A developer demonstrated Tomb Raider running on an Espressif ESP32-P4 microcontroller, which costs under $25 and consumes about 1 watt. The dual-core 400 MHz chip achieved a playable 1,024 x 600 resolution using the Open…
#Semiconductors
This week in Semiconductors · updated Mon Oct 12 2026
This week’s semiconductor news centered on AI infrastructure and advanced manufacturing. Infineon introduced a 27 kW power-supply reference design for AI server racks, while Nvidia reportedly halted GB202 production for RTX 5090 cards to redirect silicon toward data center and professional GPUs. Kioxia unveiled E1.L SSDs validated for up to 122.88TB for hyperscale AI and object storage. On the foundry side, Rapidus formed a 17-company design group for its 2nm plans, TSMC’s September revenue rose 54.6% year over year despite a monthly dip, and GlobalFoundries set 2028 production for an FD-SOI platform targeting Physical AI. Broadcom’s AI networking showcase also kept Q4 margin expectations in focus.
AI-written weekly briefing drawn from this topic's recent stories.
Technology · Open in Mobble · RSS
Retro Game Runs on Low-Power Microcontroller at Playable Resolution
Cartel's Hidden GPU Mining Operation Discovered in Mexican Mountains
Police in Puebla, Mexico raided a clandestine cryptocurrency mining facility in Tlaola, seizing 300 GPUs, satellite antennas, and electrical equipment. The operation reportedly drew power illegally from a nearby hydroele…
OpenAI May Tap Samsung Foundry for AI Chips, Adding Second Source Alongside TSMC
OpenAI's Korea general manager indicated that the company is expanding its partnership with Samsung to include chip production, potentially using Samsung Foundry for some of its AI accelerators. This would mean OpenAI so…
Leaked roadmap suggests Intel's Nova Lake-S CPUs hit mass production this quarter, launch early next year
A leaked slide indicates Intel will begin mass production of its Nova Lake-S desktop processors in Q4 2026, with the first 28-core SKU launching in Q1 2027. The flagship 52-core model, which reportedly uses two compute t…
Semiconductor Stocks to Watch: Broadcom, Micron, and NVIDIA Lead
Broadcom and Micron are set to report earnings on September 2 and September 30, respectively, following NVIDIA's strong July quarter. Hyperscaler capital expenditures are expanding, memory supply remains tight, and custo…
SK Hynix Eyes Japan Fab, Intel Foundry Partnership; AMD-Cisco AI Goes Live
AMD and Cisco's MI355X-based AI infrastructure is now operational in Saudi Arabia. SK Hynix is scouting a location in Japan for a memory fab joint venture. The company is also considering partnering with Intel's foundry …
NVIDIA Posts Record Revenue, Unveils NVHBM; AMD-Cisco AI Debuts in Saudi Arabia
NVIDIA reported $96.2 billion in revenue for Q2 FY2027, with production ramps for Vera CPU, Vera Rubin NVL72, and Groq 3 LPX for major cloud and AI customers. The company introduced NVHBM, a custom memory controller inte…
YMTC Climbs to Third in Global NAND Bit Shipments, Though Revenue Lags
Yangtze Memory Technologies (YMTC) captured 14% of global NAND bit shipments in the second quarter, breaking into the top three by volume. However, the Wuhan-based chipmaker still ranks fifth by revenue due to its heavie…
Record-Breaking SEMICON Taiwan Showcases AI-Driven Expansion Across Semiconductor Ecosystem
SEMICON Taiwan 2026 attracted over 1,300 companies and 4,300 booths, setting new records. The exhibition broadened its focus to include packaging, smart fab, quantum technology, and system integration. The event highligh…
Co-Packaged Optics Take Center Stage at SEMICON Taiwan as AI Interconnect Bottleneck
At SEMICON Taiwan 2026, co-packaged optics technology has moved from prototype to commercial production, addressing the critical interconnect bottleneck for AI chips. The event also underscored memory's emergence as a st…
AUO Unloads Singapore Facility to Western Digital for NT$8.75 Billion
AUO announced on September 2, 2026, the sale of its AFPD plant in Singapore to Western Digital for NT$8.75 billion. The move continues AUO's strategy of monetizing underused assets while redirecting funds toward advanced…
Taiwan Summit Draws 600+ to Forge AI-Driven Semiconductor Partnerships
Taiwan's Ministry of Economic Affairs convened a summit on September 2, 2026, attracting over 600 officials, executives, and researchers to discuss the semiconductor ecosystem amid AI growth. President Lai Ching-te honor…
Nvidia's RTX Spark N1X chip will come in two variants for October laptops
Nvidia confirmed that its Arm-based RTX Spark N1X chip will ship in two configurations for laptops and mini PCs starting in October. The top-tier version features a 6144-core GPU and up to 128GB of unified memory, while …
AMD's X3D processors dominate Capcom's latest benchmark, while Intel's new flagship lags older models
Tom's Hardware tested 31 CPUs using the in-game benchmark for Onimusha: Way of the Sword, revealing that AMD's X3D chips outperform flagship models by about 10%. Intel's Core Ultra 7 270K Plus trailed behind the previous…
Intel retires 'Fellow' designation, introduces 'Distinguished Engineer' to enforce business-focused technical leadership
Intel has eliminated its long-standing 'Fellow' title for top technical staff, replacing it with 'Distinguished Engineer' roles. The change, which leaves compensation unaffected, aims to align technical leadership with m…
AI boom forces TSMC to nearly double equipment orders, straining supply
TSMC's projected semiconductor equipment needs have nearly doubled since late 2025, reaching 1.9 times the original estimate by July 2026. The surge is driven by AI demand and new fab construction in Taiwan and the U.S. …
Sugon reveals thin AI workstation with 64-thread CPU and 16GB VRAM
Sugon, a Chinese supercomputer maker, has previewed a mobile workstation featuring a domestic 64-thread processor and 16GB of VRAM. The company claims the device can run local 35B Mixture-of-Experts AI models at 50 token…
Quantum Mechanics Remains Incomprehensible to Human Intuition
The article explores why quantum mechanics is so difficult to grasp, citing the double-slit experiment as an example. It notes that while physicists can predict quantum outcomes accurately, the behavior of subatomic part…
Community DLSS 5 tests show high power draw on RTX 5090
Early tests of DLSS 5 using community mods on the RTX 5090 show performance matching Nvidia's stated 8ms per frame at 4K. The modded implementation, however, may not reflect official integration via Streamline. Power dra…
Samsung Targets Doubling HBM5 Bandwidth Over HBM4E
Samsung revealed at a memory summit that its next-generation HBM5 will double the bandwidth of HBM4E and improve power efficiency by 20%. The performance jump may require a wider 4,096-bit interface rather than just fast…
Hybrid Bonding Advances in Logic, Memory Adoption Slips to Late Decade
Hybrid bonding, which connects chip layers with copper instead of solder, is now used in volume for logic chips, with TSMC at a 6-micron pitch and plans for 4.5 microns by 2029. Intel and AMD have also adopted the techni…
GoPro merges with Starman Optical to enter AI infrastructure and defense markets
GoPro will merge with optical-photonics firm Starman Optical, receiving $285 million in cash while retaining about 10% ownership of the combined company. The deal, expected to close by end of 2026, will let GoPro expand …
Wingtech Secures Court Freeze on Nexperia's Chinese Subsidiaries in Ownership Battle
A Chinese court has frozen approximately $318 million in assets belonging to Dutch chipmaker Nexperia and its equipment arm ITEC, covering stakes in five Chinese subsidiaries until August 2029. The order stems from an 8 …
Samsung outlines HBM evolution toward integrated compute-in-memory at Hot Chips 2026
At Hot Chips 2026, Samsung presented a three-phase plan to evolve high-bandwidth memory into a combined memory and compute unit. The roadmap starts with moving the HBM base die to a 4nm logic process, then progressively …
Arm CEO's $800M performance package draws proxy adviser opposition
Arm is seeking shareholder approval for a performance-based pay plan for CEO Rene Haas that could reach $800 million if the company hits a $2 trillion valuation. Proxy advisers ISS and Glass Lewis have recommended voting…
China's CXMT launches trial HBM3E output, eyes 2027 volume production
CXMT has initiated risk production of HBM3E memory, according to a report. The company is sampling the chips with Alibaba's T-Head and Cambricon Technologies. Mass production could begin in 2027, though specifications re…
G.Skill expands Ryzen memory lineup with Flare X5X kits featuring tighter subtimings
G.Skill introduced the Flare X5X DDR5 memory series, designed for AMD Ryzen processors and using AMD's EXPO ULL technology to optimize subtimings. The initial kit runs at DDR5-6000 with primary timings of 36-36-36-96 at …
Heat-driven cooling prototype could let data centers recycle waste heat for refrigeration
Researchers from Germany and Japan have built a prototype cooling device that uses shape-memory alloys to convert waste heat into mechanical motion, driving a refrigeration cycle without electricity. The system, detailed…
Nvidia invests in MediaTek to keep its AI infrastructure central as custom chips rise
Nvidia is pouring $3.5 billion into MediaTek, which will adopt Nvidia's NVLink Fusion technology to design custom AI chips for cloud providers and AI labs. This move lets Nvidia remain the backbone of AI data centers eve…
Nvidia driver 615.56 crashes mVolt+ tool, blocking extreme power-limit tweaks on RTX 50-series
Nvidia's newest driver release, version 615.56, causes mVolt+ to crash whenever users attempt to adjust core power limits on RTX 50-series graphics cards. The open-source utility previously let enthusiasts push power dra…
Taiwanese PCB maker Unimicron under investigation for allegedly relabeling China-made boards
Taiwanese prosecutors raided Unimicron's headquarters and a plant on August 28, questioning 14 employees as suspects in an alleged scheme to disguise China-made PCBs as Taiwanese products. The company, a major supplier t…
Chinese memory maker CXMT claims first LPDDR6 production, targets Xiaomi flagship
ChangXin Memory Technologies announced it has begun mass-producing LPDDR6 memory for a Xiaomi smartphone, according to a social media post cited by Reuters. The move places the Chinese firm roughly on par with establishe…
CXMT takes Pentagon to court over military-linked designation
ChangXin Memory Technologies has filed a lawsuit against the U.S. Department of Defense to contest its classification as a Chinese military company. The DRAM manufacturer asserts it has no ties to the People's Liberation…
Decoding TPUs: Google's Custom AI Chip vs Traditional GPUs
Google's Tensor Processing Unit (TPU) is a proprietary AI accelerator originally built for data center workloads, but the Pixel 11's Tensor G6 chip uses a TPU to handle on-device AI tasks, replacing the typical NPU. Unli…
Nvidia's Edge in AI Shifts from GPUs to System Orchestration
Nvidia's competitive moat in AI is increasingly tied to its data center system orchestration, not just its GPUs. The company's Vera Rubin architecture includes specialized components like the Vera CPU to manage data flow…
Trump's Unexpected Call Disrupts Nvidia's Staff Meeting
During Nvidia's all-hands meeting on Thursday, CEO Jensen Huang paused to take a phone call from President Donald Trump, whose voice was audible through the onstage microphone. The brief interruption occurred hours befor…
Micron Taiwan employees threaten strike over profit-sharing demands
A preliminary survey shows 80% of Micron workers in Taiwan support a strike if the company doesn't change its bonus system. The labor union wants a profit-sharing scheme similar to Samsung and SK hynix, which pay out 10%…
Unofficial patch lets DLSS 5 run on Ada Lovelace GPUs
A modder has modified Nvidia's DLSS 5 DLL to run on RTX 4000 series GPUs by swapping out CUDA instructions that only work on Blackwell hardware. The patch, applied to a leaked early build from NBA 2K27, enables Neural Re…
Flash storage's enduring trade-offs: density gains vs. wear limits
Flash memory, particularly NAND, dominates portable storage due to its compact size, low latency, and shock resistance compared to hard drives. However, it costs more per gigabyte, degrades with repeated write-erase cycl…
SK hynix starts construction of first US HBM assembly facility in Indiana
SK hynix held a groundbreaking ceremony for its HBM production base in West Lafayette, Indiana, which will focus on advanced packaging, testing, and R&D. The $4 billion facility is expected to open its cleanroom by Octob…
AI sector warns Trump's chip tariffs could harm data center growth
The tech industry is alarmed by reports that the Trump administration plans to impose sweeping semiconductor tariffs, potentially affecting products containing chips. The Computer and Communications Industry Association …
YMTC sets 2027 goal to dethrone Samsung and SK hynix in NAND flash market
YMTC, a Wuhan-based NAND flash maker, has informed investors of its goal to become the world's largest producer by the end of 2027, surpassing Samsung and SK hynix. The company recently filed for a 33 billion yuan ($4.9 …
Glass substrate packaging still lacks commercial adoption despite progress
Glass-core substrates, once touted as a replacement for organic packaging, are now in final qualification but have not appeared in any commercial product. SKC's Absolics plant is undergoing package-level reliability test…
US considers broader chip tariffs on consumer electronics and servers
The administration is exploring a second round of semiconductor tariffs that would also cover finished products like laptops, consoles, and servers. Commerce Secretary Howard Lutnick favors a quota system that ties duty-…
OpenAI's Jalapeño AI accelerator focuses on efficiency over raw power
At Hot Chips 2026, OpenAI detailed its first AI accelerator, Jalapeño, which doesn't beat Nvidia's Blackwell in raw performance but offers better performance-per-watt and low latency for inference. The chip was developed…
Lawsuit alleges EPA fast-tracked dangerous chip-making chemicals
Earthjustice is suing the EPA for approving two photoacid generators used in semiconductor manufacturing without adequate safeguards. The chemicals are said to be acutely toxic, with exposure potentially causing sudden d…
RTX 3060 12GB revival backfires as prices soar to $500, matching newer RTX 5060
Nvidia's reintroduced RTX 3060 12GB has seen its price climb roughly 45% since late June, with most models now near $500. For example, MSI's Ventus 2X OC rose from $300 to $480, while Asus's Dual V2 OC jumped from $330 t…
Corsair's 32GB DDR5-6000 kit drops to $401.99, becoming the best-priced option at that speed
Corsair's Vengeance 32GB DDR5-6000 memory kit has been discounted by $215, bringing its price to $401.99. That makes it the most affordable 32GB kit at that speed currently available. The deal comes amid ongoing DRAM pri…
Nvidia's quarterly revenue climbs to $96B as it locks in massive memory supply deals
Nvidia reported fiscal Q2 2027 revenue of $96.2 billion, a 106% year-over-year increase, driven by demand for AI hardware. The company also committed to purchasing up to $160 billion in memory, including a pact with SK h…
Amazon to deploy 2 million additional Nvidia GPUs as AI compute demand soars
Amazon Web Services will add 2 million Nvidia GPU chips to its data centers in 2027 and 2028, tripling an earlier commitment made five months ago. The expanded partnership also integrates Nvidia's networking, software, a…
Nvidia unveils tailored HBM design for AI partners, boosting bandwidth and cutting power
Nvidia introduced NVHBM, a custom high-bandwidth memory base die designed for its NVLink Fusion partner program, offering up to 30% more bandwidth per stack than standard HBM4e while reducing power consumption by 15%. Th…
Nvidia projects $108B quarterly revenue as data center sales surge
Nvidia reported record revenue of $96.2 billion for the last quarter, with data center sales reaching $89 billion and profits more than doubling to $59.7 billion. The company forecasts $108 billion in revenue for the upc…
Fujitsu's Monaka server CPU uses stacked 5nm cache die and dual 256-bit SVE2 units
At Hot Chips 2026, Fujitsu revealed its 144-core Monaka server CPU, which places its entire last-level cache on a separate 5nm SRAM die stacked beneath a 2nm compute die via hybrid bonding. The chip runs dual 256-bit SVE…
Nvidia's power management tech promises more compute per megawatt in data centers
Nvidia highlighted its DSX MaxLPS power management suite during a Hot Chips presentation on the Rubin GPU. The company claims that using these technologies allows operators to fit 40,000 Rubin GPUs within a 100MW power b…
At Hot Chips, Nvidia details Groq 3 LPX rack and publishes independent inference test
Nvidia's vice president of hardware, Igor Arsovski, presented the Groq 3 LPX rack architecture at Hot Chips 2026. The company also released its first third-party benchmark for the hardware. The LP30-based rack is already…
High Bandwidth Flash Faces Narrow Use Cases, OXMIQ Says at Hot Chips
OXMIQ Labs presented at Hot Chips 2026 that High Bandwidth Flash (HBF) cannot replace High Bandwidth Memory (HBM) for the vast majority of workloads. The technology may serve as a specialized memory tier for large, relat…
d-Matrix's Raptor accelerator bonds compute die to custom DRAM, achieving 100 TB/s per card
At Hot Chips 2026, d-Matrix introduced Raptor, which it claims is the first 3D DRAM accelerator for generative inference. The design places a TSMC 4nm compute die face-to-face on a custom DRAM die with a 36-micron pitch,…
Arm Unveils AGI Server CPU Architecture: 136 Cores, 2TB/s Interconnect, 300W TDP
At Hot Chips 2026, Arm disclosed technical details of its upcoming AGI data center CPU, which uses two chiplets with up to 136 Neoverse V3 cores. Each chiplet contains 70 cores, a six-channel memory controller supporting…
DIY Semiconductor Enthusiast Creates LEDs Using Homemade Cleanroom and eBay Laser
Matthew Hartensveld, known as Dr. Semiconductor, has expanded his backyard semiconductor lab to fabricate LEDs. He used a sapphire wafer coated with gallium nitride, then etched patterns with a powerful laser purchased o…
NVIDIA's Jetson Orin Nano 2 brings frontier-level AI to edge devices with doubled speed
NVIDIA introduced the Jetson Orin Nano 2, an entry-level edge AI computer that doubles the inference performance of its predecessor while consuming 40% less power in 15-watt mode. The system delivers 78 TOPS of AI comput…