Samsung’s chip division wants the top spot back: What’s happening?

“We will retake the world’s No.1 spot in the next two to three years,” said Kyung Kye-hyun, the CEO and head of Samsung’s Device Solution (DS) Division, the tech giant’s semiconductor division, in his speech during the company’s annual shareholder’s meeting this month.

In the Q&A session that followed, Kyung was forced to fend off barbed questions from shareholders, the majority of them aimed at the chip division, with responses, in so many words, that boiled down to “we are going to do better, we will do better.” The time allotted to such a session was the longest in recent memory, showing Samsung’s commitment and anxiety over the future of its chip division. Executives of Samsung’s other half, the Device Experience (DX) Division, its consumer electronics division responsible for smartphones and TVs, barely received any questions by comparison.

For the past decade, it has been rare for any Samsung business unit to state otherwise than to reaffirm their already top position, as a leader in smartphones, consumer electronics, and memory chips. When it did happen, as with Intel and Apple, where Samsung would concede the top spot and become ‘No.2’ based on revenue from market figures, it would stay reticent.

So what’s happening? And why two to three years instead of putting a pin on it?

It was only in 2017 when Samsung took over the top spot for the first time, buoyed by unprecedented demand for memory chips, from Intel, and there has been a back-and-forth between them since then. While Intel wasn’t considered a direct competitor at these times and facing its struggles in the mobile space, it has long been considered the premier semiconductor company for decades, some would argue it still is. So, it was more of a symbolic win for Samsung, a culmination of its decades-long struggle to become a semiconductor giant. A long road from the 80s when it was one of many memory suppliers that were dwarfed by large Japanese conglomerates.

Unfortunately, this isn’t the case anymore.

‘No. 3′

According to analyst firm Omdia, Samsung was ranked third in terms of revenue in the global semiconductor market in 2023, ceding its top spot of the prior year. Intel took the top spot with $51 billion, having been second in 2022. Samsung recorded $44 billion. Trailblazing Nvidia took second place to surpass Samsung with $49 billion with a year-on-year growth of 133%, having been eighth place in 2022.

As noted by Omdia, the semiconductor industry saw revenue down 9% from $597 billion to $544 billion, “highlighting the cyclical nature” of the market, the downturn notably affecting the major memory makers.

The semiconductor market encompasses a diverse array of chips, which can be roughly categorized between memory and logic (which takes a considerably larger portion of the market), digital and analog. Companies are also divided into fabless and those with fabs, the former being companies that don’t have production facilities that only design chips and the latter that can manufacture them. Samsung and Intel are what the industry calls integrated device manufacturers, or IDM, which do both. Companies that only manufacture like TSMC (which was absent in Omdia’s report but if included would have pushed Samsung’s ranking further down) are called pure-play foundries, also known as contract chip manufacturers.

Because of this, unlike end products such as smartphones, these companies are not necessarily just rivals but are customers and partners with each other. However, there is intensifying competition in segments of the market which is affecting Samsung, as an IDM, in all directions.

Accordingly, the Samsung DS Division is divided into three business units: Memory Business, the division’s crown jewel and arguably the entire conglomerate’s, Samsung Foundry, which manufactures chips, and Samsung System LSI, which handles logic design. How these three business units perform in their respective segments as well as collaborate will determine whether Samsung can put the ignominious “No. 3” behind it.

Memory upcycle starting

Memory, comprised of the products DRAM and NAND, is the most important for Samsung, the world’s largest memory maker since 1993. While the company always only announces the earnings of the DS Division, the Memory Business is its most important unit accounting for the lion’s share of revenue and the overwhelming share in profit for the division. The semiconductor market is cyclical, but this has been especially true for memory chips as they are more of a commodity than their logic counterparts.

Hence Samsung CEO Kyung’s “two to three years” as the memory upcycle that started late last year is yet to go full swing and will be the key determining factor on whether Samsung will meet its stated goal. The CEO said that the revenue of the DS Division this year will return to 2022 levels. He also said the semiconductor industry as a whole was projected to earn a revenue of $630 billion, even larger than in 2022. 

While the memory market has always been cyclical, from 2017 to 2022, it seemed to have a dose of chaos theory from Covid, which extended the upcycle longer than usual with some analysts calling it a “bubble”, fueled by high demand from data centers, consumer electronics, and stay-at-home trends at the time.

If it was a bubble, it burst in 2023, causing memory makers to lower production output for the first time in 20 years (at least when it comes to when they were confirmed to do so), a drastic measure for fabs that run 24/7 that have eye-watering high-fixed costs. For over a decade, the global memory market has been an effective triopoly of Samsung, SK Hynix, and Micron (in NAND alone there is more competition, however). Because of this, all three companies have been careful in readjusting their production output, no doubt in the hope of steadily returning the memory prices to pre-2023 levels.

Because the stakes are high, while Samsung, SK Hynix, and Micron rarely directly mentioned each other in the past, there have been subtle jabs among each other during the downturn.

Last year in March, SK Hynix boss at the time Park Jung-ho said during the company’s annual shareholder’s meeting that the memory chip sector was stuck in a “prisoner’s dilemma”. At the time, Samsung had yet to reduce its production output, unlike the others. Park’s comment meant that the rational thing to do was for all three companies to lower the supply of DRAM to the market together so that the downcycle could start and end faster but one company __ Samsung __ was acting differently.

Before this, when the memory downcycle began in late 2022, Samsung also broke away with tradition in December of that year by directly referring to its latest DDR5 DRAM as 12-nanometer (nm). Companies have grouped their DRAMs as 10nm class before this, so Samsung’s move was designed to highlight that its DRAM was better than its rivals more so than before.

Samsung would argue that it didn’t start this battle of words and numbers. Before the downturn, SK Hynix and Micron started marketing the number of layers in their NAND. How many cells are stacked in NAND chips is important, like the nm in DRAM that reflects the gate width, but it doesn’t tell the full story and is mostly a marketing term today. But its rivals’ marketing forced Samsung to reiterate during multiple conference calls in 2022 that “layers aren’t important” and eventually announce the layers of its own latest NAND. Sources within Samsung told ZDNet that the company found the entire situation “irksome and annoying”.

Nvidia and HBM

For those outside of South Korea, Samsung’s rivalry with SK Hynix may be unfamiliar compared to its famous one with LG. But the long-held perception of SK Hynix in South Korea has been similar to LG, a perennial No.2 to Samsung in memory chips.

Enter HBM, or high bandwidth memory, which according to research firm TrendForce, SK Hynix is currently the exclusive supplier of HBM3E to Nvidia for their GPUs aimed at AI applications, with over 80% market share. Samsung is now going through qualifying tests for its HBM3E with Nvidia while Micron seems to have gotten the OK from the GPU maker for theirs.

HBM is a chip comprised of multiple DRAM dies stacked vertically connected through microscopic wires called through-silicon-vias (TSV). The high bandwidth provided by its namesake caused a surge in demand for them in 2023 despite the downturn from the AI storm.

Stacking as a technology instead of scaling to push the performances of various chips over their limits started earnest in the 2010s due to the increased difficulty of just scaling chips. What is unfortunate for Samsung is that before the 2020s, it was the leader in stacking and TSV. It led the way in 3D NAND against its rivals. The South Korean tech giant was the first to introduce HBM2 to the market in 2015, though the demand for them wasn’t there at the time.

The memory market is highly conservative and the mass dissipation of new form factors is usually a generational affair. The market performance of Intel’s Optane memory is a good example. Most of this comes down to economic feasibility for the memory makers as well as their customers and less with technological breakthroughs. New fabs or lines are not made unless there are sure customers. That is why the performance of SK Hynix’s HBM3 which launched in mid-2022 and was specifically designed for Nvidia’s H100 has been a surprise of sorts. This proved that HBM is economically viable, hence Samsung and Micron’s sudden readiness to spend billions in expanding their production capacity for the chips. Last month, Samsung introduced its 12-stack HBM3E, clearly aiming to distinguish it from SK Hynix and Micron’s 8-stack counterparts, somewhat ironically.

SK Hynix seems to be enjoying its victory lap. During its annual shareholder’s meeting this month, the South Korean memory maker said it expected HBM to take up double-digit portions of its overall DRAM revenue. The company also stressed that HBM must be “customized and specialized” for customers, which is true, but it was essentially highlighting its relationship with Nvidia.

But Samsung has also announced bullish plans for its HBM production output with one executive saying it plans to near triple its production output of the chips this year. Nvidia CEO Jensen Huang confirmed during the unveiling of Blackwell this month that it was qualifying Samsung’s HBM, causing share prices of the South Korean tech giant to jump 5% the next day on the Korean bourse.

While Samsung will heavily rely on the overall upcycle of the memory market, gaining a level playing field or surpassing SK Hynix when it comes to HBM in the coming years will no doubt determine whether it can become the top vendor in semiconductors within three years.

Besides, SK Hynix also has a huge problem in NAND. Curiously absent during its annual shareholders’ meeting also held this month were discussions over its acquisition of Intel’s NAND and SSD business units for $9 billion, the first phase of two having been completed in 2021. A year later, NAND prices began plummeting and they are currently at less than half the price of their peak. In other words, in hindsight, it was the worst possible timing. Sources familiar with the matter said these newly acquired business units were in a state of capital erosion and may end up costing SK Hynix more to save them than what it paid for.

“We will develop 1c DRAM, ninth-generation V-NAND, HBM4, and other advanced chips with the utmost competence to lead the industry again,” Samsung CEO Kyung declared.

Foundry: TSMC dominates, Intel joins fray again

Samsung Foundry is facing a more uphill battle. Last year, the foundry industry in general was hard hit by the economic slowdown as demand for smartphones, consumer electronics, and cars dropped. This meant orders for processors, microcontroller units, image sensors, and other chips, mostly made with legacy nodes, used in them also dropped. Returning its fab operation rates to two years prior will be Samsung Foundry’s priority this year.

In terms of advanced nodes, while Samsung Foundry was the first to introduce a 3nm process node to the market in 2022, South Korean analysts believe that the business unit has been in the red for consecutive quarters, being unable to gain large customers for the node due to alleged low yield rate.

But to explain away the foundry, or contract chip manufacturing, market in terms of just process nodes __ which are important __ is a gross oversimplification.

TSMC, or Taiwan Semiconductor Manufacturing Company, the world’s largest foundry, seems indomitable at the current time. According to TrendForce, as of the fourth quarter of last year, the company held a 61.2% market share in foundry, with Samsung trailing behind with 11.3%. GlobalFoundries and UMC each took third and fourth places with shares of 5.8% and 5.4%. SMIC was fifth with 5.2%.

The success of TSMC, founded in 1987 which counts Apple, Nvidia, and Qualcomm as its customers, has many layers but the primary reason is probably the fostering of its ecosystem. Contract chip production has over the years developed into an entire value chain of its own. Major customers’ chip design became more complicated while the number of chip startups that lacked expertise also increased, which increased the need for efficiency and closer collaboration with the customers on TSMC’s part.

The chip giant was the first to ally with design house partners, known as chipless companies which act as matchmakers between TSMC and its fabless customers that offer design work for the customer. These alliances are divided into tiers, with Value Chain Aggregators (VCA) TSMC’s being the closest partners that not only design chips but conduct sales activity on its behalf, providing incentives for these subcontractors. TSMC also provides the rights over its IPs to VCA and lower-tier partners as customers want to use existing and proven chip IPs to increase manufacturing efficiency.

Global Unichip Corporation (GUC) is its most famous VCA, founded in 1998 and acquired by TSMC in 2003. Known for its advanced SoC design, GUC can handle initial engineering for customers or the entire design-production process in turn-key for customers.

Samsung launched its similar alliance called Samsung Advanced Foundry Ecosystem (SAFE) in 2018 and has begun extensively sharing its IPs with its design house partners. However, this ecosystem is still in its infancy compared to TSMC’s and in terms of the number of IPs is drastically behind.

Another part of the value chain that the Taiwanese giant is arguably ahead of is packaging __ where the chip die is encased in a container that allows it to connect with the motherboards of devices. Packaging, along with assembling and testing, is what is called a back-end process in chip production. The front-end is where a silicon wafer goes through various steps from lithography to chemical processes to become chip dies on the wafer.

When chip scaling was the primary goal in the past, the frond-end of chip production was considered the more important part. However, because the semiconductor industry is reaching the limits of scaling, advanced packaging has risen in importance as an affordable way to boost the performance of chips. Foundries before TSMC also mostly outsourced their packaging work, but the Taiwanese chip giant was the first to offer it to customers to increase their convenience.

The logic, memory, and other parts are packaged in “2.5D” __ in effect, bundling more chips closely together on an interposer to increase data transfer speed between each other to prevent bottlenecks. As a pure-play foundry that consistently deals with major customers such as Apple and Nvidia, TSMC is a leader in these 2.5D packages and 3D package technologies that are being developed. Samsung has begun playing catch up by forming its advanced packaging team, which CEO Kyung said is expected to contribute $100 million in revenue this year. This segment is only expected to grow as AI inference requires more and more bandwidth, so packaging chips closely is a must.

A decade ago, Samsung had a chance to narrow its gap with TSMC more aggressively. In the 2010s, Apple gave the majority of the production work for its A series of processors to Samsung, which was already a supplier of other components such as displays and memory to the iPhone maker. The South Korean tech giant’s foundry business had two major customers at this point that could ride high on the fast-expanding smartphone market: Apple and its own logic business. This changed when the well-known Apple and Samsung patent lawsuits began in 2011.

Gradually, the South Korean tech giant lost orders from Apple to TSMC, but it was able to surpass then-second-place GlobalFoundries thanks to the manufacture of its own Exynos processors used in Galaxy smartphones. Samsung claimed that there was a “great wall” between its chip and smartphone business units to assuage Apple but to no avail.

In response to customer concerns, Samsung Foundry split off from Samsung System LSI in 2017. But Samsung Foundry’s fortunes didn’t improve much since then, as Samsung System LSI’s Exynos processors, which are manufactured by Samsung Foundry, failed to become a consistent rival to Qualcomm’s Snapdragon processors, which Samsung Foundry and TSMC manufacture. “We don’t compete with our customers,” is an often repeated refrain from TSMC, something Samsung can’t say in its current structure.

Last month, Intel announced that it had $15 billion worth of foundry orders and Microsoft as its latest customer. This is an incredible development in its own right, as the US chip giant only reentered the foundry market in 2021 after Covid, security, and the subsequent supply chain issues caused governments across the world to adopt policies to increase domestic chip production. The company also confirmed that it would be a beneficiary of the US CHIPS Act. Intel intends to become the world’s second-largest foundry by 2030. Notice how it didn’t say the largest foundry.

Realistically, Samsung Foundry’s goal will be to stave off competition from Intel while maintaining its runner-up position. CEO Kyung seemed keenly aware of this as he said narrowing its gap with the market leader __ TSMC __ was one of the challenges it had to solve while making the business “sustainable”, or in other words, fostering its own ecosystem. It is also expanding convenience for customers through partnerships with Arm in Cortex-X.

The good news is that the global foundry market is projected to double in size in the next decade. According to Allied Market Research, it was valued at $106.9 billion in 2022 and is expected to reach $231.5 billion by 2032. There just might be enough for everybody. Samsung Foundry and Intel’s competition is also good news for customers. More vendors mean more competitive prices and options. As long as the supply is stable, the primary concern of customers.

On Intel’s mention of a 1.4nm process node in development, Siyoung Choi, the president and general manager of Samsung Foundry, also at the meeting, said the node was also planned by TSMC and Samsung, underplaying its significance. Choi also added that compared to Intel, the South Korean tech giant had a foundry field record of developing and supplying a more diverse range of products such as CPUs, mobile application processors, SoCs, and GPUs. Samsung Foundry, like TSMC, also offers more production capacity in legacy nodes for less sophisticated chips.

“The operation rate [of our fabs] is improving compared to last year and we expect a meaningful recovery in the second half of the year,” the general manager said. “We are preparing to mass produce 2nm process nodes next year and are conducting specific discussions centered around US customers.”

Logic: Beyond smartphones

Logic chip is a term that encompasses a wide variety of chips from the most advanced CPU to the simplest controller unit. Likewise, Samsung System LSI also offers many chips for different applications.

Key among them are the image sensor, which powers the camera, and the display driver IC, or DDI, which is used in display panels. Samsung CEO Kyung made a point during the meeting to say it earns 1.5 trillion won in revenue annually from DDI, as it is the market leader, as Samsung Display uses them for OLED panels. Samsung is also the second-largest image sensor maker in the world after Sony, which provides them for iPhones. Samsung also has its well-known SoC brand, Exynos.

All three products, in essence, grew on the back of the smartphone market. More specifically, Samsung’s own Galaxy brand. For Samsung System LSI, then, the goal is to maintain its competence in the smartphone market, which has slowed down in growth but is not going anywhere, while finding a new growth-generating product category.

AI, such as artificial generative intelligence (AGI), could be one of them. Samsung announced its own AI accelerator called Mach 1 during the meeting. The company claimed that the chip can be used with existing memory while reducing the bottleneck with the GPU by one-eighth of the current level. Systems that use the chip will launch next year. However, it is important to note that other major tech companies are also working on or have their own AI chips: AMD, Intel, Amazon, IBM, and Google among them.

Kyung also said something interesting during the meeting. “In SoC, we will make it a meaningful business in five to ten years.” The date range is wide, so his statement translates to: “We definitely want to do something with SoCs, we have a general idea, but we aren’t sure yet.”

This is likely because the potential for SoC to evolve is huge. But the technologies aren’t there yet.

The development of SoCs in smartphones in the 2010s is a good example. SoC, or system on a chip, is a single die (the part that is diced out of the wafer) that contains a variety of functions including digital, analog, mixed-signal, RF modem, and others. It works best when the company has competence in all these. Qualcomm’s Snapdragon dominated smartphones because it integrated the modem, which it had its proprietary technology for.

Chipmakers continue to try to pack more functions on a SoC. This has been incredibly difficult, hence most hardware accelerators such as AI, neural processing, and graphic processing are interconnected with the SoC die on the packaging level instead. Because packaging technology has evolved so much, companies like TSMC have been able to drastically improve the bandwidth.

Intel and AMD’s solution to this has been the chiplet, a comprise of sorts but still innovative. In a chiplet, the SoC die is separated into multiple dies according to different functions. So the general performance of the system continues to be upgraded in an economically feasible matter. Research now is focused on bonding these dies more closely together before packaging, or hybrid bonding, an area Samsung like the others is working on. Despite these developments, the effort to pack more functions on a single SoC die will continue as this sort of integration, if it can be done cost-effectively, is still the most advantageous.

New device solutions for a new era

The part that memory plays in these chip packages has only grown, as shown in the example of HBM. Samsung is also working on what it calls processor in memory, or PIM, where some functions of the processor are embedded in the memory unit itself, again for bandwidth. Compute Express Link is also another with similar aims for data centers.

These kinds of projects mean the memory, logic, and foundry parts have to work closely together. This is why Samsung’s chip division being an IDM could work in its favor.

Samsung Foundry reportedly won the order from Tesla to manufacture their self-driving chip. While it is true that Samsung is a competitor to many of its customers, there are also large companies that are not. Some competitors are arguably less concerned with protecting their proprietary technologies when it comes to chips than others (Google, for example).

Besides Tesla, the automotive industry is also presenting new opportunities. According to Omdia, the automotive sector of the semiconductor market grew 15% in 2023 from a year earlier to $75 million. The increase in electric vehicles and integration of intelligence into automobiles are driving up the demand for semiconductors, the research firm noted. Samsung also happens to make the display panels and batteries used in these vehicles.

Like AI, the idea of connected cars has been floating around for years. But the success of ChatGPT, as opposed to AlphaGo, shows that market conditions are rapidly changing and any of these concepts can quickly become a reality today. Nvidia’s success is also a great example. Before its surge from 2022 to 2023, the GPU maker also saw a boom in demand from cryptocurrency.

The semiconductor industry ran on product cycles during the PC era and in smartphones. This meant that “first to market” was a priority for chip suppliers like Samsung. And most chips, even logic, were general-purpose chips, rather than being customized. Why the semiconductor market will continue to run on cycles, today it looks more volatile than ever before.

“Accelerated computing has reached the tipping point, general purpose computing has run out of steam,” Nvidia CEO Jensen Huang said during the unveiling of Blackwell. “We need another way of doing computing, so that we can continue to scale so that we can continue to drive down the cost of computing, so that we can continue to consume more and more computing while being sustainable. Accelerated computing is a dramatic speedup over general-purpose computing, in every single industry.” 

Memory alone might be enough for Samsung to retake the No.1 position in the short term. “We cannot maintain the No.1 position with exiting businesses alone,” Samsung CEO Kyung admitted.

The device in Device Solutions refers to semiconductor devices such as transistors. It also refers to devices such as smartphones that use them, which Samsung’s chip division, by its name, is propositioning that it will provide the solutions for. While being the first to market is important, this has mostly been the mindset of the Memory Business. Coming up with chip solutions that leverage its combined strength may be needed if it wants to not only take the top spot but keep it.

Source Link