According to industry media outlet seminalysis, Samsung Electronics, or more precisely, Samsung's various semiconductor divisions, is adding fuel to the fire. Over the past decade, Samsung has been on top of the world. Their share of wafer foundry has increased rapidly, and they have had the fastest multi-logic node conversion in the foundry field.
Samsung's LSI design team has launched the best mobile chips, and Apple has relied entirely on Samsung to make all the key components. In terms of DRAM, Samsung is many years ahead of other manufacturers in terms of production costs.
But now, these technological advantages are disintegrating.
The report pointed out that the cultural problems of Samsung Electronics have shaken its core. Samsung has suffered declines in every aspect of technology development, including one area where they have historically beaten all their competitors – DRAM.
The mobile SOCs they produce can no longer squeeze into the top 3, or even be surpassed by MediaTek. As for the foundry business, the two largest customers have successively switched to TSMC.
Some reports have even pointed out that there are lies and deceptive behavior in the Samsung foundry. Even Intel's fledgling new foundry business can attract one or two customers from Samsung!
To be clear, some of the details reported in the media are unconfirmed, but the overall problem is obvious.
Samsung Electronics clearly has a problem
These problems stem directly from Samsung's cultural makeup. Before diving into the details, we thought the general trend was that these issues had been bubbling up for years, but now they have reached a boiling point.
First, let's start with the mildest part of the story.
Samsung has a service called Game Optimizing Service, which restricts most apps, in addition to common benchmarks, and even non-gaming apps like Netflix and Instagram.
Admittedly, we didn't expect to be angry, as Android OEMs have a long history of cheating on benchmark apps. As a result of the above behavior, Samsung is facing a class action lawsuit.
Why did Samsung launch this "Game Optimizing Service"?
That's to control heat and power consumption, of course, because the heat and power consumption of past generations of products have always been very moving. In the final analysis, it is because the node problem has already appeared in Samsung.
Dating back a few years ago, Samsung's internal CPU design failed due to poor leadership. Samsung also had a GPU team, but eventually disbanded. These two examples illustrate why developing your own chip is much more difficult than it seems. The culture of design companies and factories is critical to success. These talented engineers need the right motivation, direction, and leadership.
After the failure of the internal GPU architecture, many fans of Samsung are excited about moving to AMD's RDNA-based GPU IP. This excitement is very short-lived. The current generation of Exynos 2200 does raise issues related to foundry and process nodes. Its performance per watt is very bad. The performance and power consumption of RDNA-based GPUs isn't the only issue with this SOC.
Samsung initially planned to roll out Exynos chips more widely around the world. Some analysts and industry publications even expect 60% of the total galaxy S22 series sales to be based on the Exynos 2200 and 40% on the Qualcomm S8G1.
In the end, this proved ineffective, and exynos ended up accounting for less than 25 percent of the actual share. While we don't think the program will be raised to 60%, we did hear that Samsung wants to increase the share of Exynos processors to 40%. The shortage of quantities is mainly due to performance and production.
Samsung's Exynos 2200 production is reportedly extremely low. This is partly due to the fact that they use 4LPE nodes. The node is innovative on Samsung's 7/5nm series, offering a 198nm UHD cell height, compared to the previous node with a 218nm UHD cell height. The yield of this node is rumored to be as low as 20%.
These look too low, but we've heard that parametric yields are terrible, even if the catalyst yields are good.
Another source told us that the chips that were eventually shipped had higher power consumption and lower performance targets, resulting in about 80 percent of parametric yields. Regardless, there are rumors of an agreement between executives to ship the latest node even if it has no economic or power/performance implications.
Here, we need to parse some nouns, the so-called catalyst yield loss is that the transistor, through hole or metal layer part does not work at all. Parametric yield refers to the loss of parametric yield when these features work, but the question is whether it meets your performance, power, voltage, etc. goals.
Because the exynos 2200's parametric yield is so low, the chip's target must be set low. In fact, there are rumors that Samsung has reduced the clock frequency of the GPU from the planned 1.69GHz to 1.49GHz and eventually to 1.29GHz.
These issues are not just due to foundries and SOC teams missing out on their technical goals. But these different units are allegedly pointing fingers at each other. The Samsung LSI (design) is blamed on the Samsung Foundry, while Samsung Mobile is blamed on the S.LSI.
Samsung LSI executives even seem to blame changes in South Korea's labor law. Instead of letting engineers do ridiculous work at critical times, limit employees to a maximum of 52 hours per week. While we've heard that this isn't fully complied with, it reduces the overwork of many Samsung engineers.
Samsung's resistance is so strong that some have even pushed for legislation to relax these labor laws.
Samsung Semiconductor's culture has become so toxic that the foundry even allegedly lies about production.
South Korean media reported that Samsung is conducting management reviews and audits in response to the above issues. The outcome of this review is likely to be a reshuffle of management and teams, similar to the restructuring of the wireless division last year. Later reports of lies went even deeper to claim that Samsung's foundries lied to customers and Samsung's chairman on 5nm, 4nm and 3nm yields. Given that there are quite a few media reports in South Korea, as well as many local experts involved, these various reports seem to be quite credible.
We can be sure that Qualcomm is angry with Samsung.
According to TechInsights, Qualcomm used a variant of Samsung's 5nm node, which is called 4LPX, rather than a denser 4LPE node like the Exynos 2200. Multiple sources have also told us that the S888 and S8G1 processors have a very poor Parametric yield, causing Qualcomm to need to push these SOCs to higher power levels to achieve certain performance goals.
Although the reported Qualcomm S8G1 does not have a lower yield than the Exynos 2200, it is not enough. As a side note, this works well for Qualcomm (and NVIDIA). We were told that the two customers had negotiated to pay for each finished chip instead of each wafer manufacturing.
Due to the problems with the S765G, S780G, S888 and S8G1, Qualcomm decided to abandon Samsung's high-end SOC entirely. Qualcomm even has a dedicated team working around the clock to prepare for the production of the S8G1+based on TSMC's N4 process nodes.
For the foreseeable future, the S8G2 and future high-end Qualcomm chips will be on the shelves of TSMC. Despite the slowdown in the smartphone market in recent months, TSMC will be able to maintain growth in the smartphone segment this year and next due to the shift in share with Qualcomm and continued content growth.
Samsung Mobile has been racing to find alternatives, while the S.LSI division is struggling with smartphone SOCs. South Korean rumors even say that Samsung has turned to evaluate the MediaTek Tianji lineup of the Galaxy A series lineup.
Dr. TM Roh, president of Samsung Mobile, has said that there is a new application processor that is only available for Galaxy phones. This is strange because most S.LSI Exynos SOCs, while externally available, are basically exclusive to Samsung Galaxy smartphones. This suggests more infighting and drama between Samsung Mobile and S.LSI.
Samsung's foundry problem is deeper. As we reported last year, their 3nm GAP node's OEM products won't even ship to external customers until 2024. 3GAE, the first full surround (GAA) node, has been silently postponed, and may even be canceled.
Yields are said to be extremely low. You would be sorely mistaken for anyone who thinks Samsung can catch up with TSMC because of TSMC's N3 problem and N2 production at the end of 2025.
Samsung made mistakes and lost their biggest foundry customer, Qualcomm, and their second largest foundry customer, Nvidia.
However, the Samsung LSI is not completely chaotic. They gained share of the 5G infrastructure market with well-designed, affordable, and efficient chips. They have also won many victories in the automotive sector in infotainment systems such as Hyundai and Volkswagen.
S.LSI seems well suited to working closely with customers, such as Google on the Tensor smartphone SOC. Despite the success, S.LSI seems to have been frustrated at every opportunity, as Samsung Foundry seems to have cost them the contract for the next-generation Cisco Silicon One network ASIC, which has already lost to Intel!
LSI has been working closely with Tesla for years on Autopilot/ADAS HW 3.0. This is a chip that Tesla co-designed with the assistance of Samsung, and both sides contributed meaningful IP to the final chip design. This chip design has shipped millions of units to Tesla's vehicles. HW 4.0 was originally scheduled for production at the end of last year, but appears to have been postponed to this year. In addition to Ambarco, Tesla is S.LSI/Foundry's only major external customer.
LSI has also recently made other mistakes, such as in the image sensor market. Their ISOCELL smartphone sensors are slow to incorporate hybrid bonding, unlike Sony, which has been shipping in bulk since 2017.
In addition, they failed to gain a share of Chinese smartphone manufacturers that chose to use Sony at the high end and on Omnivision. Stand-alone camera sensors and cameras belong to their own division, called Samsung NX, but have since been phased out, even after Samsung has invested in standalone cameras for decades.
Samsung DRAM disaster
Samsung Electronics' cash cow, Samsung DRAM, hasn't been all smooth sailing. 5 years ago, Samsung was undoubtedly better than Micron and SK Hynix in terms of density, performance and cost structure. In those days, some estimates advanced them by a year and a half.
Now, although the number of Samsungs is much larger compared to these two peers, in some of these indicators, Samsung can be said to lag behind Micron and SK Hynix. Samsung's act of being too aggressive in craft development for cultural reasons is to blame.
Briefly, as capacitor size slows down, DRAM density and cost scale slow down dramatically. The 1Xnm generation is the first sign of this dramatic slowdown, but since then, the cost per node has only scaled by around 15 percent.
The density growth is so tepid that DRAM manufacturers have turned to letters as suffixes instead of numbers like the 20nm generation. Relative to the competition in terms of cost, power consumption and performance, Samsung is far ahead of the 1Y generation.
This all changed with Generation 1Z. Samsung decided to be very aggressive in EUV adoption. It was a top-down decision, not an engineering one. These top-down decisions are very common among Samsung Electronics, and they are the result of the cultural issues we have been pointing out. For the 1Z, Samsung announced that they will adopt an EUV. This was done in the context of hype and media. Samsung is incredibly proud of this "achievement."
Samsung dominated early shipments of UV by absorbing about 50% of its EUV tools. Samsung's attempts to plug it into DRAM as well as their earlier 7nm logic attempts failed. 1Z DRAM nodes are never fully accelerated. This trend continues in the next generation of 1 Alpha nodes, further increasing the utilization of UV. The node is reportedly taking longer to develop. While Samsung claims that the 1 Alpha has been in mass production for a long time, it still hasn't increased significantly. Meanwhile, SK Hynix and Micron have been able to catch up with their Gen 1Z in terms of cost, performance and power, without using UV.
In addition, in the 1 Alpha generation, Micron only continues to push the DUV, while SK Hynix has begun to insert the EUV. As a result, Micron is already accelerating the 1 Alpha, while sales growth at SK Hynix and Samsung has been rather tepid. SemiAnalysis estimates that Micron currently has a cost advantage in TERMS in DRAM due to being able to move all of its capacity to Gen 1 Alpha and achieve optimal density and cost across its entire lineup.
Samsung and SK Hynix are still delivering the first-generation DDR5 using the non-EUV 1Y process, while Micron is leading the competition and delivering their generation 1Z process for the first-generation DDR5. In addition, Micron is quickly following up on the Gen 2 DDR5 on the Gen 1 Alpha process node. To be clear, EUV isn't the only culprit, but it's one of the biggest technical differences.
This is the worst-case scenario for Samsung. They can't boost 1Z. They have been unable to upgrade 1 Alpha, and now there are reports that they have canceled the development of the next generation of 1 Beta process nodes! Others reported that Samsung pushed directly to the 1 Gamma node. This cancellation report has a high degree of credibility because it is based on a disgruntled engineer at Samsung. He even posted a blog on the topic!
The engineer has proven to be part of Samsung's DRAM technology development team. He wrote a letter to Lee Jae-yong and CEO Kye Hyun Kyung, two heads of Samsung Electronics, describing the failures and problems. The blog has been taken down, but South Korean media have captured some rather worrying quotes from it.
"I've heard a lot of stories about the 'crisis,' but I think this moment is more dangerous than ever. In the midst of successive events, the top decision-makers seem to be unable to grasp the root cause of the problem. Samsung's DRAM technical development engineers wrote.
We encourage you to check it out to really grasp the gravity of the situation. For a 4-year tenured engineer who is passionate about his work, such unbridled criticism is a huge red flag. This is especially true because of the well-defined hierarchical approach to Korean work culture. Last month, we also saw a handful of Samsung DRAM technology developers move to SK Hynix. This exceeds the normal attrition rate.
Cultural issues have shook Samsung tremendously. While many parts of Samsung Electronics are still well-functioning execution machines, such as Samsung displays, NANDs, cars, and networks, the most important business is in trouble. These cultural issues eventually led to Samsung losing its technical and cost advantage in DRAM, lagging far behind TSMC in the leading process technology competition, losing its largest foundry customer, and losing out to Qualcomm and MediaTek in smartphone SOC design.
Source/Semiconductor Industry Insights