Last week, ESMC sister media wrote a technical article about the performance of Samsung's flagship mobile phone chip Exynos 2200 this year, which is a tragedy, and its performance and efficiency are at the bottom of this year's mobile phone flagship chip, successfully relieving the word-of-mouth dilemma encountered by Qualcomm Snapdragon 8 Gen 1. The chip actually provides a glimpse into the problems of at least three of Samsung Electronics' business units.
The first is the Samsung Foundry, the Samsung Foundry factory that makes the main body of the Exynos 2200 chip; the Samsung 4nm process has shown poor electrical characteristics on both the Qualcomm Snapdragon 8 Gen 1 and Samsung's own Exynos 2200 — and it's already well known. The second is Samsung LSI, the chip design team: that is, the same IP scheme, almost the manufacturing process, Qualcomm is slightly better at the specific design implementation of the chip; not to mention comparing with the MediaTek Tianji chip next door. Finally, there is Samsung Mobile, the team that made Samsung phones: the phone with the Exynos 2200 chip (Galaxy S22), which also pulls the overall performance in terms of system design.
In fact, the crotch pull performance of exynos chips is not only on this generation of products, but this year's flagship chip of Samsung mobile phones is particularly "eye-catching". This may just be a microcosm of the current problems that Samsung Electronics has.
Generally speaking, Samsung's flagship mobile phones over the years will use a dual-chip strategy, that is, some mobile phones use their own chips, and the other part uses Qualcomm chips - no matter why Samsung adopts this strategy, for this year's flagship mobile phone Galaxy S22, Samsung originally planned to use its own Exynos 2200, 40% with Qualcomm Snapdragon 8 Gen 1. The actual situation is quite different from the plan, and the Galaxy S22 phones with Exynos 2200 chips seem to account for less than 1/4 of the total. The Exynos 2200 has also become the least visible presence in any Exynos chip in its history.
Recently, the research institute SemiAnalysis wrote an article pointing from the general direction to the "culture issue" within Samsung Electronics enterprises, although it is not clearly pointed out, but it can basically be summarized by bureaucracy and serious internal friction, in order to say that Samsung Electronics may be currently encountering big trouble. However, it should be noted that this article is a collection of gossip reported by the Korean media, and some information has not been confirmed. This article attempts to render only a subset of them. This article will not focus on malicious speculation about Samsung Electronics' corporate culture, but only give some information, and add some thoughts and comments from International Electronic Business.
Internal contradictions are spreading
Regarding the situation that there are various mutual blames between different business agencies and departments of Samsung Electronics, it belongs to the drama that can be staged in the eight o'clock file of the series, and this article will not be elaborated more, and this article of SemiAnalysis has introduced it a lot. While these are certainly specific symptoms of the bureaucracy of the company, such as rumors that Samsung Mobile is very dissatisfied with the Samsung LSI department; Samsung SLI accuses Samsung Foundry of not working; Samsung LSI is also dissatisfied with the change of Korean labor law; engineers work for a long time; and are not honest with customers.
In such a cultural environment, it is said that Samsung Foundry has lost qualcomm and Nvidia's big orders. Although this matter is also expected, after all, in the continuous 7nm, 5nm, 4nm process shows significant weakness compared to TSMC, and it is very likely to further widen the gap with TSMC in the 3nm process, Qualcomm and NVIDIA, as a large customer of Samsung Foundry, should have ideas. But the story seems to be more complicated than expected.
The failure of Samsung's 4nm process is well known. South Korean media reported poor yields on the Exynos 2200, based on Samsung's 4LPE process. SemiAnalysis said that the catalyst yield yield of the 4LPE process is OK (catastrophic yield means that the transistor, via, or metal layer is not working properly), but the parametric yield (which means that although it works properly, the voltage, power consumption, and performance do not meet the target indicators) is very bad. So the chips that were finally shipped achieved lower performance at higher power consumption, and the parametric yield yield was about 80%. It is rumored that Samsung executives are just thinking about selling new processes, without considering economics, energy consumption, efficiency and other issues.
Based on this, the Exynos 2200 is actually unable to achieve the target performance and power consumption. South Korean media rumors say that Samsung had to adjust the core frequency of the GPU from the initial 1.69GHz to 1.49GHz, and even the final 1.29GHz. The credibility of this rumor seems to be relatively low, as analysts in the International Electronic Business Intelligence know about the Exynos 2200 GPU architecture (RDNA 2) and more mobile phone GPUs, reaching a frequency of 1.29GHz is not a reality (a large core, high-frequency design like the Adreno 730 does not reach 1GHz). However, this rumor may be expressing that the 4LPE process cannot achieve the original expectations of the electrical characteristics.
So the first big customer that Samsung chips may have to be cut is Samsung Mobile itself. Samsung Mobile is said to be already looking for alternatives to the phone chip. South Korean media claim that Samsung's Galaxy A-series mobile phones may be ready to switch to MediaTek Tianji chips. If this matter is true, it shows that the contradictions within Samsung Electronics are not small.
External customers are churning
Samsung Foundry has two well-known big customers: Qualcomm and NVIDIA. SemiAnalysis said that Qualcomm has been quite popular with Samsung since the last two generations of Snapdragon processors have adopted Samsung's manufacturing process. This year's Snapdragon 8 Gen 1 underperformed and was even overtaken by MediaTek, in which Samsung Foundry is clearly partly responsible. TechInsights announced that Qualcomm does not actually use the 4LPE process, but a 5nm improved process called 4LPX, which will be inferior to 4LPE in density. The Exynos 2200 uses 4LPE.
According to the news released by The TechInsights advanced technology Hello Yuzo Fukuzaki, the 4LPX process "hardly found 4nm feature" used by Qualcomm Snapdragon 8 Gen 1 is different from Samsung's real 4nm process , that is, the 4LPE process. This is a very interesting topic, and if TechInsights discloses more information in the future, I can further expand on this topic in the future.
SemiAnalysis said in the article that multiple sources indicate that the parametric yield of qualcomm Snapdragon 888 and Snapdragon 8 Gen 1 chips is very bad, causing Qualcomm to have to push up the power consumption level to achieve the target performance level. And it is said that the yield of the Exynos 2200 is worse, which is probably the reason why Qualcomm did not choose to use the 4LPE process... There are even rumors that Qualcomm and Nvidia are negotiating to pay according to the good product die, rather than in wafer.
A few months ago, the mobile phone chip master had news that TSMC had already invested in the N4 process of Snapdragon 8 Gen 1 improved chip. And the supply chain news generally mentioned that in the future, Qualcomm's flagship and high-end chips will turn orders to TSMC. It seems that the decline of the smartphone market will probably not affect TSMC in the past two years, after all, it has eaten a lot of predetermined quantities for a while.
As for Nvidia, the new generation of GPUs for the data center version of the Hopper architecture that has just been released will not unexpectedly use the TSMC 4N process (NVIDIA claims that this is a customized version of TSMC's 4nm process); and what is even more unfortunate is that the Nvidia generation of game GPUs (Ada Lovelace) for the consumer market leak specifications show that this batch of game graphics cards will be based on TSMC 4N technology.
Although Qualcomm and Nvidia always prefer the idea of multi-supplier in choosing the foundry OEM strategy, the transfer of flagship chips to TSMC is a routine operation, which does not mean that Samsung will be abandoned. But such an operation at least shows that Samsung's semiconductor cutting-edge manufacturing process is almost no longer the first choice of chip design manufacturers.
3nm problem is big, DRAM business is in trouble...
The most serious allegation against Samsung in South Korean media reports is that it is not very honest with customers about the yield of cutting-edge manufacturing processes. Coupled with various gossip news, Samsung Electronics does not seem to be on a very healthy track at present, especially its semiconductor business. At present, Samsung's first GAA structure transistor process 3GAE yield does not seem to be satisfactory, so in Samsung's public roadmap in the past two years, the 3GAE process has gradually disappeared. Instead, the 3GAP process is scheduled for mass production in 2023, and SemiAnalysis believes that 3GAP may be postponed.
Then the uncertainty is even greater, especially at a time when the competition between the three cutting-edge foundry factories is becoming more and more fierce.
On the other hand, SemiAnalysis also devotes a chapter to depicting the current difficulties encountered by Samsung DRAM. Roughly speaking, 5 years ago, Samsung was undoubtedly in a leading position in this field compared with Micron and SK Hynix. But now in some of the more comparative dimensions, Samsung is beginning to show fatigue. The root cause is also a "cultural problem".
In the 1Z generation of technology, Samsung has aggressively chosen EUV lithography. SemiAnalysis said that this was not a decision from an engineering point of view, but was influenced by "top-down", which was caused by cultural problems within Samsung Electronics, which was quite leapy and eye-catching. The 1Z DRAM node has not reached the full scale increase target. During the development of Samsung's subsequent 1 Alpha, SK Hynix and Micron took the opportunity to catch up in terms of cost, performance and power consumption of the 1Z node.
At the same time, Samsung has not been able to achieve 1Z, 1 Alpha can truly achieve the complete expansion of scale effect. Micron is already shipping the 1Z node of the original DDR5, but the second generation of the 1 Alpha process node is progressing smoothly. Not long ago, the South Korean media also reported a blog article that is said to be a Samsung DRAM development engineer, mentioning the current problems in DRAM technology.
However, Samsung Electronics is involved in a wide range of business. These are just some of the problems that may exist at the moment, and there are many gossip in the Korean media. In fact, Samsung's various businesses still have new blossoms in many fields. For example, Samsung LSI has performed well in the 5G infrastructure market, and has made new gains in the automotive field in recent years, and there are also new customers like Google in the mobile phone field (although tensor SoC performance is also a bit difficult to say); Tesla, as another big customer, Samsung has also cooperated with it to develop self-driving chips.
At the same time, Samsung Display, NAND, Networking and other business development is also good. But in the few cutting-edge manufacturing markets that have persisted so far, Samsung Foundry really needs to work harder. Analysts at International Electronics Business believe that the above crux of the matter is basically on Samsung Foundry, and there are some situations that are not listed here, such as its image sensor market is also undergoing encounters.
In the advanced manufacturing process market, Intel's promotion of IDM 2.0 plans is actually a threat to Samsung Foundry; Cisco Silicon One network ASIC orders are said to have been lost to Intel. Market analysis always says that these two years are particularly critical for Intel; but in fact, it may be more critical for Samsung, and the progress and maturity of its 3nm process to be observed are opportunities to understand Samsung's current development. Culture and the like aside, the technical problems that emerge from the outside need to be solved at least.
Editor-in-Charge: Elaine