Although Qualcomm once crushed its opponents with its advanced self-developed architecture, Qualcomm is now increasingly showing signs of fatigue. I personally believe that Qualcomm has been making mistakes since the release of 805.
This time I will write a long article to talk about what Qualcomm did wrong.
It has only been three months since the 801 was released. Qualcomm was obviously not satisfied with the 2.45g high frequency of the 801. It further improved the architecture and increased the frequency, and launched the 805, a 32-bit SoC with a frequency of up to 2.7g. The krait450 architecture is used. At the time, all parameters of this SOC were impeccable. In order to maximize performance, Qualcomm even equipped it with a 64-bit dual-channel luxury memory controller specification. This is also the first time Qualcomm has launched a SoC that supports 4k display output. This SoC was carried by the National Bank of China Samsung Note 4 and other models, and represented the strongest mobile SoC performance at the time. Except for the exynos5433, which could compete with it, all other SoCs were completely crushed by it. However, no one could have imagined that this would be the last time Qualcomm would achieve such success.
After just releasing the 32-bit 805 in the spring, Qualcomm has finally entered the 64-bit era. The release of Android 5.0 that supports 64-bit has brought great convenience to the performance of 64-bit SOC. Qualcomm, which tasted the benefits of 64-bit on the 410, released several 64-bit SOCs in the autumn, including the msm8994, which is the legendary Hell Flame Dragon 810.
According to Qualcomm’s tradition, the 8 Series should be supported by self-developed architecture. However, because it coincides with the transition from 32-bit to 64-bit, Krait is a pure 32-bit architecture, and it is impossible to force it to 64-bit by changing it. In addition, the time left for Qualcomm is too hasty, so Qualcomm had to force the public version to replace it.
If you just use the public version, then the problem will not be too big. Samsung is using it. At that time, Kirin also appeared and Kirin is also using the public version. If you do it honestly, nothing will happen. The bad thing is that Qualcomm has not made a public version flagship for a long time, and it has always lacked a rational understanding of the fever of the public version architecture (Qualcomm uses the public version on low-end SOCs, but because the low-end SOCs themselves do not generate much heat, Qualcomm simply does not Didn't realize the public version could be so hot). As a flagship, performance must be high. In order to ensure sufficient performance advantages compared to the 805, Qualcomm chose to violently increase the performance to 4xA57 4xA53.
However, the A57 is the hottest among the public versions in the past. In addition, Qualcomm has not used the public version as a flagship for a long time, so it is not sure about the maximum frequency setting of the public version. In addition, the 801 and 805 A big victory for high frequency, Qualcomm has set a high frequency of 2.0g for A57. The power consumption of A57 will rise sharply after the frequency exceeds 1.4g. The high frequency of 2.0g is destined to not be cool.
If the high-frequency A57 is due to lack of understanding, Qualcomm also did a smarter thing: chose the 20nm process. The exynos7420 of the A57 and A53 is not too bad with the help of 14nm. The 20nm used by Qualcomm this time has further promoted the birth of Hell Flame Dragon.
The old manufacturing process makes it difficult to see the energy efficiency ratio of the 810 even at low frequencies. Kotori has an shv32 on hand, which is officially locked to four large cores, but the heat generated by the four A53s is enough to be scary.
Although the actual performance of the 810 is not too weak, especially the GPU performance is excellent enough, unfortunately, due to the overturn of power consumption, all mobile phone manufacturers that use the 810, even those related to Qualcomm, have been historically criticized. rewritten. The top version of Xiaomi Note only has benchmark scores and no experience, ruining Xiaomi’s chance to hit the high-end market. Nexus6p becomes the last model of the Nexus series. Oppo had to postpone its new Find series products for nearly four years. Sharp also gave up the three-sided full-screen design on the new model. OnePlus had to barely maintain the experience by downclocking to 1.8g. Samsung has no choice but to use the 7420 globally.
Although Qualcomm tried to correct the 810, in the v2.1 version, it was still the Hell Flame Dragon. This Soc with luxurious parameters and crash experience finally got the code name of epic fail and the name of the fire dragon. , is also widely recognized among the people.
After the failure in trying to repair the 810, Qualcomm immediately devoted all its energy to the research and development of a new architecture. This new architecture is kryo.
Qualcomm still took advantage of its self-research capabilities as always, but Qualcomm made a very irrational decision at this time: to compete with x86.
Qualcomm began to conspire with Juhard to develop win10arm. Juhard was obviously not satisfied with the toothpaste factory’s behavior of squeezing out toothpaste, so it began to develop win10arm.
Unlike mobile platforms that value integer computing capabilities, desktop platforms have higher requirements for floating point computing capabilities. In order to meet the demand for high floating point, kryo has made great efforts in floating point performance and directly achieves out-of-sequence five launches. Exactly one year after the release of the 810, the 820 came out equipped with four Kryo self-developed msm8996.
Although floating point is not the most important thing for mobile devices, the powerful floating point performance allows kryo’s single-core performance to crush many public versions, and the 820 has set a record of four-core crushing eight-core processors. Nuclear's glorious record. Because the 810 is too hot, the 14nm 820 was widely praised when it was released and is widely regarded as a model of high performance and low heat generation.
However, 820 is not a completely successful SoC. Quad-core itself is not suitable for clustering, but Qualcomm made it 2 2. Since eas scheduling had not yet been born at that time, hmp often showed unreasonable scheduling on dual-same architecture clusters, causing the cores to often run at unnecessary high frequencies. . The bigger problem is that Qualcomm has cut the small core cache, and coupled with the immature branch prediction, the small core often predicts failure and explodes the cache.
These problems were not solved until the later launch of 821, which is the 820 official super version (Qinlili’s current main machine is 821). Later, Google couldn’t bear it and wrote eas scheduling by hand, and finally it was certain This greatly alleviates the efficiency issues of 82x.
The design error does not mean that a good card is ruined. What is even more exciting is that Qualcomm mentioned above that it wants to enter the desktop, right? Juhard has lived up to Qualcomm’s expectations and quickly launched win10 arm. and tested successfully on 820. We can take a look at the video from that year/watch?v=A_GlGglbu1U
(Youtube video, how to open it and figure it out yourself)
Thanks to the powerful floating point performance, the 820 in the video has It can barely run PS, and the system applications are smooth enough for daily use.
It would be completely appropriate to launch the 820 win10 device at this time, because the design of the 820 takes into account the computing needs of the desktop. However, Qualcomm did not do this and wasted such a good opportunity. It waited until later to launch the 835 win10 device. However, the floating point performance of the 835 was not enough to meet the needs of the desktop. As a result, the experience was not as good as the n3450, and the toothpaste factory was forced to After sitting on the toothpaste in a hurry, Qualcomm's crisis became even greater.
At this time, Qualcomm actually has a way to fight the crisis. Qualcomm has built-in DSP (similar to NPU) since 600, but Qualcomm charges high fees for development using DSP, resulting in only a small number of manufacturers taking advantage of DSP. If the restrictions on DSP are relaxed at this time, Qualcomm can definitely lead the AI ??trend ahead of Kirin, but after all, Qualcomm has been a patent troll for a long time and does not have this awareness.
The 82x missed the opportunity to enter the desktop market, and did not become popular on the mobile side due to design errors. At this time, Qualcomm decided to return to the public version. But after all, Qualcomm still has some backbone. It is not like the 810 that is forced to not directly release the public version on the flagship, so Qualcomm can only choose to modify the public version.
The design of the 835 is actually very conservative. Not only is the single-core performance actively reversed compared to the 821, but the GPU improvement compared to the 821 is also extremely limited, not even 15.
However, because of the use of the public version, 835 was released a full quarter earlier than the third quarter release that Qualcomm has always insisted on. Due to the higher efficiency of the contemporary public version and the first batch of 10nm tests, the 835's heat control is exceptionally good, and sales have soared. But at this time, Qualcomm has not given up the idea of ????all self-research at all.
At the same time, it seems that 820 running win10arm has given Qualcomm such confidence that Qualcomm even wants to get a share of the toothpaste factory in the server industry.
Towards the end of 2017, Centriq appeared. Using Falkor's self-developed architecture, this architecture uses out-of-sequence five launches like its predecessor kryo, but its computing power and efficiency are greatly improved, and it still maintains A75-level energy efficiency ratio despite the explosion of floating-point performance.
But Centriq failed to capture any market share at all. Qualcomm's ancestral agricultural enterprise has been tossing arm servers in the bulldozer era, and finally gave up. Competition in the server market is much more difficult than Qualcomm thought. Because most traditional server applications are compiled using x86 and cannot run on ARM, there are very few applications for ARM servers. Coupled with the pincer attack from the toothpaste factory Zhiqiangfa and the agricultural enterprise EPYC, Centriq became a failed product and there will never be a next generation. Dileep Bhandarkar, the technical president responsible for self-research on its architecture, also resigned in July this year.
The server market has been considered a failure, but Qualcomm has never given up the idea of ??sharing the toothpaste factory. As mentioned above, Juhard has prepared win10 arm. Qualcomm, which was frustrated by the server, turned to the notebook market.
Although the 835 simulates x86 efficiency loss is relatively large, but because the toothpaste factory has been so dominant in notebooks for many years that it squeezes out toothpaste, it stands to reason that Qualcomm can rely on the excellent energy consumption ratio of the 835 to dominate the market segment. It has gained a foothold in the market, but its entry into the notebook industry has not been as smooth as Qualcomm thought.
At the beginning of the year, the agricultural enterprise released the new Ryzen architecture, and the toothpaste factory could no longer sit firmly on the throne.
The 835 notebook has not yet been officially launched. The toothpaste factory, which feels the murderous attitude of agricultural enterprises, has brought the most conscientious upgrade in history, fully entering the quad-core era. The performance of the 835 instantly pales in front of the 8th generation low-voltage toothpaste factory. Powerless. As a result, Qualcomm can only regard long battery life as the only highlight of the 835 notebook, because apart from battery life, Qualcomm really can't come up with any strategies to deal with the enemy.
As 2017 draws to a close, the mobile Ryzen comes out with the most powerful core display in history, bringing huge changes to the notebook industry. Because the price is cheap and the performance is even slightly better than the 8th generation low voltage from the toothpaste factory, it became very popular in the low-end notebook market. This move by agricultural enterprises made the toothpaste factory anxious. They sat on the toothpaste and brought the descendants of the atom series - J4005/J5005. The performance of these two ultra-low voltage entry-level Us is greatly improved compared to the predecessor atom, J5005 Even a true quad-core. The N3450 was replaced, and before that, the only one that slightly lost to the 835 in some projects was the N3450 (/news-story/qualcomm-snapdragon-835-vs-celeron-n3450-benchmarks-with-windows-10.html). Qualcomm's last hope of entering the notebook world was dashed.
After 835 broke Qualcomm’s traditional New Year-style iteration rhythm, Qualcomm’s rhythm became increasingly chaotic. In the first quarter of this year, Qualcomm hurriedly released the 845. The tight schedule meant that Qualcomm had no time to improve the defects that could have been solved. The 845 had a serious cache leakage problem.
If cache leakage is not a big problem, the bigger problem is that Qualcomm’s dominance is no longer guaranteed.
The performance improvement of the 845 is still limited, but on the other side of the ocean, in a country where Pipa is said to have originated from it, a SoC called the "Light of Android" was born. With six launchers and an independent L3 of up to 4M, the 9810 ran an astonishing single-core score of 3,500 points. In the same-frequency floating-point efficiency test, the 9810 ran nearly double the efficiency of 845.
Before this, Qualcomm’s status had never been shaken so violently. Qualcomm had never suffered such a humiliation in terms of running scores. You know, the epic fail that year was not in the running points. Make no concessions.
Qualcomm’s rhythm was further disrupted.
As usual, Qualcomm will launch official super versions of a SoC 2-4 quarters after its release, such as 800 and 801, 820 and 821. However, this time Qualcomm only launched the 845 in less than a quarter. The official super version 850 was rushed out. Qualcomm, which was forced to panic, pushed the core frequency to 2.96g in one breath. However, no matter how high the frequency is, the status cannot be guaranteed.
Qualcomm’s position has been attacked again and again. Samsung successfully developed its own baseband and said goodbye to Qualcomm in CDMA. The uplink of cat18 even killed the 850. The toothpaste factory continues to advance in the baseband, directly integrating a complete x86 core to compete with Qualcomm, and grabbing orders for Apple's new iPhone baseband. From Yu Dazui's mouth, 7nm and cat21 came out.
The 845 is still the SoC with the best energy consumption ratio, but the 9810 is getting closer and closer.
Excluding the inefficient public version of the 9810 GPU, the CPU efficiency of the 9810 is even higher than that of the 845.
Of course Qualcomm is in a hurry, there is no reason why it shouldn’t be. As the saying goes, a dog jumps over the wall in a hurry, and Qualcomm, which was forced into a hurry, became more and more chaotic, and even tried to threaten the toothpaste factory. However, the toothpaste factory had no time to deal with such petty troubles.
At the same time, Qualcomm is also struggling to avoid being acquired by the closed-source rogue Broadcom.
The 5G era is just around the corner, and major operators are also accelerating the elimination of 2G and 3G. However, most of Qualcomm’s patents are limited to 2G and 3G. It is becoming increasingly difficult to make a living by selling basebands, and more and more people are leaving the technical department...
Qualcomm, whose rhythm was disrupted, began to launch a large number of castrated versions of Soc market segments . According to convention, Qualcomm should release the next generation flagship in Q4 this year. However, Qualcomm’s official media does not even have any warm-up publicity for the new flagship. The 675 is quite energetic in a small way. Samsung on the other side of the ocean has officially announced that the 9820 will be a major improvement compared to the 9810. Exynos has grown from a baby who only used the public version to a very aggressive adult.
Missing opportunities step by step, failing to turn over again and again. Although the 835 and 845 have indeed achieved good sales, Qualcomm's advantage is losing little by little. Losing independent technology is tantamount to putting your life in the hands of others, and Qualcomm is experiencing this trend.
8150 has become Qualcomm’s only hope for a comeback, and this remaining hope is still passing by little by little.
Written on 2018/11/22, simultaneously published on Zhihu//Bilibili