News Posts matching #HBM

Return to Keyword Browsing

Micron Breaks Ground on Advanced Wafer Fabrication Facility in Singapore

Micron Technology, Inc. broke ground today on an advanced wafer fabrication facility located within the company's existing NAND manufacturing complex in Singapore. This new facility represents a planned investment of approximately US $24 billion (SG $31 billion) over 10 years and is designed to ultimately provide 700,000 square feet of cleanroom space. Wafer output is scheduled to begin in the second half of calendar 2028, helping Micron address growing market demand for NAND technology driven by the rapid expansion of AI and data-centric applications.

The groundbreaking ceremony for this facility, Singapore's first double-story wafer manufacturing fab, was marked by the attendance of Gan Kim Yong, deputy prime minister and minister for Trade and Industry of Singapore, Dr Beh Swan Gin, permanent secretary of the Ministry of Trade and Industry, Jermaine Loy, managing director of the Singapore Economic Development Board (EDB) and Jacqueline Poh, CEO of JTC Corporation.

Microsoft Introduces Its Newest AI Accelerator: Maia 200

Today, we're proud to introduce Maia 200, a breakthrough inference accelerator engineered to dramatically improve the economics of AI token generation. Maia 200 is an AI inference powerhouse: an accelerator built on TSMC's 3 nm process with native FP8/FP4 tensor cores, a redesigned memory system with 216 GB HBM3e at 7 TB/s and 272 MB of on-chip SRAM, plus data movement engines that keep massive models fed, fast and highly utilized. This makes Maia 200 the most performant, first-party silicon from any hyperscaler, with three times the FP4 performance of the third generation Amazon Trainium, and FP8 performance above Google's seventh generation TPU. Maia 200 is also the most efficient inference system Microsoft has ever deployed, with 30% better performance per dollar than the latest generation hardware in our fleet today.

Maia 200 is part of our heterogenous AI infrastructure and will serve multiple models, including the latest GPT-5.2 models from OpenAI, bringing performance per dollar advantage to Microsoft Foundry and Microsoft 365 Copilot. The Microsoft Superintelligence team will use Maia 200 for synthetic data generation and reinforcement learning to improve next-generation in-house models. For synthetic data pipeline use cases, Maia 200's unique design helps accelerate the rate at which high-quality, domain-specific data can be generated and filtered, feeding downstream training with fresher, more targeted signals.

Giga Computing Announces New Server Built on NVIDIA GB200 NVL4 Platform at SCA/HPC Asia 2026

Giga Computing, a subsidiary of GIGABYTE and a leader in high-performance computing and data center solutions, today announced the launch of its next-generation AI and HPC server, GIGABYTE XN24-VC0-LA61, powered by the NVIDIA GB200 NVL4 platform. Purpose-built for its heterogeneous architecture of CPU and GPU, the system will make its debut at SCA/HPC Asia 2026 in Osaka, Japan, showcasing Giga Computing's latest innovation in accelerated computing with liquid cooling technologies.

The XN24 server has been selected for the RIKEN Center for Computational Science (R-CCS) next-generation HPC-Quantum hybrid platform. The new platform integrates GIGABYTE servers into the development of FugakuNEXT, the eventual successor to Japan's flagship supercomputer, Fugaku. This integration will use the NVIDIA CUDA-Q platform to support the development of hybrid quantum-GPU supercomputing systems and research into advanced scientific applications that bridge quantum and traditional high-performance computing.

Samsung Memory Distributor Reportedly Hikes Prices by 80%

Everyone is aware of the massive memory shortage the industry is currently experiencing, yet prices are reportedly still rising. According to an anonymous Samsung memory distributor, an immediate price hike of 80% has been applied. Later, an employee from the Device Solutions Campus in Giheung, located south of Seoul, confirmed this information. However, we still lack official confirmation from other third-party sources. If the 80% hike is accurate, it will likely apply to the latest batch of contract prices Samsung has imposed on its distributors. As new contract negotiations occur every quarter, it seems the 80% price increase is applied for Q1, given that the effect is noted as immediate.

The scope of this adjustment includes all Samsung memory products, which encompass a diverse range of offerings such as DDR, HBM, LPDDR, and GDDR solutions. These products are used in mobile devices, desktops, embedded systems, graphics cards, AI accelerators, and more. However, since some memory products are supplied directly to Samsung's customers without distributors, it is likely that only certain types of these products will experience a significant price increase, if the reports are accurate. We still have a few weeks left before the trickle-down economics of the supply chain start to manifest, so we will know for sure if these rumors are true.

Update 18:36 UTC: : According to UDN, Samsung has denied these price hikes, claiming that the company has not issued a comprehensive price increase across its memory products.

Samsung Reportedly Developing 2 nm Custom HBM Logic Dies

This afternoon, a ZDNet South Korea news article has disclosed early details of another intriguing 2 nm project reportedly underway at Samsung Semiconductor. Specific customers are not mentioned by industry insiders, but the South Korean megacorporation's HBM (High Bandwidth Memory) development team could be tailoring custom logic dies for individual product requirements. There are rumors circulating about the computer memory R&D department selecting "foundry process nodes as advanced as 2 nm" for next-wave HBM designs. The ZDNet piece does not define whether engineers have opted into using the in-house foundry's SF2 or SF2P nodes.

Samsung's sixth-gen HBM product line (aka HBM4) is said to be based on a 4 nm process node, likely picked out from the SF4 family. According to an anonymous company insider: "Samsung Electronics is designing a custom logic die for HBM under the leadership of the custom SoC (system-on-chip) team newly established within the System LSI Business Division last year...We are building a portfolio ranging from 4 nm to 2 nm to respond to the needs of various customers." ZDNet adds that next-gen ultra-high-performance AI accelerators are expected to rely on the very best futuristic HBM modules. "Strong" AI enterprise industry demand is anticipated once 2 nm logic dies come into fruition, possibly post-2027, after the arrival of HBM4E (seventh gen).

NVIDIA Upgrades "Vera Rubin" Memory Bandwidth Ahead of AMD Instinct MI400 Launch

As we anticipate the first shipments of NVIDIA's "Vera Rubin" in late summer, it has been reported that NVIDIA upgraded its superchip multiple times since March to better compete with AMD's upcoming Instinct MI400 series of accelerators. According to SemiAnalysis, NVIDIA's initial target for the "Vera Rubin" VR200 NVL72 system was 13 TB/s in March, which was upgraded to 20.5 TB/s by September. However, at CES 2026, NVIDIA confirmed that the VR200 NVL72 system is now operating at 22 TB/s of bandwidth. Compared to AMD's Instinct MI455X accelerator, which has 19.6 TB/s, NVIDIA initially had inferior system bandwidth. They addressed this by using faster DRAM and improving interconnects between CPUs, GPUs, and the entire system.

In November, AMD compared the MI400 lineup to NVIDIA's upcoming "Vera Rubin" series, claiming similar compute performance and memory bandwidth but 1.5 times higher memory capacity and scale-out bandwidth. AMD claims it will deliver up to 40 FP4 and 20 FP8 PFLOPs, roughly twice the compute performance of the current MI350. The GPUs also transition to HBM4 memory from HBM3e, increasing capacity from 288 GB to 432 GB and raising total bandwidth from 8 TB/s to 19.6 TB/s. Each GPU provides 300 GB/s of scale-out bandwidth and adds broader AI data format support along with expanded AI pipelines. Two models are planned: the Instinct MI455X for large-scale AI training and inference, and the MI430X for HPC and AI.

Global AI Server Shipments Forecast to Grow Over 28% YoY in 2026, with a Rising Share of ASIC-Based Systems

North American CSPs' continued investments in AI infrastructure are expected to increase global AI server shipments by more than 28% YoY in 2026, according to the latest market research from TrendForce. The rapid growth of AI inference services is boosting demand for general-purpose servers, supporting both replacement and expansion efforts. Consequently, TrendForce predicts that total global server shipments, including AI servers, will accelerate from 2025, with a 12.8% YoY growth in 2026.

The server market from 2024 to 2025 primarily centered on training advanced large language models (LLMs), utilizing AI servers with GPUs and HBM for parallel computing. However, starting in the second half of 2025, the growth of AI agents, LLaMA-based applications, and Copilot upgrades shifted CSP's focus toward inference services as a monetization strategy. AI inference workloads are now deployed not only on dedicated AI server racks but also on general-purpose servers that handle pre- and post-inference computing and storage.

Tesla Completes AI5 Chip, Revives Dojo 3 with Intel Packaging Deal

Tesla has finalized the development of its AI5 automotive processor, allowing for production at TSMC and Samsung facilities. This advancement also revives the previously halted Dojo 3 supercomputer project. CEO Elon Musk announced that the AI5 chip will offer performance on par with NVIDIA's "Hopper" architecture, with two AI5 units equaling the power of a single "Blackwell" processor. The revival of Dojo 3 project was rumored with Intel joining as a key packaging partner, marking a shift from Tesla's earlier dependence on TSMC for complete manufacturing. Intel will manage assembly and testing using its EMIB technology, which connects multiple dies via silicon bridges instead of a full-wafer interposer. This method is better suited to Tesla's large Dojo modules that combine several 654 mm² chips into a single package.

Also based on previous information, Samsung will produce the D3 training chips at its Texas facility using a 2 nm process, leaving Intel to oversee packaging operations exclusively. This division of labor addresses capacity constraints and provides Tesla with more flexibility in customizing interconnect layouts. For the AI5 automotive chips, both Samsung and TSMC will create distinct versions, but Tesla aims to ensure identical software performance across both. Initial projections suggest the AI5 could operate at just 150 W while matching the performance of NVIDIA's H100 accelerator that requires 700 W. This efficiency is achieved by removing general-purpose graphics subsystems and optimizing the architecture specifically for Tesla's neural processing.

SK Hynix Shows 16-Layer 48 GB HBM4 Memory Modules at CES 2026

SK Hynix has showcased its most advanced high-bandwidth memory, packaged as a 16-high 48 GB HBM4, designed for high-capacity AI and HPC accelerators that require lots of memory. This is the first time SK Hynix has shown a memory module more advanced than the 12-high 36 GB HBM4, which used to run at 11.7 Gbps. For the new 16-high HBM4, we don't have official data or any speed references, but we can assume SK Hynix managed to extract a bit more bandwidth from the four additional DRAM layers. Interestingly, the announcement was somewhat vague, possibly because Micron and Samsung are also enhancing their HBM4 modules before they become available to companies like AMD and NVIDIA, allowing for potential increases in speed if needed.

The company also demonstrated the custom base die HBM called cHBM. This includes a customized base die sitting at the bottom of the DRAM stack that includes some logic usually associated with the GPU/ASIC logic die, and not memory. For example, this could include HBM PHY and memory controllers that are usually embedded on the GPU die and occupy space, which could be used for more compute logic on the die and maximize performance. In SK Hynix's demonstration and illustration, this is manifested as die-to-die PHYs, embedded memory controllers, HBM PHY, and much more. Customers can basically choose whatever they want and can even embed processing logic inside the die area.

Memory Makers Prioritize Server Applications, Driving Across-the-Board Price Increases in 1Q26

TrendForce's latest investigations indicate that DRAM suppliers in 1Q26 will continue to reallocate advanced process nodes and new capacity toward server and HBM products to support rising AI server demand. This shift has significantly limited supply in other markets, causing conventional DRAM contract prices to increase by about 55-60% QoQ. Meanwhile, disciplined capacity management by NAND Flash suppliers, along with robust server demand that is displacing other applications, is likely to increase contract prices across all NAND Flash product categories by 33-38%.

Despite weaker notebook shipments and slower growth in memory demand due to potential specification downgrades, PC DRAM prices are still set to rise sharply in the first quarter. DRAM suppliers have simultaneously tightened supply to PC OEMs and module makers, forcing some OEMs to procure memory at higher prices through modules makers. This is expected to raise DRAM suppliers' module pricing and significantly drive up PC DRAM prices.

Intel Envisions 16 Compute Dies and 24 HBM5 Modules in a Single Package

Intel Foundry has published a short video demonstration of its advanced packaging technologies, demoing silicon scaling beyond reticle limits of 830 mm². According to the video, Intel's Foveros 3D and EMIB-T interconnect can scale silicon to 12x reticle size, packaging up to 16 compute dies paired with 24 HBM5 memory modules in a single package. All of this will leverage Intel's 18A—including 18A-P and 18A-PT—and 14A nodes that the company is preparing for mass production and external customers.

The approach involves a sophisticated layering technique where base dies, manufactured using the 18A-PT process, incorporate backside power delivery to enhance logic density and reliability. These base dies house SRAM structures similar to Intel's "Clearwater Forest" architecture. They serve as foundations for compute tiles built on the advanced 14A/14A-E nodes, which feature second-generation RibbonFET transistors and PowerDirect technology. Foveros Direct 3D enables vertical stacking through ultra-fine pitch hybrid bonding, while EMIB-T incorporates through-silicon vias to facilitate higher bandwidth connections between chiplets. This combination achieves scalability beyond reticle limits and will support all HBM standard, including HBM4, HBM5, and future versions.

Micron Forecasts DRAM Shortages Beyond 2026

Despite achieving a record $13.64 billion in revenue for the first quarter of fiscal 2026, which ended on November 27, 2025, Micron addressed the ongoing DRAM shortage in its recent earnings call, noting that it is not a short-term issue. The company stated, "Sustained and strong industry demand, along with supply constraints, are contributing to tight market conditions, and we expect these conditions to persist beyond calendar 2026." Consequently, major customers are securing supply through multi-year contracts, driving up spot prices in both consumer and enterprise memory segments.

A significant factor in this imbalance is the increased demand for HBM memory. Micron mentioned that it "has completed agreements on price and volume for our entire calendar 2026 HBM supply," indicating that HBM customers are securing capacity well in advance. HBM requires more wafer area than standard DDR5 products as its stacks up to 12 DRAM dies, so its usage shifts scarce resources towards data center and AI workloads, reducing availability for traditional PC volumes. This results in a reallocation of limited dies and wafers to buyers willing to pay premiums and commit to long-term agreements, which are still a thing despite limited supply.

Rapidus Explores Glass Panel-Level Packaging to Rival TSMC

It seems that the Japanese foundry Rapidus is gearing up for strong competition with TSMC. The company has developed a panel-level packaging (PLP) prototype using a glass interposer, with plans to move towards mass production by 2028. At this week's SEMICON Japan 2025 in Tokyo, the prototype will be showcased. It replaces conventional round silicon wafers with large square glass panels, intended for use in high-performance AI packages that combine multi-chip GPUs and more than a dozen of HBM dies. This approach could reduce material waste and enable larger, denser multi-chip assemblies necessary for next-generation AI accelerators, without breaking the traditional laws of 300 mm wafer.

The company is developing 600x600 mm glass panels, which offer a significantly larger usable area than 300 mm silicon wafers and enable the production of more interposers from a single sheet. Glass substrates provide advantages in electrical performance and flatness compared to organic package materials, making them appealing for dense interconnects. However, glass also introduces new manufacturing challenges, such as fragility and potential warpage as panel sizes increase. Intel has notably discontinued its in-house development of glass substrates and has chosen to license the technology to others. The company can thus monetize its innovations, though large-scale Intel-based technology production is not anticipated until after 2030.

SEMI Reports Global Semiconductor Equipment Sales to Reach $156 Billion in 2027

Global sales of total semiconductor manufacturing equipment by original equipment manufacturers (OEMs) are forecast to reach a record high of $133 billion in 2025, growing 13.7% year-on-year, SEMI announced today in its Year-End Total Semiconductor Equipment Forecast - OEM Perspective at SEMICON Japan 2025. Growth in semiconductor manufacturing equipment sales is expected to continue in the two following years of the forecast period, with projections of $145 billion in 2026 and $156 billion in 2027. This growth will be driven primarily by investments related to AI, particularly in leading-edge logic, memory, and the adoption of advanced packaging technologies.

"Global semiconductor equipment sales show robust momentum, with both the front-end and back-end segments projected to see three consecutive years of growth, culminating in total sales surpassing $150 billion for the first time in 2027," said Ajit Manocha, SEMI president and CEO. "Investments to support AI demand have been stronger than anticipated since our midyear forecast, leading us to boost the outlook for all segments."

Samsung Denies NAND Flash Exit as Sapphire PR Manager Calls for Calm

The gaming hardware industry has been in a bit of a state of late as a result of increased demand for DRAM causing a supply shortage and massive price hikes. As a result of the aforementioned issues, Samsung has been rumored to be converting some of its HBM3E and NAND production capacity to DRAM in order to meet demand. The ensuing rumors claimed that Samsung was planning an exit from the NAND, and thus SATA SSD market, as a result of the shift in focus. Micron's recent exit from the consumer space, and the closure of its Crucial memory and SSD brand, lent credence to these rumors, however, these rumors have recently been addressed by a Samsung spokesperson directly in a response to Wccftech. The spokesperson simply said "The rumor regarding the phasing out of Samsung SATA or other SSDs is false," apparently declining to expand any further. However, this is just one indication that the DRAM crisis may not be as long-lived as some have claimed.

Around the same time, in an interview with Hardware Unboxed, Edward Crisler, the PR manager for Sapphire, cautioned PC gamers and potential buyers against panic buying, saying that, "the good news is, I don't think the real pain, that we're suffering now and for the next six months or so, is going to last much longer than that," although he goes on to say that the actual issue at hand is uncertainty of the situation. The implication is that the market will eventually begin to stabilize within the next six to eight months. The implication isn't necessarily that DRAM prices will return to normal, but rather that DRAM supply will eventually catch up to supply. It could also be the case that the massive AI and datacenter boom currently causing the shortages will slow in the coming months, which would also help to stabilize things somewhat. Crisler is careful to note that DRAM prices may still remain somewhat elevated after the market stabilizes, but he seems to be convinced that the sky-high prices we're seeing for consumer memory will fall to some degree and that the gaming industry will adapt to whatever the end result is of the market shake-up.

Intel to Acquire SambaNova AI Chip Startup

According to sources close to Wired, Intel has signed a new agreement with SambaNova, an AI chip startup, to acquire the company under undisclosed terms. The signed term sheet is non-binding, meaning it can be canceled under any condition, and a final deal has not yet been reached. Intel's current CEO, Lip-Bu Tan, has been serving as chairman of the board for SambaNova Systems, effectively giving him insider knowledge of the company's operations. The company has raised a total of $1.14 billion in funding as of early 2025, meaning that a lot of investors have shown confidence into its technology.

SambaNova develops AI chips based on its Reconfigurable Dataflow Unit (RDU), which takes a novel approach to enable trillion-parameter inference on its SambaRack rack-scale solutions. The latest 4th generation RDU features 1,040 RDU cores, 653 TeraFLOPS of BF16 (similar to FP16) compute, 520 MB of on-chip memory, 64 GB of HBM3, and is paired with a 1.5 TB external DDR memory pool to accommodate large language models for inference. Since it is a private company, we don't have sales figures officially published anywhere.

December 13, 20:35 UTC: Bloomberg reports that the deal is worth about $1.6 billion, including debt. Additionally, the deal is reportedly nearing its final stages.

SK Hynix Forecasts Tight Memory Supply Lasting Through 2028

SK Hynix held an internal company meeting, where the company reportedly presented some tough pills to swallow for many gamers. The company now forecasts the tight memory supply to last through 2028 for the commodity DRAM, which includes DDR5/DDR4, GDDR6/GDDR7, and LPDDR5x/LPDDR6. All the aforementioned DRAM variants are essential for PC and console components, making millions of gamers at risk of absorbing a massive price increase as a result. As memory supplier inventories deplete, production capacity will not increase to meet demand as it has in the past. This marks a departure from the usual response by memory manufacturers, who typically ramp up capacity in response to demand.

However, an interesting thing that SK Hynix has noted is that this situation will exclude the company's most advanced memory solutions—HBM and SOCAMM. These products are expected to get an additional capacity expansion, without tight supply impacting this part of the DRAM supply chain. These products are even in high demand as they are mostly consumed in higher volume than the regular commodity DRAM, due to their integration into AI products such as GPUs and servers.

Huawei Ascend 950 AI Accelerator Pictured

Huawei's next-generation Ascend 950 AI accelerator has been pictured for the first time, showcasing the company's custom silicon and HBM memory. The chip combines Huawei's first self-developed HBM memory with a new generation of AI acceleration. Huawei aims to compete through scale rather than focusing solely on single-chip performance. Although it currently trails NVIDIA in per-chip performance, Huawei's system-scale solutions can still be competitive. The company has announced the Ascend 950 family for early 2026, featuring two variants.

The 950PR model includes 128 GB of in-house HBM with around 1.6 TB/s of bandwidth, while the 950DT model increases memory to 144 GB and boosts bandwidth to nearly 4 TB/s. Both chips target one PetaFLOP of FP8, and two PetaFLOPS of FP4. Huawei's competitive strategy emphasizes dense packaging and aggressive networking, rather than relying solely on raw per-chip performance. We have no information on the node selection, but it will likely be SMIC's newest N+3 node with 5 nm-class features. Chinese SMIC has officially achieved volume production of its newest 5 nm node relying on the deep ultraviolet (DUV) to manufacture its silicon. Since the first customer for N+3 was Huawei with Kirin 9030 SoC, it is only logical that the more important Ascend AI accelerator family is manufactured using the same node.

JEDEC Prepares SPHBM4 Standard to Deliver HBM4-Level Throughput with Reduced Pin Count

JEDEC Solid State Technology Association, the global leader in standards development for the microelectronics industry, today announced it is nearing completion of a new standard for Standard Package High Bandwidth Memory (SPHBM4). SPHBM4 devices are similar to the HBM4 devices commonly used in artificial intelligence accelerators, using the same DRAM dies on a new interface base die which can be mounted on standard organic substrates. In contrast, HBM4 is typically mounted on silicon substrates.

As planned, SPHBM operates at the same aggregate data throughput as HBM4 using fewer pins by operating at a higher frequency. Where the HBM4 interface has 2048 data signals, when published, SPHBM4 will define 512 data signals with 4:1 serialization to achieve the same bandwidth. This change allows the relaxed bump pitch required for connection to organic substrates.

Semiconductor Quarterly Revenue Surpasses $200bn for the First Time as Industry-Wide Growth Accelerates

New research from Omdia shows that the semiconductor market delivered a record breaking performance in 3Q25 with industry revenue reaching $216.3bn, up 14.5% quarter-over-quarter (QoQ). This marks the first time the global semiconductor market has exceeded $200bn in a single quarter, following an already strong 2Q25, which saw 8% QoQ growth. At this pace, the industry is on track to exceed $800bn in total revenue for 2025.

3Q25 far exceeds seasonal expectations
Demand for AI and memory products remains strong, with both segments continuing to outpace the wider market. While AI has dominated recent industry narratives, Omdia notes that more segments contributed to growth in 3Q25 than in previous quarters.

NVIDIA Gets H200 Export License for China, With 25% of Revenue Going to the U.S. Government

The Trump administration has approved a framework allowing NVIDIA to resume exports of its H200 data center accelerators to China, subject to a revenue-sharing condition that allocates 25% of sales proceeds to the U.S. government. Rather than lifting export restrictions outright, the arrangement requires that recipients be vetted through Commerce Department licensing, so shipments will be controlled and limited to approved commercial customers. The decision reopens a major market for high-end accelerators while maintaining a regulatory channel to monitor distribution.

The H200 chip remains a powerful solution for large-scale AI workloads even today. Manufactured using TSMC's N4 process, it combines NVIDIA's "Hopper"-derived architecture with 141 GB of HBM3E memory and 4.8 TB/s memory bandwidth, making it ideal for large model training and dense inference tasks. Although it doesn't match the performance of NVIDIA's current "Blackwell" family and other upcoming accelerators, the H200 is a mature product that has been shipping since spring 2024 and benefits from a well-developed driver and software ecosystem. Its raw compute power is roughly double that of solutions like the Huawei Ascend 910C, making it very competitive even today.

Imec 3D HBM-on-GPU STCO Study Shows Major Thermal Improvements

This week, at the 2025 IEEE International Electron Devices Meeting (IEDM), imec, world-leading research center in advanced semiconductor technologies, presents the first thermal system-technology co-optimization (STCO) study of 3D HBM-on-GPU (high-bandwidth memory on graphical processing unit), a promising compute architecture for next-gen AI applications. By combining technology and system-level mitigation strategies, peak GPU temperatures could be reduced from 140.7°C to 70.8°C under realistic AI training workloads - on par with current 2.5D integration options. The result demonstrates the strength of combining cross-layer optimization (i.e., co-optimizing the knobs at all the different abstraction layers) with broad technological expertise, a combination that is unique to imec.

Integrating high bandwidth memory (HBM) stacks directly on top of graphical processing units (GPUs) offers an attractive approach for building next-gen compute architectures for data-intensive AI workloads. This 3D HBM-on-GPU promises a huge leap forward in compute density (with four GPUs per package), memory per GPU, and GPU-memory bandwidth compared to current 2.5D integration options where HBM stacks are placed around (one or two) GPUs on a silicon interposer. However, the aggressive 3D integration approach is prone to thermal issues because of higher local power density and vertical thermal resistance.

SK Hynix Slows Down HBM4 Ramp, Prepares 300+ Layer NAND Flash

Interesting developments are unfolding in the world of DRAM and NAND flash, as one of the largest players, SK Hynix, has announced product updates for 2026. The company appears to be following its competitor Samsung by slowing the ramp-up of its next-generation HBM4. Initially, SK Hynix planned to increase its HBM4 capacity in the second quarter of 2026. However, this has been postponed to the third quarter of 2026, as the company plans to expand its HBM4 manufacturing capacity. The delay is due to the high demand for its current HBM3E capacity, necessitating the continued operation of HBM3E manufacturing sites longer than expected. SK Hynix's largest customer for HBM chips, NVIDIA, is reportedly facing challenges with the volume production of "Rubin," delaying the need for HBM4, while the HBM3E-equipped "Blackwell" remains in strong demand.

Samsung Could Convert Some HBM3E Capacity to Regular DRAM to Meet Demand

As Samsung transitions from its HBM3 and HBM3E variations of HBM memory to HBM4, the company is reportedly reallocating some of its HBM production capacity to regular DRAM to meet demand. Discussions within the company focus on converting approximately 30-40% of the current capacity from the 10 nm-class, fourth-generation 1a DRAM process to the 10 nm-class, fifth-generation 1b lines used for general-purpose DRAM. This includes DDR5, LPDDR5X, LPDDR6, and GDDR7 memory. With additional investment to repurpose mature process nodes like 1z, Samsung expects to free up the equivalent of about 80,000 wafers per month of 1b capacity.

While HBM typically offers higher margins for companies like Micron and SK Hynix, Samsung's internal estimates show that the operating profit for a 12-hi HBM3E stack is around 30%. In contrast, the near-term expectations for general-purpose DRAM exceed 60%. These margin differences, along with forecasts of a significant decline in average selling prices for 12-hi HBM3E and limited incremental demand from the largest HBM customer, have reduced the incentive to expand HBM3E production.

Samsung HBM4 Development Complete, Mass Production Pending NVIDIA Approval

Samsung has reportedly finished development of its first sixth-generation HBM (HBM4) chips and is now in the final pre-mass-production stage. According to new reporting from AjuNews and Korean industry sources, Samsung's Device Solutions division has cleared Production Readiness Approval (PRA), the last internal checkpoint before full-volume manufacturing. HBM4 is expected to deliver roughly 60% higher bandwidth than the current HBM3E chips. Samsung has already shipped HBM4 samples to NVIDIA, which is evaluating them for use in its upcoming Rubin platform. Initial samples have already exceeded NVIDIA's next-gen GPU requirement of 11 Gbps per pin. The company is using improved 1c-class DRAM paired with a 4 nm logic base die, a combination that helps close the gap with rivals while keeping thermals and power in check at these higher speeds. Once NVIDIA signs off, Samsung can move straight into volume production, with manufacturing lines already prepared.

Samsung had already signaled strong progress during its Q3 2025 earnings call on October 30, noting that HBM4 samples were in the hands of global customers and that mass production is planned for 2026. "HBM3E is currently in mass production and being sold to all related customers, while HBM4 samples are simultaneously being shipped to key clients," the company said at the time. Samsung also confirmed that its foundry arm will prioritize stable 2 nm GAA output and HBM4 base-die production in 2026, alongside the ramp of its new Taylor, Texas fab. The company is also said to be working on a faster HBM4 variant targeting another 40% performance uplift, with an announcement possible as early as mid-February 2026.
Return to Keyword Browsing
Nov 17th, 2026 13:49 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts