With the advancements in AIGC models such as ChatGPT and Midjourney, we are witnessing the rise of more super-sized language models, opening up new possibilities for High-Performance Computing (HPC) platforms.

According to TrendForce, by 2025, the global demand for computational resources in the AIGC industry – assuming 5 super-sized AIGC products equivalent to ChatGPT, 25 medium-sized AIGC products equivalent to Midjourney, and 80 small-sized AIGC products – would be approximately equivalent to 145,600 – 233,700 units of NVIDIA A100 GPUs. This highlights the significant impact of AIGC on computational requirements.

Additionally, the rapid development of supercomputing, 8K video streaming, and AR/VR will also lead to an increased workload on cloud computing systems. This calls for highly efficient computing platforms that can handle parallel processing of vast amounts of data.
However, a critical concern is whether hardware advancements can keep pace with the demands of these emerging applications.

HBM: The Fast Lane to High-Performance Computing

While the performance of core computing components like CPUs, GPUs, and ASICs has improved due to semiconductor advancements, their overall efficiency can be hindered by the limited bandwidth of DDR SDRAM.

For example, from 2014 to 2020, CPU performance increased over threefold, while DDR SDRAM bandwidth only doubled. Additionally, the pursuit of higher transmission performance through technologies like DDR5 or future DDR6 increases power consumption, posing long-term impacts on computing systems’ efficiency.

Recognizing this challenge, major chip manufacturers quickly turned their attention to new solutions. In 2013, AMD and SK Hynix made separate debuts with their pioneering products featuring High Bandwidth Memory (HBM), a revolutionary technology that allows for stacking on GPUs and effectively replacing GDDR SDRAM. It was recognized as an industry standard by JEDEC the same year.

In 2015, AMD introduced Fiji, the first high-end consumer GPU with integrated HBM, followed by NVIDIA’s release of P100, the first AI server GPU with HBM in 2016, marking the beginning of a new era for server GPU’s integration with HBM.

HBM’s rise as the mainstream technology sought after by key players can be attributed to its exceptional bandwidth and lower power consumption when compared to DDR SDRAM. For example, HBM3 delivers 15 times the bandwidth of DDR5 and can further increase the total bandwidth by adding more stacked dies. Additionally, at system level, HBM can effectively manage power consumption by replacing a portion of GDDR SDRAM or DDR SDRAM.

As computing power demands increase, HBM’s exceptional transmission efficiency unlocks the full potential of core computing components. Integrating HBM into server GPUs has become a prominent trend, propelling the global HBM market to grow at a compound annual rate of 40-45% from 2023 to 2025, according to TrendForce.

The Crucial Role of 2.5D Packaging

In the midst of this trend, the crucial role of 2.5D packaging technology in enabling such integration cannot be overlooked.

TSMC has been laying the groundwork for 2.5D packaging technology with CoWoS (Chip on Wafer on Substrate) since 2011. This technology enables the integration of logic chips on the same silicon interposer. The third-generation CoWoS technology, introduced in 2016, allowed the integration of logic chips with HBM and was adopted by NVIDIA for its P100 GPU.

With development in CoWoS technology, the interposer area has expanded, accommodating more stacked HBM dies. The 5th-generation CoWoS, launched in 2021, can integrate 8 HBM stacks and 2 core computing components. The upcoming 6th-generation CoWoS, expected in 2023, will support up to 12 HBM stacks, meeting the requirements of HBM3.

TSMC’s CoWoS platform has become the foundation for high-performance computing platforms. While other semiconductor leaders like Samsung, Intel, and ASE are also venturing into 2.5D packaging technology with HBM integration, we think TSMC is poised to be the biggest winner in this emerging field, considering its technological expertise, production capacity, and order capabilities.

In conclusion, the remarkable transmission efficiency of HBM, facilitated by the advancements in 2.5D packaging technologies, creates an exciting prospect for the seamless convergence of these innovations. The future holds immense potential for enhanced computing experiences.

Samsung recently announced that they will ahead of TSMC in the foundry market within 5 years. At the same time, Intel also claimed to become the second-largest player in the market before 2030. Currently, both Samsung and TSMC are adapting 3nm process to do the chip manufacturing, with the technology of GAA(Samsung) and FinFET(TSMC) respectively.

Samsung sees GAA technology as a crucial key to surpassing TSMC. Currently, Samsung’s 4nm lags behind TSMC by about 2 years, and its 3nm is about a year behind. However, this situation will change when TSMC turns to 2nm. Industry insider sources indicate that TSMC plans to use GAA technology in 2nm process, and Samsung believes that they can seize the chance to catch up with TSMC since TSMC may have a hard time when turning to 2nm process.

Industry insiders have revealed that AMD has shifted some of its 4nm CPU chip orders from TSMC to Samsung. It is reported that AMD has signed an agreement with Samsung to manufacture some of its mobile SoC by using Samsung’s 4nm node, and Samsung may also manufacture AMD’s Chromebook APU.

The Fight in the Foundry Market is On

According to TrendForce, the top 10 global foundry players in 4Q22 with TSMC account for 58.5% of market share by revenue, far ahead of Samsung’s 15.8%. Industry insiders suggest that Samsung still has a long way to go to catch up with TSMC. Some sources say that TSMC’s 2nm process will be mass-produced as scheduled in 2025, while Samsung’s plans are still to be observed.

Intel is also striving for the top spot in the wafer foundry market. Since the beginning of 2021, Intel has implemented a series of measures in its foundry business after announcing its “IDM 2.0” strategy. Last July, Intel stated that it will manufacture chips for MediaTek, and the first batch of products will be produced within the next 18 to 24 months using more mature manufacturing technology (Intel 16). In addition, Intel said that Qualcomm and Nvidia are also interested in having them manufacture their chips. To regain its leading edge in chip manufacturing, Intel has unveiled its 5 process technology stages to be launched in the next few years, including 10nm, 7nm, 4nm, 3nm, and 20A.

And TSMC has no competitive relationship with their clients by not doing the wafer design, apparently, this is also a significant advantage for TSMC and other foundry manufacturers. In recent years, more companies have recognized the importance and highly profitable nature of foundry manufacturing, leading to the independent establishment of foundry manufacturing operations. Samsung and Intel have also followed this trend, as foundry manufacturing can optimize production technology and provide major companies with more opportunities for trial and error.

The market started worrying about the oversupply in semiconductor 2023, when the demand will start growing again depends on two factors: the situation of the macroeconomy and the inventory status.

Since foundries’ capacity utilization rates started drop in 3Q22, chip supply as a whole has decreased significantly. This, in turn, has helped limit inventory growth across the supply chain. However, the global economy is still at risk of a mild recession, so consumers may allocate more of their spending to daily necessities. They may also spend more on tourism due to easing of the pandemic. This could lead to weak sales for consumer electronic products.

Not to mention that most consumers already purchased the electronic products that they need for working or studying at home during the pandemic. Assuming that the overall inventory level of the supply chain will return to a healthier level, TrendForce believes that chip demand will begin to rebound to a certain extent in 2Q23. Then, the demand growth will become more obvious from 3Q23 onward. Nevertheless, this demand growth may not be too strong due to uncertainties in the global economy.

If we observe the situation from the perspective of the foundry industry, smartphones represent the largest application segment in terms of wafer consumption. The smartphone supply chain started inventory correction earlier, so demand rebound might be more obvious initially for smartphone-related chips compared with chips used in other consumer devices. On the other hand, with different benchmarks, the demand for HPC chips will show more significant growth compared with the demand for smartphone chips.

The output value of the world’s top 10 foundries in 4Q21 reached US$29.55 billion, or 8.3% growth QoQ, according to TrendForce’s research. This is due to the interaction of two major factors. One is limited growth in overall production capacity. At present, the shortage of certain components for TVs and laptops has eased but there are other peripheral materials derived from mature process such as PMIC, Wi-Fi, and MCU that are still in short supply, precipitating continued fully loaded foundry capacity. Second is rising average selling price (ASP). In the fourth quarter, more expensive wafers were produced in succession led by TSMC and foundries continued to adjust their product mix to increase ASP. In terms of changes in this quarter’s top 10 ranking, Nexchip overtook incumbent DB Hitek to clinch 10th place.

TrendForce believes that the output value of the world’s top ten foundries will maintain a growth trend in 1Q22 but appreciation in ASP will still be the primary driver of said growth. However, since there are fewer first quarter working days in the Greater China Area due to the Lunar New Year holiday and this is the time when some foundries schedule an annual maintenance period, 1Q22 growth rate will be down slightly compared to 4Q21.

Top 5 foundries account for nearly 90% of global market share, Samsung recovers share with advanced processes

Looking at the top five industry players, TSMC’s 4Q21 revenue reached US$15.75 billion, a QoQ increase of 5.8%. Although 5nm revenue spiked thanks to the new iPhone, 7/6nm revenue dropped due to a weak Chinese smartphone market, becoming the only TSMC node in decline in 4Q21, and inducing a contraction in TSMC revenue growth in 4Q21, though TSMC still accounts for more than 50% of global market share. As one of TSMC’s few competitors in advanced processes below 7nm, Samsung strengthened 4Q21 revenue to US$5.54 billion, a quarterly increase of 15.3% owing to the gradual completion of new advanced 5/4nm process capacity and the mass production of new flagship products from major client Qualcomm. Although Samsung’s foundry business has posted record revenue, the slower ramp-up of advanced process capacity continues to erode overall profitability. Therefore, TrendForce believes that improving advanced process capacity and yield in 1Q22 is one of Samsung’s top priorities.

Constrained by limited growth in new production capacity and the fact that the new wave of wafers contracted at the latest pricing has yet to be produced, UMC’s revenue stalled slightly in 4Q21, to US$2.12 billion, up 5.8% QoQ. GlobalFoundries benefited from the release of new production capacity, product mix optimization, and new long-term agreement (LTA) pricing, pushing up ASP performance. Revenue in 4Q22 hit US$1.85 billion, up 8.6% QoQ. SMIC posted 4Q21 revenue of US$1.58 billion, 11.6% QoQ, due to mounting demand for products such as HV, MCU, Ultra Low Power Logic, and Specialty Memory as well as factors such as product mix adjustment and appreciating ASP.

Surpassing DB Hitek, Nexchip officially breaks into the top 10 in 4Q21

The foundries ranked 6th to 9th are HuaHong Group, PSMC, Vanguard International Semiconductor (VIS), and Tower Semiconductor (Tower), respectively. Each has benefiting from factors such as a utilization rate uniformly at full capacity, release of new production capacity, and adjustment of ASP and product mix, sustaining the growth of revenue performance. It is worth mentioning, the acquisition of Tower by Intel netted Intel mature process technologies and a customer base and expanded the diversity and production capacity of its foundry business. However, before this transaction is officially completed, Tower is still considered an independent entity in terms of the accounting process. TrendForce states, after Intel’s foundry business is properly integrated with Tower, Intel will officially enter the ranking of top ten foundries.

Coming in 10th on the top 10 foundry ranking is Nexchip with revenue of US$352 million and a quarterly growth rate of 44.2%, the fastest growth rate among the top ten, and officially surpassed DB Hitek. According to TrendForce investigations, the primary reason Nexchip was able to break into the top 10 in 4Q21 was the company’s diligent production expansion. Nexchip also plans to develop more advanced processes such as the 55/40/28nm nodes and multiple product lines including TDDI, CIS, and MCU, to compensate for its current single product line and limited customer base. Since Nexchip is currently ramping-up operations quickly, its growth performance in 2022 should not be underestimated.

Intel and AMD will be releasing new CPUs that support DDR5 DRAM solutions for PCs and servers this year. In response, the DRAM industry led by South Korean suppliers is developing solutions to complement the arrival of the new CPUs. In the midst of the gradual shift to DDR5, DRAM suppliers will also scale back the supply of DDR3 solutions, according to TrendForce’s latest investigations. With Korean suppliers accelerating their withdrawal from DDR3 production, Taiwanese suppliers yet to kick off mass production using newly installed capacities, and Chinese suppliers falling short of their expected yield rate, the global supply of DDR3 solutions will undergo an impending decline. With respect to the demand side, however, not only has the supply of networking chips been ramping up, but material shortage issues are also gradually easing. As such, buyers are now procuring DDR3 solutions ahead of time, resulting in a tight supply and demand situation in the DDR3 market. TrendForce therefore expects DDR3 DRAM prices to recover from a bearish first quarter and undergo a 0-5% QoQ increase in 2Q22.

On the supply side, Samsung and SK hynix have begun scaling back their DDR3 production while also planning to declare EOL (end of life) for their DDR3 offerings, such as 1/2Gb and 4Gb chips. It should be noted, however, that Micron’s DDR3 solutions will not reach EOL even by 2026, meaning the company will still offer DDR3 solutions long after its two Korean competitors have stopped doing so, according to TrendForce’s understanding. Also worth noting is that Micron is migrating its DDR3 production to a US-based fab that mainly manufactures specialty DRAM solutions. Nevertheless, since this fab’s production capacity will be divided between products for consumer and automotive applications, TrendForce believes that the aforementioned migration will tighten Micron’s supply of consumer DRAM solutions because the US fab will give priority to automotive DRAM solutions that offer a higher gross margin and are currently enjoying surging demand.

Although Taiwan-based DRAM suppliers that focus on promoting DDR3 solutions, namely, Nanya Tech and Winbond, are in the process of capacity expansion, their new production lines will not be operational until 2023-2024. Hence, the contribution from the newly added capacities is not expected to drive up DDR3 supply substantially this year. Chinese suppliers, including CXMT and GigaDevice, are continuing to collaborate in DDR3 development, though their capacity increases and yield rate improvements have both fallen short of market expectations. After being added to the Entity List, JHICC, yet another China-based DRAM supplier, is now dealing with severe restrictions with respect to procuring equipment, making it difficult for JHICC to raise its wafer input. Furthermore, the company has no spare resources that can be allocated to R&D and pilot runs. As a result, JHICC still primarily manufactures DDR4 4Gb chips at its initial 25nm node, with no DDR3 production at the moment.

With regards to demand, DDR3 consumer DRAM is primarily used in end-devices such as STBs and networking products (e.g., GPON, routers, and modems), which do not require high-performance SoCs. While the foundry industry suffered a severe shortage of wafer capacities allocated to logic ICs in 4Q21, production capacities for relatively low-margin chips were noticeably impacted in turn. Along with a preexisting component mismatch situation, most manufacturers found themselves unable to assemble end-devices. Moving into early 2022, however, the supply of certain materials, including those used in foundry operations, saw a gradual improvement. As various components needed for device manufacturing became available after Lunar New Year, certain buyers have once again kicked off their consumer DRAM procurement activities.

In addition, DRAM spot prices shifted from a prior decline to a strong upturn at the end of last year as the Chinese government ordered a month-long lockdown in Xi’an. The ensuing price hike, which has lasted for two months, subsequently led buyers to procure even more DRAM ahead of time in anticipation of further price hikes. Hence, although the demand for end-products has yet to make a full recovery, buyers are now slowly and steadily procuring consumer DRAM in order to avoid either higher upcoming prices or even an inability to secure consumer DRAM inventory.