News Posts matching #HBM

Return to Keyword Browsing

TSMC Showcases Custom C-HBM4E, N3P Logic Dies Target Double Efficiency

At last week's Open Innovation Platform Ecosystem Forum in Amsterdam, TSMC provided more details over its custom HBM for the HBM4 generation, according to reports from HardwareLUXX. The company's custom C-HBM4E logic die is expected to shift to the N3P node, alongside a voltage drop from 0.8 V to 0.75 V. According to TSMC's presentation, the move targets roughly 2× better power efficiency versus today's DRAM processes. Even standard HBM4 base dies will see upgrades. Instead of using a conventional DRAM process as in HBM3E, TSMC plans to manufacture HBM4 base dies on its N12 logic node, reducing operating voltage from 1.1 V to 0.8 V. The company expects this alone to provide around a 1.5× efficiency gain. For C-HBM4E specifically, the base die doesn't just move to the N3P node, it also integrates the memory controllers directly into the stack. Those blocks normally sit on the host SoC, so the PHY becomes a fully custom design as well.

Moreover, TSMC is expanding its advanced packaging lineup with updated InFO and SoW options, but CoWoS remains the company's main growth driver. After moving from 1.5× to 3.3× reticle sizes with support for eight HBM chips, CoWoS now heads toward CoWoS-L, enabling up to 12 HBM3E/HBM4 stacks for 2026 AI parts, followed by a larger A16 generation version planned for 2027. TSMC is also securing major customers for its custom HBM logic dies. Micron has selected the foundry to build the logic base die for its HBM4E parts, with volume production planned for 2027. SK Hynix is also reportedly preparing its first custom HBM4E products for the second half of next year. According to Korea Financial Times, SK Hynix will use TSMC's 12 nm process for mainstream server-grade HBM base dies, while NVIDIA's top-end GPUs and Google's TPUs will step up to a 3 nm node.

Micron Reportedly Drafting New $9.6 Billion HBM Plant in Japan

Insiders believe that Micron is planning a major expansion of operations at an existing Japanese facility. The North American computer memory manufacturer runs numerous production hubs around the world, but mid-October news reports indicated an upcoming exit from China. Late last week, a Nikkei Asia article alleged that a major $9.6 billion investment is linked to the creation of a speculative new plant at the company's Hiroshima site. Present day operations—located in West Japan—focus on the manufacturing of DRAM, but industry moles reckon that an additional factory will specialize in high-bandwidth memory (HBM).

Leaked information points to possible construction work starting around May 2026. Initial product shipments—perhaps HBM4 and HBM4E—are supposedly projected to begin in 2028. An expanded Japanese manufacturing hub could take the pressure off the firm's Taiwanese factories—currently tasked with pumping out various forms of HBM. Nikkei Asia's report alludes to incentives, leveraged by the Japanese government. The nation's Ministry of Economy, Trade and Industry (METI) could be preparing significant subsidies—secretive parties have whispered of a 500 billion yen total. Global demand for AI-oriented computer memory seems to be growing, with little respite. Suppliers are seemingly still pivoting to meet steep requests from important clients.

SK Hynix Showcases Advanced AI Memory From HBM4 to Next-Gen Storage at SC25

SK hynix showcased advanced memory technologies for the era of AI and high-performance computing (HPC) at Supercomputing 2025 (SC25), held in St. Louis, the U.S., from November 16-21. Held annually since 1988, SC is the world's largest global HPC conference. It brings together experts from industry, academia, and research institutions to share the latest trends, create opportunities for collaboration, and discuss the future direction of technological advancement. At this year's event, the convergence of AI and HPC emerged as a central topic. Under the theme "Memory, Powering AI and Tomorrow," SK hynix presented its innovative memory lineup set to lead the AI and HPC markets and shared a new technological vision to accelerate data analysis in computing systems.

Innovative Products and Tech Driving AI and HPC Performance
At the event, SK hynix presented core products including HBM, DRAM, and eSSD solutions. Visitors to the booth could also see demonstrations of the products in AI and HPC environments, highlighting the company's technological competitiveness. At the front of the booth, SK hynix displayed its latest HBM products including the 12-layer HBM4 which it developed as a world-first in September 2025. HBM4 features 2,048 input/output (I/O) channels, twice that of the previous generation, delivering a significant increase in bandwidth. Also offering more than a 40% boost in power efficiency, HBM4 is the most suitable solution for ultra-high-performance AI computing systems. In addition, the 12-layer HBM3E, the highest performing commercialized HBM currently available on the market, was presented with NVIDIA's next-generation GB300 Grace Blackwell GPU.

Next-Gen HBM Targets Embedded GPU Cores

Tech companies are exploring a major shift in HBM design seeking to embed GPU cores directly into next-generation memory stacks. According to Korean industry reports, Meta and NVIDIA are evaluating "custom HBM" architectures that integrate GPU cores into the base die of future HBM devices, with SK Hynix and Samsung said to be involved in early discussions. HBM stacks multiple DRAM dies on top of a base die that handles external I/O. HBM4 is expected to reach mass production next year and will include an onboard controller to improve bandwidth and efficiency. Integrating GPU cores would take this concept several steps further, distributing compute into the memory itself to reduce data movement and cut power usage. Industry sources say the approach could boost both performance and energy efficiency for AI workloads by cutting the distance between compute and memory. But the design still faces major struggles such as limited die area in Through-Silicon Vias (TSV) based stacks, power delivery constraints, and the difficulty of cooling those compute-heavy GPU logic inside the base die.

Kim Joung-ho, a professor in the School of Electrical Engineering at KAIST, said, "The speed of technological transition where the boundary between memory and system semiconductors collapses for AI advancement will accelerate," and added, "Domestic companies must expand their ecosystem beyond memory into the logic sector to preempt the next-generation HBM market." AMD's recently unveiled Instinct MI430X accelerator built on the next-generation AMD CDNA architecture supports 432 GB of HBM4 memory and 19.6 TB/s of memory bandwidth. NVIDIA's "Vera Rubin" Superchip takes a different approach. Each Rubin GPU integrates two reticle-sized compute chiplets paired with eight HBM4 stacks, delivering around 288 GB of HBM4 per GPU and roughly 576 GB of HBM4 across the full Superchip. From those potential changes companies that have strong packaging and logic capabilities will benefit, while pure memory vendors may need to expand into system-level semiconductor technologies to stay competitive.

Global DRAM Revenue Jumps 30.9% in 3Q25, Micron's Market Share Climbs by 3.7 Percentage Points

TrendForce's latest research shows that significant increases in conventional DRAM contract prices, higher bit shipments, and growing HBM volumes drove the global DRAM industry revenue to US$41.4 billion in 3Q25, marking a strong 30.9% QoQ growth.

Looking ahead to Q4 2025, DRAM suppliers' inventories are almost depleted, and growth in bit shipments will decrease substantially. Regarding pricing, CSPs continue to be relatively flexible with procurement costs, which causes other applications to also increase their prices to secure supply. Consequently, contract prices for both advanced and legacy nodes, as well as across all major applications, are anticipated to increase rapidly. TrendForce predicts that conventional DRAM contract prices will rise by 45-50% QoQ, and total contract prices (including DRAM and HBM) will increase by 50-55%.

SK Hynix Shares Next-Generation AI Memory Strategy and Roadmap at the SK AI Summit 2025

SK hynix shared its next-generation AI memory strategy and roadmap at the SK AI Summit 2025 held in Seoul from November 3-4. At the summit held under the theme "AI Now & Next," the company elevated its vision from becoming a full-stack AI memory provider to embracing a broader role as a full-stack AI memory "creator." Hosted annually by SK Group, the event was formerly known as the SK Tech Summit but has evolved with a greater focus on AI to become the SK AI Summit since 2024. This year's event brought together global AI experts and industry professionals to discuss the present and future of AI technologies, as well as its applications across industries.

Among those in attendance were SK Group Chairman Chey Tae-won, SK hynix CEO Kwak Noh-Jung, and SK Telecom CEO Jung Jaihun. They were joined by influential tech leaders including OpenAI CEO Sam Altman and Amazon CEO Andy Jassy, who delivered keynote speeches via video. Furthermore, global tech giants including NVIDIA, Amazon Web Services (AWS), Google, TSMC, and Meta, together with numerous domestic members of the K-AI Alliance, participated to discuss collaboration opportunities across the AI ecosystem.

AMD Instinct MI430X Targets HPC Workloads with 432GB of HBM4 Memory

AMD has unveiled the Instinct MI430X, a data center accelerator designed for the intersection of high-performance computing and large-scale AI. Deployments are already underway at two of the world's leading research facilities. The GPU is being integrated into the Discovery supercomputer at Oak Ridge National Laboratory and the Alice Recoque exascale system in Europe. Both systems will utilize the chip alongside AMD's upcoming EPYC "Venice" processors. By launching directly into production systems rather than through a preliminary announcement, AMD is making the MI430X as the immediate successor to its previous Instinct lineup and a key element for countries and organizations aiming to develop self-reliant, open AI infrastructure.

The MI430X tackles the memory bottlenecks that challenge modern AI and simulation workloads with an impressive 432 GB of HBM4 memory, delivering approximately 19.6 TB/s of bandwidth. These specifications essentially transform the accelerator into a near-memory-compute device for trillion-parameter models. The chip is powered by AMD's next-generation CDNA architecture, which maintains hardware-level double-precision FP64 support for traditional scientific computing while introducing native FP8 and FP4 precision modes for efficient AI training and inference. The design also focuses on performance-per-watt improvements, catering to environments where power efficiency is as crucial as raw performance for sustained exascale operations.

AMD Unveils First MLPerf 5.1 Training Results on Instinct MI350 Series GPUs

The latest MLPerf 5.1 Training results mark an important milestone—the first MLPerf training submission using AMD Instinct MI350 Series GPUs. These new benchmarks demonstrate breakthrough generational performance gains and highlight broad ecosystem participation across some of today's most demanding AI training workloads.

Building on the strong MLPerf 5.1 Inference results released earlier this year, this submission represents the first time the AMD Instinct MI350 Series—including both the MI355X and MI350X GPUs—have been publicly benchmarked for AI training. The results show clear progress in scalability, efficiency, and compute performance, underscoring how the AMD Instinct MI350 Series is accelerating the development of next-generation generative AI models.

Memory Industry to Maintain Cautious CapEx in 2026, with Limited Impact on Bit Supply Growth

TrendForce's latest investigations reveal that despite higher ASPs boosting profitability across the memory industry, capital spending on DRAM and NAND Flash is only anticipated to increase modestly in 2026. This limited investment growth is unlikely to significantly affect bit output. Instead, the emphasis is shifting from capacity expansion to advancements such as process technology upgrades, higher-layer stacking, hybrid bonding, and high-value products such as HBM.

In the DRAM sector, total capital expenditure is expected to reach US$53.7 billion in 2025 and rise to $61.3 billion in 2026, representing about 14% YoY growth. For NAND Flash, spending is projected to rise from $21.1 billion in 2025 to $22.2 billion in 2026, a roughly 5% increase.

SK hynix Showcases Full-Stack AI Memory Portfolio at 2025 OCP Global Summit

SK hynix showcased its full-stack AI memory portfolio for enhancing AI and data center infrastructure at the 2025 OCP Global Summit in San Jose, California from October 13-16. The OCP Global Summit is a conference hosted by the Open Compute Project (OCP), the world's largest open data center technology collaboration community. Held under the theme "Leading the Future of AI", this year's event welcomed experts from leading global companies and developers to share the latest trends in data center and AI infrastructure and discuss developing industry solutions.

Under the slogan "MEMORY, Powering AI and Tomorrow," SK hynix presented a range of innovative technologies which improve the performance and efficiency of AI infrastructure. The company's interactive booth was organized into four sections - HBM, AiM, DRAM, and eSSDs - and featured product-based character designs, 3D models of key technologies, and live demonstrations. The centerpiece of SK hynix's booth was its groundbreaking 12-layer HBM4. In September 2025, the company became the first in the world to complete development of HBM4 and establish the product's mass production system. Featuring 2,048 input/output (I/O) channels, twice that of the previous generation, HBM4 offers increased bandwidth and over 40% greater power efficiency. These specifications make it optimized for ultra-high-performance AI computing systems.

NVIDIA and SK Group Build AI Factory to Drive Korea's Manufacturing and Digital Transformation

NVIDIA today announced that it is working with SK Group to build an AI factory to advance semiconductor research, development and production, as well as cloud infrastructure to support digital twin and AI agent development. SK Group is building an AI factory featuring more than 50,000 NVIDIA GPUs, with the first phase planned for completion by late 2027. Once complete, the system is expected to be one of Korea's largest AI factories.

The new factory will serve SK subsidiaries - including SK hynix and SK Telecom (SKT) - as well as external organizations through a GPU-as-a-service model, accelerating digital transformation and industrial innovation for Korea's industries. Further expanding the NVIDIA and SK Group partnership, the companies are collaborating on the development of SK hynix high-bandwidth memory (HBM) and next-generation advanced memory solutions for NVIDIA GPUs, semiconductor manufacturing and telecommunications infrastructure.

NVIDIA and Samsung Build AI Factory to Transform Global Intelligent Manufacturing

NVIDIA today announced plans with Samsung Electronics to build a new AI factory, representing a new era where intelligent computing and chip manufacturing converge. The state-of-the-art AI factory will combine Samsung's semiconductor technologies with NVIDIA platforms to establish the foundation of next-generation, AI-driven production. Powered by more than 50,000 NVIDIA GPUs, Samsung's semiconductor AI factory will be a centerpiece of the company's digital transformation, integrating accelerated computing directly into full-fledged advanced chip manufacturing.

Through this collaboration, Samsung and NVIDIA are setting a global benchmark for AI-driven semiconductor manufacturing at scale, integrating data from physical equipment and production workflows to achieve predictive maintenance, process improvements and increased operational efficiency in autonomous fab environments.

Samsung Ships HBM4 Samples to Customers, Mass Production in 2026

Samsung has just reported its third quarter 2025 results, and the company has confirmed that its next-generation HBM4 memory samples has been shipped to customers worldwide, with mass production expected in 2026. "HBM3E is currently in mass production and being sold to all related customers, while HBM4 samples are simultaneously being shipped to key clients," notes Samsung, confirming that its memory business is thriving and advanced memory solutions are in great demand. In addition, the company has confirmed that "In 2026, the Foundry Business will focus on providing a stable supply of new 2 nm GAA products and the HBM4 base-die, and beginning operations at the Company's fab in Taylor, Texas in a timely manner."

In an HBM stack, consisting of stacked DRAM dies up to a 12-high, connected with TSVs, there is the possibility of embedding an optional base die with customized logic/accelerator circuitry tailored to a specific need. Companies usually opt for off-the-shelf HBM memory for the standardized pricing model implemented across suppliers like Samsung, Micron, and SK Hynix. However, when ordering quantities so large, like NVIDIA and AMD do, they can request special accommodation features. While this may not be necessarily a compute die that brings TeraFLOPS of power, it will likely be a data processing/logic die that helps route data packets more efficiently, cutting latency and improving performance. Especially during inference, where latency is the most important factor, having a "smarter" HBM could yield sizable, double-digit gains in throughput of tokens.

Samsung Electronics Announces Third Quarter 2025 Results

Samsung Electronics today reported financial results for the third quarter ended Sept. 30, 2025. The Company posted KRW 86.1 trillion in consolidated revenue, an increase of 15.4% compared to the previous quarter. Operating profit increased to KRW 12.2 trillion. The Device Solutions (DS) Division reported a 19% increase in sales quarter-on-quarter (QoQ), with the Memory Business setting an all-time high for quarterly sales, driven by strong growth of HBM3E and server SSDs. Meanwhile, the Device eXperience (DX) Division posted a revenue increase of 11% QoQ due to the successful launch of new foldable phones and solid flagship sales.

Looking ahead to Q4, the rapid growth of the AI industry is expected to open up new market opportunities for both the DS and DX Divisions. The DS Division plans to focus on enhancing its performance by increasing sales of high-value-added memory products tailored to AI. The semiconductor market is expected to remain strong, driven by ongoing AI investment momentum. Meanwhile, the DX Division will strengthen its efforts to launch AI products equipped with the most innovative technologies through open collaborations with leading global partners in respective business segments.

NVIDIA's "Vera Rubin" Superchip System Pictured for the First Time

NVIDIA yesterday had a GTC conference in Washington D.C., showing its latest "Vera Rubin" Superchip symbiote. Pictured for the first time is the combination of two "Rubin" GPU paired with a single "Vera" CPU carrying 88 custom NVIDIA cores and 176 threads in a single package. NVIDIA quoted performance targets of roughly 50 PetaFLOPS of FP4 compute per Rubin GPU, which yields about 100 PetaFLOPS FP4 for the two-GPU Superchip. The company said engineering samples are already moving through labs and set mass production goals for 2026 with broader shipments and deployments into 2027.

Each Rubin GPU appears to integrate two reticle-sized compute chiplets (2x830mm²?) paired with eight HBM4 stacks, delivering about 288 GB of HBM4 per GPU and roughly 576 GB of HBM4 on the full Superchip. NVIDIA also populated the board with SOCAMM2 LPDDR5X modules to provide large, low-latency system memory, with some older briefings indicating around 1.5 TB of LPDDR5X per Vera CPU on typical trays. The Vera CPU itself uses an 88-core, 176-thread Arm-based custom design and shows signs of a multi-chiplet layout with a distinct I/O chiplet nearby. With "Grace," NVIDIA relied on Arm's Neoverse design, but with Vera, the design team brought this CPU core in the house to extract maximum performance. Additionally, NVLink bandwidth climbs to approximately 1.8 TB/s to sustain heavy CPU-to-GPU traffic in system-demanding workloads such as AI inference and training.

Tight DRAM Supply to Boost DDR5 Contract Prices—Profitability in 2026 Expected to Surpass HBM3e

TrendForce's latest investigations show that server DRAM contract prices are strengthening in 4Q25, driven by ongoing data center expansion among global CSPs. This momentum is lifting overall DRAM pricing. Although final contract pricing for the quarter is still being negotiated, suppliers are showing a greater willingness to raise quotes as CSPs increase order volumes.

TrendForce has accordingly revised its 4Q25 outlook for conventional DRAM pricing upward, from an earlier forecast of 8-13% growth to 18-23%, with a strong likelihood of further upward revision.

SK hynix Announces 3Q25 Financial Results

SK hynix Inc. announced today that it has recorded 24.4489 trillion won in revenues, 11.3834 trillion won in operating profit (with an operating margin of 47%), and 12.5975 trillion won in net profit (with a net margin of 52%) in the third quarter. The company achieved its highest-ever quarterly performance, driven by the full-scale rise in prices of DRAM and NAND, as well as the increasing shipments of high-performance products for AI servers. In particular, operating profit exceeded 10 trillion won for the first time in the company's history.

As demand across the memory segment has soared due to customers' expanding investments in AI infrastructure, SK hynix once again surpassed the record-high performance of the previous quarter due to increased sales of high value-added products such as 12-high HBM3E and DDR5 for servers. Driven by surging demand for AI servers, shipments of high-capacity DDR5s of 128 GB or more have more than doubled from the previous quarter. In NAND, the portion of AI server eSSD, which commands a price premium, expanded significantly as well. Building on this strong performance, the company's cash and cash equivalents at the end of the third quarter increased by 10.9 trillion won from the previous quarter, reaching 27.9 trillion won. Meanwhile, interest bearing debt stood at 24.1 trillion won, enabling the company to successfully transition to a net cash position of 3.8 trillion won.

Intel's Advanced Packaging Could Be America's Answer to Silicon Dominance

Intel might be the key to re-realizing the American dream of advanced chip production and packaging on U.S. soil. Under the Trump administration, the United States has been making significant efforts to establish leading-edge chip manufacturing domestically. Achieving this goal is challenging for several reasons. In the race for the most advanced silicon, only a few major veteran players remain: Intel, Samsung, and TSMC. Among these, Taiwan-based TSMC has consistently overcome obstacles to become the leader in the semiconductor industry. Through a strategic approach to semiconductor development, TSMC has excelled in both leading-edge node production and advanced chip packaging, enhancing performance.

Historically, Intel has faced difficulties with leading-edge semiconductor node production, even outsourcing some chip manufacturing to TSMC. However, there is a significant opportunity for Intel to not only produce silicon but also establish itself as a major packaging partner for many manufacturers, including TSMC. TSMC's facilities in Arizona address the issue of USA-based manufacturing only partially. While TSMC's Arizona Fab 21 produces 4 nm wafers, these wafers must be sent back to Taiwan for packaging, disrupting the sovereign supply chain that is crucial for domestic manufacturing. Addressing these issues could present a good opportunity for Intel, even if Team Blue doesn't manufacture the underlying silicon.

NVIDIA "Vera Rubin" NVL144 Servers Set for 2026 Volume Production

Foxconn has begun engineering validation for NVIDIA's "Vera Rubin" NVL144 MGX liquid-cooled racks, with plans for mass production in the latter half of 2026, according to internal supply-chain documents obtained by Taiwan Economics Daily. Each NVL144 unit features 144 Rubin GPUs with HBM4, updated 88-core Vera CPUs, and ConnectX-9 NICs and NVLink 6 interconnect for CPU-to-GPU and GPU-to-GPU communication, effectively doubling the compute density of the current NVL72 platform. The Taiwanese contractor Foxconn currently handles about 60% of NVIDIA's AI server output and is expanding its immersion-cooling and copper-plating lines at its Wisconsin and Houston facilities to meet new domestic manufacturing requirements.

The gap between GB300 peak shipments and Rubin's first silicon will be six to eight months, the shortest generational gap in NVIDIA's history. Despite the accelerated roadmap, Blackwell Ultra remains NVIDIA's main revenue source through mid-2026. Procurement data shows that cloud providers have reserved around 400,000 GB300 nodes for delivery in late 2025 and early 2026, indicating strong near-term demand. At the same time, OpenAI has placed a large preorder for Vera Rubin silicon, securing initial production volumes and supporting Foxconn's capacity investments. The rapid transition from Hopper to Blackwell and then to Rubin within 30 months marks the fastest acceleration of datacenter equipment cycles, leading hyperscale operators to speed up depreciation schedules.

Samsung Targets 13 Gbps HBM4e, Posts $8.5 Billion Q3 Operating Profit

Samsung more details of its next-generation HBM4e memory at the OCP Global Summit aiming for per-pin speeds above 13 Gbps and a total bandwidth of 3.25 TB/s, roughly 2.5× faster than HBM3e. The company also claims more than double the power efficiency, with energy use dropping below 1.95 pJ/bit, less than half that of HBM3e. According to SeDaily and Tweaktown, the development was driven by NVIDIA's request for higher-speed memory to power its upcoming Vera Rubin AI accelerator. While the JEDEC HBM4 standard caps bandwidth at 8 Gbps per pin, NVIDIA has reportedly urged Samsung, SK hynix, and Micron to push beyond 10 Gbps. Chosun Daily reports that Samsung may dissolve its 1c DRAM yield-improvement task force to fast-track HBM4 mass production, even as early cold tests show yields below 50%. The company has reportedly skipped its usual internal production review and moved directly toward ramping production. TrendForce adds that Samsung shifted its HBM4 base die process to a 4 nm FinFET node, positioning itself ahead of competitors in both speed and manufacturing scale. Volume production of 10 Gbps-capable chips is expected to begin by late 2025.

Moreover, on Tuesday, Samsung announced its earnings for the third quarter of 2025, and surprisingly the company's operating profit rose 32% from last year to 12.1 trillion won ($8.5 billion), far exceeding market predictions of about 10 trillion won, according to TrendForce citing Reuters, ZDNet and Chosun Daily. While Samsung didn't share detailed numbers, analysts mentioned in the report estimate the semiconductor division (DS) alone earned around 5 trillion won ($3.5 billion), over 10 times its Q2 profit of 400 billion won (($281 million). A major reason for Samsung's comeback in the third quarter is the big improvement in its foundry division. This part of the business has boosted performance by running at higher capacity and winning new customers. Samsung will reveal full results, including business-by-business breakdowns, on October 30.

Intel Phases Out "Ponte Vecchio" and "Arctic Sound" Data Center GPUs

Intel has apparently started a slow deprecation phase of its "Ponte Vecchio" Data Center GPU Max and "Arctic Sound" Data Center GPU Flex series. According to the changelog in the Intel XPU Manager—a free and open-source tool for monitoring and managing Intel data center GPUs—version 1.3.3 deprecates Data Center GPU Flex Series and Data Center GPU Max Series. Owners of these GPUs are now advised to stick with the older version 1.2.42 for maximum feature support, and not update to the latest release. There are several reasons why this is happening, and a part of the suspicion is that Intel is preparing for "Jaguar Shores" AI accelerator transition, even as its has been only two years since its Data Center GPU Flex/Max series have launched.

Built on the Xe-HPC microarchitecture using Intel's 10 process, Ponte Vecchio is a massive chip, measuring about 1,280 mm² and containing roughly 100 billion transistors. It features 16,384 shading units, 1,024 tensor cores, and has a peak TDP of up to 600 W per card. The Max 1100, 1350, and 1550 SKUs, which offer 48 to 128 GB of memory, once made the design appealing for specialized HPC and AI applications. However, its large scale results in high costs, significant power consumption, and integration challenges. Intel's "Arctic Sound," also based on the Generation 12.5 GPU microarchitecture and manufactured on the Intel 10 node, is a much smaller chip, measuring around 190 mm² with about 8 billion transistors, 8,192 shading units, and 128 ROPs, and has a TDP close to 500 W. The Arctic Sound 1T, 2T, and M SKUs are compact 16 GB designs aimed at midrange data center tasks, with the GPU first appearing in 2021.

NAND Flash Shortages May Last 10 Years, Phison CEO Warns

The massive AI infrastructure buildout is placing a significant strain on the entire semiconductor supply chain, with one often overlooked component being NAND-based storage. According to Phison Electronics CEO Pua Khein-Seng, the NAND flash shortage could last an entire decade. CEO Pua warned that an upcoming NAND flash shortage is estimated to begin in 2026 and could stretch across the next decade. This claim is much higher than anything we have seen, where these shortages last only a year or two. Speaking in a recent interview, Pua said that the combination of surging AI-driven storage demand and limited recent investment in flash capacity has placed the memory market on a collision course with extended turnaround times.

Pua Khein-Seng put much of the blame on an earlier investment hangover. He argued that after prior cycles where heavy spending led to price collapses, many flash makers pulled back, slowing expansion. He noted that starting around 2023, a significant share of capital was redirected into high-bandwidth memory used for model training, as the margins there were more attractive. As a result, NAND received less focus just when demand began rising again. He also emphasized a shift inside cloud operations from training to inference, adding that inference workloads and model storage create large and persistent requirements for NAND flash. As one of the major controller designers for SSDs, the Phison CEO knows precisely where the demand is headed, many quarters ahead of the public.

Raspberry Pi Increases Prices for Certain 4 GB/8 GB Products by $5 to $10

As many of you are aware, memory prices have been rising rapidly for the last six months. Insatiable demand for High-Bandwidth Memory (HBM) for AI applications is competing for fab space with the commodity LPDDR memory used by Raspberry Pi, leading to shortages and price rises. At this point, memory costs roughly 120% more than it did a year ago. We came into this year holding substantial stockpiles of memory, which has allowed us to hold prices flat. However, we've now reached the point where we have to pass some of this cost on. With effect from today, we're making the following changes to pricing:
  • 4 GB Compute Module 4 and Compute Module 5 variants increase by $5
  • 8 GB Compute Module 4 and Compute Module 5 variants increase by $10
  • Raspberry Pi 500 unit-only increases by $10 to $100
  • Raspberry Pi Development Kit for Compute Module 5 increases by $5 to $135

Tight Upstream Supply and Restocking Drive 2024 DRAM Module Revenue Growth of 7%

TrendForce reports that following the completion of inventory digestion in the downstream consumer market concluded at the end of Q4 2023, DRAM suppliers shifted focus towards HBM and server DDR5 products, leading to tighter supply for other DRAM types. This increase in demand pushed overall DRAM prices higher, encouraging module manufacturers to replenish stocks and boost procurement. Consequently, the global DRAM module market achieved USD 13.3 billion in revenue in 2024, a 7% YoY increase that reversed the 28% decline experienced in 2023.

In the latter half of 2024, demand momentum declined due to rising module prices. Module manufacturers struggled to transfer these high chip costs to distribution channels, limiting sales. To stay competitive, some aimed to enhance costs structure by increasing their purchase of cheaper server reball chips.

ASML Schedules Delivery of 56 Low-NA and 10 High-NA EUV Tools in 2027

ASML is about to manufacture significantly more Low-NA EUV and High-NA EUV machines than it originally anticipated. As you may be aware, acquiring an EUV scanner is a year-long process, during which companies make commitments that can span several years to procure these tools. According to ASML, the company estimates that in 2027, based on current order information and demand, it will deliver 56 Low-NA EUV scanners. This includes five units for Intel, seven units for Samsung, and as many as 20 units for SK Hynix. According to South Korean media, SK Hynix plans to install 20 Low-NA EUV units in the next two years, all designed for HBM memory and advanced storage solutions.

For High-NA, ASML plans to deliver 10 High-NA EUV scanners, each costing around $380 million. Intel will add to its facilities manufacturing the 14A node, and SK Hynix is expected to use two of those as well for its memory operations. Intel is betting big on its next-generation 14A node, adding additional EUV tools to its already existing High-NA portfolio. In the past, Intel disclosed that it processed over 30,000 wafers within a single quarter while achieving simplified manufacturing by cutting the required steps for a particular layer from 40 to fewer than 10, which naturally led to much faster cycle times. Similarly, Samsung has seen impressive results in its operations, documenting a 60% reduction in cycle time for one of its specific applications.
Return to Keyword Browsing
Nov 17th, 2026 13:49 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts