by specialists surprised by the artificial limitation of displaying images on the screen in games and complex applications to 30 frames per second. Until recently, HiSilicon chipsets mysteriously refused to run the simplest applications.
Guess where the Snapdragon 810 sits. Photo tweakers.net
All smartphone chipset manufacturers purchase a license from ARM, after which they have all the available documentation, informational support, and the right to manufacture and sell. The next stage is the design of the chipset, and then the platform. The manufacturer selects the number of cores and the technological process for their manufacture, draws and manufactures a stencil, which is later embodied in a prototype. It should be understood that the ARM company does not manufacture pilot batches, in many respects their work is purely theoretical. Let's remember one of the first incarnations of the ARM big. LITTLE – chipset Qualcomm Snapdragon 615 (MSM8939) which was very odd. Two clusters of four Cortex A-53 cores had to quickly switch among themselves depending on the load, solving energy saving issues without losing performance, but something went wrong, and all smartphones based on this chipset warmed up and drained the battery literally before our eyes. For 2014, it was a completely new technology, not tested and thoroughly studied by the developers themselves. By the way, each new generation of Cortex is distinguished not so much by a transition to a new level of performance, but by additional equipment integrated into the chipset. In the case of the A-53, LTE and Wi-Fi 802.11ac became such integrated modules, and this was a rather radical design change. Everything points to the fact that the theoretical calculations of power consumption, made by the engineers of ARM, in real conditions were greatly underestimated. The outdated and 'hot' 28 nm process technology should be referred to real conditions.
At the next stage of development, we observed a continuation of the trend. So, the Snapdragon 650 chipset introduced faster Wi-Fi and LTE modules (with support for two SIM cards), a signal processor and a new video chip, and the RAM controller began to support RAM LPDDR3 with a frequency of 933 MHz versus 800 MHz in the 615th. With all this, the technical process remained the same – 28 nm, and the chipset warmed up, despite the removal of two cores from the design, of which there were six against eight. The technical description of the chipset – the Cortex type – became a marketing gimmick, with the help of which a smartphone seller could prove the superiority of a new product over an old one, but in practice it meant nothing. A comparison of the performance of Cortex A-15 (Krait ™ 400) in Snapdragon 800 and Cortex A-72 in Snapdragon 650 is quite indicative. Despite a significant gap in the version number (72 versus 15), in the tests for one core the 650 was weaker by 3%. The technology has not radically changed, and the objective acceleration of smartphones was caused by the appearance and availability of faster RAM, new flash drives and the transfer of additional functions to video chips. At the same time, all control over the operation of the chipset, overclocking and parking of cores, the choice of the power supply mode and going to sleep was carried out at the software level. The manufacturer called such software with a beautiful name and demonstrated it as an achievement, one example is CorePilot from MediaTek. But when managing hardware with software, there is always the possibility of failure and incorrect operation. And they advertised it like this:
Last fall, the HiSilicon Kirin 980 chipset was announced, which is used in Honor View 20, Huawei Mate 20 and 20 Pro. Earlier, it became a habit to understand that the increase in pure computing performance when switching to a more delicate technical process was about 10-15%, despite the promises of chip makers about the superiority of new products by 30-40%. This was the case when going from 28 nm to 16 and from 16 to 10 nm. This growth was fully explained by the increase in memory bandwidth (transition from DDR3 to DDR4) and the increase in the processor clock frequency. But this time the growth was real 29% compared to the closest pursuer – Snapdragon 845. And there can be only one explanation for this – the use of the new Cortex A-76 architecture and, perhaps, the inability of modern tests to put old and new chipsets in the same conditions. A distinctive feature of the new Cortex A-76 architecture is the use of DynamIQ technology.
Typical marketing picture, in the case of Cortex A-76, which turned out to be almost true
DynamIQ is a hardware-based power management technology for computing cores and a kind of hardware task manager. Among the distinctive features (according to ARM) are the slowdown of the computational core to minimum values without a complete shutdown and instant start when an external API is called. ARM created a new architecture very flexible, and, apparently, those who like to delve into the specifications will find a new benchmark for evaluating chipsets from different manufacturers – the size of L3 Cache for ARM, which in A-76 can vary from 512 KB to 4 MB. It seems that the new 855 chipset from Qualcomm will also be a variant of the Cortex A-76 with numerous modifications. Although the manufacturer does not recognize what exactly is hidden under the designation Kryo 485 in the SDM855, the very scheme of computing cores 1 + 3 + 4 would be impossible using the Cortex A-75 architecture.
The evolution of ARM big.LITTLE, expressed in the general access to the processor memory of all cores at once. Avoiding the RAM cycle accelerated cluster switching
And at the same time, the changes that new premium chipsets bring to our lives do not affect the well-being of the vast majority of people. These chipsets are too expensive and very few people can afford, and the entry of new technologies into the middle and budget price segments takes years. They do not affect the mass market for software for the same reasons. The performance of the good old 4-core Cortex A-53 chipset will be enough for most tasks, if the smartphone has a normal amount of RAM and the read / write speed of the internal drive is not the lowest.
Instead of an afterword
Everything that has happened since the advent of the Cortex architecture can be called optimization and quantitative improvement. One core was replaced by ten, and the requirements were imposed on the power of the video processor based on the screen resolution, which had already reached its logical limit. It makes no sense to complicate the displayed image with new small elements, the user will not see them. At the same time, the problem of parallel computing has not yet been resolved, a lot is written and talked about at conferences, which means that the performance of one core is still the most important parameter. All last year's stories about AI (AI) turned out to be the same optimization of calculations, and sometimes they meant a part of the chipset that can be programmed to solve specific problems. The scaling of ARM cores in the Cortex A-76 has reached the top, reaching odd circuits and using shared processor memory by all cores. By itself, the further development of the ARM architecture cannot make a truly qualitative leap in performance in our smartphones, and, what is most unpleasant, it is not very clear why this leap should be made.
What for?!
Friends, what do you think can stimulate the further development of mobile chipsets?