HBM
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 49
- Pat Gelsinger Backs Nagare Labs’ Z2RAM Research
Former Intel CEO Pat Gelsinger is supporting Nagare Labs and its Z2RAM memory technology, according to TechNews and Wccftech. The laboratory-stage technology is designed to combine SRAM-like speed, DRAM-like density and Flash-like scalability. Z2RAM uses an oxide semiconductor and a two-transistor cell, and Nagare Labs is researching vertical stacking and direct integration with logic chips through 3D hybrid bonding. The company says its first generation could provide more than five times the bit density of SRAM, with comparable bandwidth. Nagare Labs has raised a $5 million seed round and is seeking research and manufacturing partners. The technology remains a laboratory prototype, and no standards yet exist for integrating this type of memory with accelerators.
科技新报 · 🔥 0 - Bank of Korea official sees higher chip output on HBM demand
36Kr, citing First Financial, reported that a Bank of Korea official expects chip production to rise next year as HBM demand grows. The official also said chip prices would continue increasing, although the rate of increase would slow.
36氪快讯 · 🔥 0 - RTX 5090 Discontinuation Rumor Fuels Huaqiangbei Price Surge
Nvidia denied reports that it had discontinued the GeForce RTX 5090, yet the rumor triggered rapid price changes in China’s Huaqiangbei GPU market. IT之家 reported that RTX 5090 channel prices rose from roughly 30,000 yuan to around 50,000 yuan or as high as 55,000–60,000 yuan, while RTX 5080 and RTX 5060 prices also increased. The report attributed part of the supply pressure to memory makers prioritizing HBM for AI servers, alongside a broader shortage of GDDR7 memory for consumer GPUs.
IT之家 AI · 🔥 0 - ITIF urges capacity expansion for AI-driven memory shortage
The Information Technology and Innovation Foundation argues that AI data-center demand is shifting memory production toward HBM, squeezing DRAM and NAND supply and exposing a structural mismatch with existing capacity. A translation published by Digital Economy Development Review says the response should prioritize long-term capacity expansion, faster substitute-chip certification and R&D rather than tariffs, preferential allocation or broad export controls. The analysis also considers temporary licensing for Chinese memory suppliers as a way to balance procurement costs against long-term dependency.
虎嗅 AI · 🔥 0 - RTX 5080 and RTX 5070 Ti prices rise in China amid tighter GDDR7 supply
NVIDIA partners and retailers in China raised prices for the GeForce RTX 5080 and RTX 5070 Ti 16 GB, according to Wccftech. Board Channels reported that the RTX 5080 16 GB moved from roughly 12,700–13,500 RMB to 14,500–15,000 RMB, while the RTX 5070 Ti 16 GB rose from about 9,000–9,150 RMB to 9,650–9,850 RMB. Wccftech attributed the increases to tighter GDDR7 supply as memory manufacturers prioritize higher-margin HBM for AI markets.
Wccftech · 🔥 0 - Nvidia's Huang sees no memory glut by 2029
Nvidia CEO Jensen Huang told CNBC that the memory-chip industry is unlikely to face a supply glut by 2029, arguing that users can always consume more memory and that HBM will remain widely adopted. He also said Nvidia was the first company to bring smartphone LPDDR memory into data centers, enabling performance with twice the energy efficiency of any product available at the time, though no benchmark methodology or independent measurements were provided. IT之家 AI reported that experts cited by Maeil Business instead expect new capacity to intensify competition around 2029.
IT之家 AI · 🔥 0 - Jensen Huang Sees No Memory Supply Glut by 2029
Nvidia CEO Jensen Huang told CNBC that a memory-supply surplus is unlikely by 2029, arguing that AI infrastructure demand could sustain continued memory consumption. Reported follow-up coverage says HBM will remain in use for years and Nvidia will keep using large amounts of LPDDR, which the company first brought from phones into data centers. TechNews also reported that Huang remains open to working with Chinese memory suppliers such as CXMT, while noting that his 2029 outlook is an executive forecast rather than an independently verified market forecast.
CNBC Technology · 🔥 0 - Powerchip Semiconductor Q3 profit nearly doubles as 3D AI lines resume
Powerchip Semiconductor reported third-quarter revenue of NT$21.3 billion, up 23% sequentially and 80% year over year, while net profit rose 98% to NT$6.5 billion and gross margin reached 48%. Its 3D AI foundry revenue share fell from 5.4% to 1.4%, partly because of equipment relocation and a production pause at the Tongluo plant. The company said related production lines had resumed in Hsinchu, with the Wafer-on-Wafer line in small-volume trial production; it expects 3D AI foundry revenue momentum to return in the fourth quarter. Powerchip forecasts double-digit fourth-quarter revenue growth and plans to raise automotive-process foundry prices by 10% to 20% for 2027 while negotiating with customers. These are company forecasts and plans rather than completed results.
科技新报 · 🔥 0 - Samsung said to plan 2nm M2800 5G modem, freeing 4nm capacity for HBM base dies
TechNews, citing THEELEC and industry sources, reports that Samsung's System LSI Business has decided to build its next external 5G modem, M2800, on a 2nm process; the chip is slated for completion in 2027 and is expected to pair with Exynos 2800 in a dual-chip setup in the 2028 Galaxy S28 series. The report says the move frees 4nm capacity that Samsung Foundry is redirecting toward HBM base dies, the logic-bearing layer at the bottom of an HBM stack, while the M2700 modem for the 2027 Galaxy S27 stays at 4nm and the old Shannon brand has been retired in favor of M-series naming. cnBeta's analysis adds that Samsung is validating its own 2nm process for Nvidia custom NVHBM base dies, which suppliers have reportedly been asked to build on TSMC 3nm, and says there is no evidence Qualcomm has decided to place flagship Snapdragon orders on Samsung 2nm. TechNews also cites Samsung MX internal tests praising the 4nm M2700's efficiency and data speeds over the Shannon 5410, a vendor claim.
科技新报 · 🔥 0 - TrendForce sees Q4 2026 DRAM contract prices up about 20% on AI inference demand
TrendForce estimates that DRAM contract prices will rise about 20% quarter-over-quarter in the fourth quarter of 2026, according to IT之家 and a 36氪 brief citing 财联社. That pace is slower than the roughly 80% increase the research firm cites for the second quarter and about 40% for the third quarter, though DRAM prices remain high in a cycle of tight supply and strong demand. IT之家 attributes the increase mainly to expanding AI inference demand, with larger models, longer context windows and growing KV cache needs pulling up server DRAM and enterprise SSD capacity and bandwidth requirements while memory makers prioritize high-end products such as HBM. The slower pace is put down to an already high price base, weaker end-market purchasing power and price ceilings in some long-term procurement agreements; IT之家 also reports that Nvidia is considering cutting Rubin Ultra's HBM stack count from 12 to 8 layers due to tight HBM supply.
IT之家 AI · 🔥 0 - Sunrise shows Qiwang S3 inference GPU and 256-card supernode on CCTV Dialogue
On October 17, Chinese GPU maker Sunrise (曦望) appeared on CCTV Finance's Dialogue programme, where it showed off its new inference-focused GPU, Qiwang S3, along with a 256-card supernode built on it, according to a PingWest report. Co-CEO Wang Zhan said the company went "all-in on inference" at the end of 2024 and is targeting a lower cost per token, arguing that agents' repeated planning, retrieval and tool calls will drive inference demand. Sunrise says the S3 trims training-related modules, is the first domestically produced part to use LPDDR6 instead of HBM, and, by the company's own calculation, delivers roughly 5x the inference performance of the previous S2 at about 90% lower inference cost per token.
品玩 实时要闻 · 🔥 0 - Micron develops Edge Inference Memory for physical AI
According to The Elec, Micron executive Lee Cheol said at the SEDEX 2026 forum on semiconductors for humanoid robots in Seoul that Micron is working with major customers on memory for physical AI, targeting humanoid robots, autonomous driving and AR/VR devices. The product, Edge Inference Memory (EIM), stacks low-power LPDDR vertically on a logic die to reduce data-transfer latency and raise bandwidth between AI processors and memory; Micron has not disclosed performance targets but is working to integrate EIM into LPDDR6. The company says humanoid robots currently run roughly 1B–2B parameter models that LPDDR can still serve, but expects a move toward HBM as vision-language models grow, while industry estimates put EIM mass production in the second half of 2028 — a timeline Wccftech rates as a roughly 55% plausible rumor.
cnBeta · 🔥 0 - Dell CEO warns memory shortage will worsen in 2027 as $95B AI server backlog awaits
Dell chairman and CEO Michael Dell told The Wall Street Journal that shortages of high-bandwidth memory and conventional DRAM are expected to worsen in 2027 as AI data center buildouts keep absorbing supply, with new memory fab capacity not expected to reach full output until 2029 or 2030. Dell holds a record backlog of about $95 billion in enterprise and AI server orders, and the reports say memory availability, not demand, is what limits how quickly those systems can ship. Dell says it buys DRAM close to final assembly and signs long-term supply agreements to protect working capital, while IT之家 notes that Micron and SK hynix executives have offered similar views on extended tightness.
cnBeta · 🔥 0 - Micron opens second Taiwan advanced packaging and test plant for HBM
Micron held an inauguration ceremony in Taichung on the 15th for its second advanced packaging and test facility in Taiwan, Micron Advanced Assembly and Test Taiwan 2 (AATT2), with Taiwan President Lai Ching-te, economics minister Kung Ming-hsin, Micron President and COO Manish Bhatia, and customers and partners including Asus, MSI and MediaTek attending, TechNews Taiwan reported. The site is intended to strengthen Micron's advanced packaging and test capacity for high-bandwidth memory (HBM) and AI-driven demand. In its latest financial report, Micron said most of its 2027 HBM capacity has already been booked by customers, expects HBM demand growth to keep outpacing conventional DRAM through 2028, and said its HBM advanced packaging plant in Singapore is expected to begin contributing capacity from 2027.
科技新报 · 🔥 0 - SK hynix unveils 3D FeNAND for analog compute-in-memory AI
SK hynix has published research showing how ferroelectric 3D NAND (3D FeNAND) can perform analog compute-in-memory (A-CiM), with the work appearing in an IEEE paper. The approach stores data and controls current via the polarization state of ferroelectric material, allowing a single cell to produce multiple current states suited to AI weight data, and lets NAND carry out multiply-accumulate (MAC) operations inside the memory to cut data movement. Simulated results under die-size constraints show up to 20x higher throughput and up to 5,000x better energy efficiency than conventional 2D non-volatile memory arrays; the company also presented the technology at OCP Global Summit 2026, saying NAND could eventually support data center and AI workloads beyond storage.
IT之家 AI · 🔥 0 - SK hynix forecasts AI infrastructure performance to grow about 20x by 2030
SK hynix development chief Ahn Hyun said in a keynote at SEDEX 2026 in Seoul on October 14 that AI infrastructure performance scale will grow about 20 times by 2030, according to Yonhap. He said the AI inference market is expanding along three axes — task complexity, context length and request volume — swelling the KV cache and forcing data to be spread across HBM, DRAM and SSD, with the resulting data movement becoming a new bottleneck on performance and power. SK hynix said it is developing 3D stacked DRAM that places DRAM directly atop logic chips such as GPUs and NPUs, plus high-bandwidth flash (HBF), which applies HBM's through-silicon-via stacking and high-speed I/O to NAND; it aims for HBF bandwidth at about 50% of same-generation HBM with roughly 10x the capacity, targets that have no independent measurement yet.
IT之家 AI · 🔥 0 - Ex-Intel CEO Gelsinger calls HBM 'terrible memory'
Former Intel chief executive Pat Gelsinger, in an a16z interview reported by IT之家 via Wccftech, again criticized high-bandwidth memory (HBM), saying it has low bit density, bandwidth limited by chip-edge "shoreline" resources, and power and thermal problems — calling it "disastrous" for memory-intensive AI workloads while conceding it is the best option currently achievable. A 36Kr piece from the account 硅基观察Pro, based on the same a16z podcast, adds that Gelsinger argues AI has made chip design faster while manufacturing and advanced packaging have become the bottleneck, and predicts most of the hundreds of AI inference chip companies will be eliminated. The article also relays his views on memory innovation and optical interconnect, including a 2028–2029 window for NPO and CPO adoption; these are his opinions rather than verified results.
IT之家 AI · 🔥 0 - KLA at SEMICON Taiwan: AI turns chip race into a 'learning velocity' contest
Ahmad Khan, president of KLA's Semiconductor Products and Customer Group, said in a CEO Summit keynote at SEMICON Taiwan 2026 that Moore's Law still matters but AI is changing how semiconductors innovate: transistor scaling alone can no longer carry performance gains, so innovation is widening from a single chip to system-level integration of logic, HBM and 3D DRAM, advanced packaging and high-speed interconnects. He noted that leading-edge GPU dies are so large that a wafer may yield only about five chips, so a single yield-critical defect costs far more, and once stacked DRAM becomes HBM integrated with the GPU on an interposer, "the final yield is the sum of all the parts." Beyond fatal defects, long, high-load AI training and inference runs make latent defects a growing concern; KLA's broadband plasma optical inspection systems generate roughly 125 TB of data per hour, nearly 3 PB a day, which must be sifted with GPU-based high-performance computing plus AI models and physics algorithms. Khan argued the next phase of competition depends on how fast the ecosystem turns faster innovation, richer information and deeper collaboration into "learning velocity."
科技新报 · 🔥 0 - AI datacenter demand revives DDR4 and old PC platforms as memory prices climb
Ifanr reports that as AI datacenters absorb memory capacity, consumer PCs are moving backwards: Gigabyte announced on October 9 that existing B760 and H610 motherboards can support Intel's upcoming LGA1700 processors, expected in early 2027, via BIOS updates, with DDR4 still on the compatibility list, while Intel has repeatedly said DDR4-supporting Raptor Lake remains a core strategy with stable supply for years. AMD launched a 10th-anniversary Ryzen 7 5800X3D at Computex 2026 at a $349 suggested price (2,599 yuan in China), with board makers adding new B550X series boards, and the 2021-era RTX 3060 12GB returned to retail this summer at about $339.99, rising to $460–500 at most retailers by late August. IDC data cited by the outlet put Q3 2026 global PC shipments below 63 million units, down 20.1% year over year and 9.1% quarter over quarter; Omdia says DRAM and SSDs now approach 40% of mainstream PC bill-of-materials costs, and TrendForce expects Q4 conventional DRAM contract prices up another 10%–15% and NAND Flash up 15%–20%.
爱范儿 · 🔥 0
Experience and discussion from the community
Share my HBM experienceAsk about HBM
Nobody has shared their experience with HBM yet.