AI applications is driving the memory market forward, with HBM (High Bandwidth Memory) undoubtedly being a sought-after product of the industry, attracting increased capital expenditure and production expansion from memory manufacturers. At the meantime, a new force in the memory market has quietly emerged: GDDR7 is expected to drive the memory market steadily forward as HBM amid the AI wave.

• The Differences between GDDR7 and HBM

GDDR7 and HBM both belong to the category of graphics DRAM with high bandwidth and high-speed data transmission capabilities, providing strong support for AI computing. However, GDDR7 and HBM differ slightly in terms of technology, application scenarios, and performance.

GDDR7 is the latest technology in the GDDR family primarily used to enhance the available bandwidth and memory capacity of GPU. In March 2024, JEDEC, the Solid State Technology Association, officially released the JESD239 GDDR7 standard, which significantly increases bandwidth, eventually reaching 192GB/s per equipment.

It can be calculated that the memory speed is 48Gbps, double that of GDDR6X, the number of independent channels double from 2 in GDDR6 to 4 in GDDR7, and it supports densities ranging from 16-32 Gbit, including support for 2-channel mode to double system capacity.

Additionally, JESD239 GDDR7 is the first JEDEC-standard DRAM to use a Pulse Amplitude Modulation (PAM) interface for high-frequency operation. Its PAM3 interface improves the signal-to-noise ratio (SNR) in high-frequency operations while improving energy efficiency.

GDDR7 is mainly applied in graphics processing, gaming, computing, networking, and AI, particularly in gaming, where its high bandwidth and high-speed data transmission capabilities can significantly improve frame smoothness and loading speed, enabling a better experience for game players. In the field of AI, GDDR7 boasts great potential, capable of supporting rapid data processing and computation for large AI models, thus speeding up model training and inference.

Michael Litt, chairman of the JEDEC GDDR Task Group, has stated that GDDR7 is the first to focus not only on bandwidth but also on integrating the latest data integrity features to meet the market demands for RAS (Reliability, Availability, and Serviceability). These features allow GDDR devices to better serve existing markets like cloud gaming and computing, and expand its presence to AI sector.

Based on memory stacking technology, HBM connects layers through Through-Silicon Via (TSV), and features high capacity, high bandwidth, low latency, and low power consumption. Its strength lies in breaking the memory bandwidth and power consumption bottleneck. Currently, HBM is mainly used in AI server and supercomputer applications.

Since the introduction of the first generation in 2013, HBM has developed the second generation (HBM2), third generation (HBM2E), fourth generation (HBM3), and fifth generation (HBM3E).

This year, HBM3e will be the mainstream in the market, with concentrated shipments expected in 2H24. Besides, the sixth generation HBM4 is anticipated to make its debut as early as 2025. Reportedly, HBM4 will bring revolutionary changes, adopting a 2048-bit memory interface, which theoretically can double the transmission speed again.

• Three Memory Giants Scramble for the Initiative in GDDR7 Market

Due to high technical barriers, HBM market share is firmly at the helm of the three major memory players: SK Hynix, Samsung, and Micron. With the ongoing influence of AI, their competition has been expanding from HBM to GDDR field.

Since the beginning of this year, the three manufacturers have successively announced the availability of GDDR7 memory samples. It’s expected that some of them will start mass production of GDDR7 between 4Q24 and 1Q25.

Photo credit: Samsung Electronics

In March, Samsung and SK Hynix announced their respective GDDR7 specifications. Samsung’s GDDR7 chip, using PAM3 signal for the first time, can achieve a speed of 32Gbps at a DRAM voltage of only 1.1V, exceeding the JEDEC GDDR7 specification of 1.2V.

SK Hynix’s latest GDDR7 product, compared to its predecessor GDDR6, offers a maximum bandwidth of 160GB/s, double that of the previous generation, with a 40% improvement in power efficiency and a 1.5 times increase in memory density.

In June, Micron announced it already begun sampling its new generation of GDDR7, achieving a speed of 32Gbps and a memory bandwidth of 1.5TB/sec, a 60% improvement over GDDR6, boasting the industry’s highest bit density. Micron’s GDDR7 utilizes 1β DRAM technology and an innovative architecture and has four independent channels to optimize workloads, offering faster response time, smoother gaming experience, and shorter processing time.

Additionally, Micron’s GDDR7 improves energy efficiency by 50% relative to GDDR6, which hence enhances thermal performance for portable devices (Like laptop) and extends battery lifespan. The new sleep mode can reduce standby power consumption by 70%. Micron claims its next-generation GDDR7 can deliver high performance, increasing throughput by 33% and reducing response time for generative AI workloads (Text and image creation included) by 20%.

Photo credit: Micron

Recently, rumor has it that NVIDIA RTX 50 series will fully adopt the latest GDDR7, with a maximum capacity of 16GB, including models GN22-X11 (16 GB GDDR7), GN22-X9 (16 GB GDDR7), GN22-X7 (12 GB GDDR7), GN22-X6 (8 GB GDDR7), GN22-X4 (8 GB GDDR7), and GN22-X2 (8 GB GDDR7). The industry believes that GDDR7 will become a new arena in the memory market following HBM, in which manufacturers will continue to battle for NVIDIA GPU orders.

Read more

• [News] SK hynix to Kick off Mass Production for GDDR7 in Q4 2024
• [News] Micron Reportedly Targets 25% HBM Market Share by 2025

(Photo credit: Samsung Electronics)

U.S. Commerce Secretary Gina Raimondo previously mentioned during an online Intel foundry event that the U.S. must continue investing to regain global leadership and requires “Chip Act 2.”

According to a report from TechNews citing from global media Tom’s Hardware, the U.S. Department of Commerce plans to announce additional subsidies for the semiconductor bill as soon as this week.

Raimondo is scheduled to attend the “Revitalizing American Innovation” conference hosted by the Center for Strategic and International Studies (CSIS) in Washington on February 26th and will unveil the latest subsidies under the “Chip Act.”

In this regard, Intel is expected to receive a government subsidy of USD 10 billion, while TSMC and Samsung may also be included in the latest subsidy list. Samsung Electronics is, according to its own expectation, investing USD 17 billion to construct a foundry in Taylor, Texas, while TSMC is investing roughly USD 40 billion to build a foundry in Phoenix, Arizona. However, it’s rumored that due to the U.S. prioritizing domestic companies, the expected subsidy amounts may differ from those of Intel.

The U.S. government enacted the “Chip Act” in 2022, but subsidies have been modest, with only three American companies currently benefiting, including BAE Systems, GlobalFoundries, and Microchip Technology.

Due to Intel’s investment of USD 43.5 billion in the United States since 2021, constructing new semiconductor plants, sources cited by the report believe that the likelihood of Intel receiving USD 10 billion (equivalent to 23% of the investment amount) is quite high.

Read more

• [News] GlobalFoundries Reportedly Becomes Third Recipient of CHIPS Act Subsidy, Receives USD 1.5 Billion
• [News] US Allocates USD 39 Billion Subsidy to Semiconductor Industry for Establishing Plants

(Photo credit: TSMC)

The latest financial reports for 4Q23 from six leading global semiconductor foundries signal optimism for the semiconductor industry’s recovery in 2024.

In 2023, the semiconductor sector underwent significant adjustments. As the industry worked towards normalizing its inventory levels amidst ongoing high inflation risks, the short-term market outlook remained unclear. #TrendForce has analyzed the latest financials from these six foundries to provide insights into what 2024 might hold for the industry.

TSMC

TSMC reported a slight YoY revenue decrease of 1.5% to US$19.62 billion in 4Q23, though it saw a 13.6% increase from the previous quarter. With an anticipated CAGR of 15–20%, TSMC’s 2024 capital expenditures are expected to be between $28 billion and $32 billion.

The company forecasts more than 10% growth in the semiconductor market (excluding memory) and around 20% growth in the wafer fabrication sector for 2024.

Samsung Electronics

Samsung Electronics’ 4Q23 consolidated revenue fell 3.81% YoY to ₩67.78 trillion. Its DS division reported revenues of ₩21.69 trillion but faced an operating loss of ₩2.18 trillion.

Despite the challenges, Samsung is focusing on advancing 3nm and 2nm GAA process technologies, expecting a revival in smartphone and PC demand in 2024 to rejuvenate the foundry market to its former prosperity.

Intel

Intel’s 4Q23 earnings saw a 10% revenue increase to $15.406 billion, with its foundry business, Intel Foundry Services, jumping 63% to $291 million in revenue.

Despite seasonal demand slumps in its core PC and server segments, Intel’s AI chips have accumulated $2 billion in orders, with sales forecast to improve in the second half of the year.

Global Foundries

GlobalFoundries reported a 12% revenue drop in 4Q23 to $1.85 billion, with a net income of $356 million. The company anticipates 1Q24 revenues to range between $1.5 billion and $1.54 billion, primarily due to the current industry-wide chip inventory adjustments.

Nevertheless, GlobalFoundries expects its 2023 automotive market revenue to surpass $1 billion, forecasting continued growth into 2024.

UMC

UMC disclosed a 19% YoY decrease in 4Q23 revenues to $1.79 billion. The company cited an extended semiconductor industry inventory adjustment period due to a challenging global economic climate, leading to a slight reduction in wafer shipments and capacity utilization. UMC expects a gradual uptick in wafer demand through 1Q24.

SMIC

SMIC reported a modest increase in 4Q23 revenues to $1.68 billion, with a 0-2% growth projection for 1Q24. Despite last year’s cyclical lows and competitive pressures, SMIC anticipates its 2024 revenue growth will at least match the industry average, with capital expenditures mirroring those of 2023.

TrendForce had earlier forecasted a delayed recovery in the end-market by the fourth quarter of 2023. However, they noted that inventory stocking by Chinese Android firms for the year-end sales rush—particularly for mid-to-low-end 5G and 4G smartphone application processors—alongside the influence of new Apple iPhone releases, might surpass initial expectations.

This indicates that the revenues of the world’s top ten semiconductor foundries are poised for growth, potentially surpassing the growth rates observed in the third quarter.

(Photo credit: Samsung)

The competition for dominance in 2nm semiconductor technology has intensified at the beginning of 2024, marking a crucial battleground among global foundry companies.

As per a report from IJIWEI, major foundry enterprises such as Samsung Electronics, TSMC, and Intel are set to commence mass production adopting 2nm process starting this year. Consequently, the fierce competition for supremacy in 2nm technology is expected to escalate from 2025 onwards. Currently, the most advanced production technology globally is at the 3nm level.

• TSMC

TSMC’s 2nm products will be manufactured at the Fab 20 in the Hsinchu Science Park in northern Taiwan and at a plant in Kaohsiung.

The Fab 20 facility is expected to begin receiving related equipment for 2nm production as early as April, with plans to transition to GAA (Gate-All-Around) technology from FinFET for 2nm mass production by 2025.

During TSMC’s earnings call on January 18th, TSMC revealed that its capital expenditure for this year is expected to fall between USD 28 billion and 32 billion, with the majority (70% to 80%) allocated to advanced processes. This figure is similar to that of 2023 (USD 30.4 billion), indicating stable investment to ensure its leading position in 2nm technology.

• Intel

After announcing its re-entry into the foundry business, Intel is actively advancing its foundry construction efforts. The plan includes the introduction of the Intel 20A (equivalent to 2nm) process in the first half of 2024 and the Intel 18A (1.8nm) process in the second half of the year. It is understood that the Intel 18A process will commence test production as early as the first quarter of this year.

Intel’s 2nm roadmap is more ambitious than originally anticipated, being accelerated by over six months. In response to criticisms of its “overly ambitious” plans, Intel swiftly began procuring advanced Extreme Ultraviolet (EUV) equipment.

• Samsung Electronics

Samsung Electronics has devised a strategy to gain an advantage in the more advanced process war through its Gate-All-Around (GAA) technology. Currently, it is mass-producing the first-generation 3nm process based on GAA (SF3E) and plans to commence mass production of the second-generation 3nm process this year, significantly enhancing performance and power efficiency.

Regarding the 2nm process, per a report from Nikkei, Samsung plans to start mass production for mobile devices in 2025 (SF2) and gradually expand to high-performance computing (HPC) in 2026 and automotive processes in 2027.

Currently, Samsung Electronics is producing GAA products for the 3nm process at its Hwaseong plant and plans to manufacture products for both the 3nm and 2nm processes at its Pyeongtaek facility in the future.

• Rapidus

Rapidus, a chip manufacturing company supported by the Japanese government, is expected to trial-adopt 2nm process at its new plant by 2025 and begin mass production from 2027.

If Rapidus’ technology is validated, the global foundry market may expand beyond the Taiwan-Korea duopoly to include Taiwan, Korea, the United States, and Japan.

The technology competition to become a “game-changer” ultimately depends on the competition for customers. It’s rumored that TSMC holds a leading position in the 2nm field, with Apple speculated to be its first customer for the 2nm process. Graphics processing giant NVIDIA is also considered a major customer within TSMC’s client base.

According to TrendForce data as of the third quarter of 2023, TSMC’s revenue share accounted for a dominant 57.9%, with Samsung Electronics trailing at 12.4%, a gap of 45.5 percentage points.

However, Samsung Electronics is not sitting idly by. With continuous technological investment, Samsung’s foundry customer base grew to over 100 in 2022, a 2.4-fold increase from 2017. The company aims to expand this number to around 200 by 2028.

Particularly, Samsung’s early adoption of GAA technology is expected to give it an advantage in achieving early production volumes for advanced processes.

Read more

• Top 10 Foundries Experience 7.9% QoQ Growth in 3Q23, with a Continued Upward Trend Predicted for Q4, Says TrendForce
• [News] Intense Competition with Samsung and Intel in Advanced Processes; TSMC Speeds Up 2nm Progress

(Photo credit: TSMC)

Nvidia CFO, Colette Kress, recently hinted again that the next-gen chips might be outsourced to Intel Corp. During the call with semiconductor analyst Tim Arcuri at the UBS Global Technology Conference on November 28th, she was asked whether Intel would be considered as a foundry partner for the next-gen chips.

In response, she stated that there are many powerful foundries in the market. TSMC and Samsung Electronics have been great partners. She said, “we’d love to have a third one,” when answering whether Nvidia want a third partner.

Kress also mentioned that, TSMC’s and others’ US fab may also be their options, and “there is nothing necessarily different but again in terms of different region. Nothing will stop us from potentially adding another foundry.”

Kress highlighted that Nvidia’s current data center GPUs designed for AI and high-performance computing (HPC) are predominantly outsourced to TSMC. However, in the previous generation, Nvidia’s gaming GPUs were mainly entrusted to Samsung for fabrication. According to Sedaily, Samsung’s foundry was responsible for manufacturing Nvidia’s GeForce RTX 30 series gaming GPUs based on the Ampere architecture.

Speaking of foundry partners for AI products, Nvidia anticipates that TSMC will remain a crucial foundry partner for producing AI Hopper H200 and Blackwell B100 GPUs. Any additional orders might be entrusted to Samsung.

Nvidia CEO previously said Intel’s next-gen process test chips “look good”

Additionally, reports from Barron also mentioned that on May 30th, during an interaction with journalists in Taiwan, Nvidia CEO Jensen Huang was asked whether Nvidia is considering diversifying its supplier base given the rising tensions between the U.S. and China. In response, Huang referred to Nvidia’s long-standing collaboration with TSMC and Samsung Electronics, stating, “We have a lot of customers depending on us. And so our supply chain resilience is very important to us. We manufacture in as many places as we can.”

At that time, Huang also expressed, “We’re open to manufacturing with Intel. And (Intel CEO) Pat (Gelsinger) has said in the past that we’re evaluating their process, and we’ve recently received the test chip results of their next generation process and the results look good.”

From Nvidia CFO’s talk in November and Nvidia CEO’s response in May, it is obvious that, beyond TSMC and Samsung, Nvidia is thinking about a potential third foundry partner.

(Image: NVIDIA Hopper Architecture – H100 SXM)