Hyosung Heavy Industries was selected for KEPCO's project to localize and test high-voltage HVDC systems, responsible for valves and controllers. At a Busan exhibition, the company showcased converter valves, LPIT, STATCOM, SST, and a hydrogen generator.
Singular Photonics released the Litavis CMOS SPAD sensor with integrated digital processing of photon events on-chip and in-pixel. It delivers 47% photon detection efficiency at 785 nm, 37 ps timing resolution and 4.3 ns dead time; availability dates were not given.
According to Silicon Data, the residual value of Nvidia B200 GPUs is 158% of their original launch price—58% above launch—about a year after mass shipments. The reason is shortage and high demand for inference workloads.
GMO Internet Group has launched Japan's first 'humanoid ambulance' in Tokyo—a van with diagnostics, spare parts, tools, and engineers for on-site repair of broken robots. It carries a backup humanoid for replacement, while faulty ones are sent to the Humanoid Lab repair base in Shibuya. The service currently covers only GMO AIR robots and operates on request.
AEye and Swedish AB Swiftstrike AI agreed to integrate the STRATOS lidar into Swiftstrike autonomous counter-drone systems. The sensor detects objects at up to 0.93 miles (1.5 km) and offers twice the resolution of AEye's Apollo, helping identify drones without radio signals.
A team from Xi'an University with City University of Hong Kong and Fudan University identified the cause of degradation in AlScN ferroelectric memory — mobile nitrogen vacancies. A new structure design achieved 10 billion rewrite cycles, about 100 times the previous limit of 100 million cycles.
Nvidia unveiled DSX MaxLPS and DSX Flex at AI Infra Summit. In a test with Lambda, 19 nodes fit into the same power budget instead of 16, boosting throughput and performance per watt by about 24%. DSX Flex allows temporarily reducing data center load on grid request.
Nvidia uses 10-40 RISC-V cores per GPU for power management, video and security. Since 2018 one core handles much of the graphics driver, replacing the proprietary Falcon controller. Nvidia estimates over a billion such cores shipped in 2024.
Intel CEO Lip-Bu Tan said risk production for the 14A process will begin in Q1 2027 instead of the second half. The defect density of 14A reached 0.5 in June 2026, with a target of 0.1–0.2 by launch.
Intel's CEO said AI infrastructure expansion is making memory supply a bottleneck for the semiconductor industry. He warned of a memory shortage and tighter CPU supply constraints in 2027.
A KAIST team used a single material for efficient charge injection into atomically thin semiconductors. The technology could enable vertical integration of ultrathin layers, increasing density and energy efficiency of future AI chips.
24 public Taiwanese chip packaging and testing companies collectively earned NT$768.9 billion for January–August 2026. Growth in the semiconductor back-end segment is broader than that of the largest player, ASE.
Onsemi announced the EPP platform, which embeds heterogeneous Si, SiC, and GaN dies directly into 12-inch silicon wafers with precision RDL. The technology replaces wire bonds and molding compounds, enabling up to 5x power density.
GE Aerospace is preparing to assemble the first test unit of the XA102 adaptive engine under the US Air Force's NGAP program: component procurement and unit tests are underway. The engine is GE's first designed and assembled using a digital model (MBD) instead of blueprints; the project builds on XA100 experience and competes with Pratt & Whitney's XA103.
DIGITIMES reports rapid growth in demand for silicon carbide power devices driven by the EV market, triggering a wave of capacity expansion among SiC substrate makers. Prices for 6-inch substrates are expected to start recovering in 2027.
MLPerf Inference v6.1 results are out. AMD tested a 512-GPU Instinct MI355X cluster for the first time, Nvidia previewed Vera Rubin NVL72 with up to 3.7x gain over GB300, and Intel expanded participation with Xeon 6 and Arc Pro B70.
TSMC will raise 2nm wafer capacity from about 90,000 per month in late 2026 to about 110,000 by mid-2027, a 22% increase, according to supply chain sources. 3nm output will rise from about 180,000 to about 210,000 wafers per month, with part of the gain from its Arizona fab.
A mobile RTX 5070 Ti in a Lenovo Legion Pro 5 Gen 10 was overclocked from a stock 140W to 160W via VBIOS and voltage curve modification. The 3DMark Steel Nomad score rose about 16% to 4,839 points.
Nvidia expanded its Blackwell workstation lineup with the RTX PRO 5500 card on the GB202 GPU with 84 GB GDDR7 and 21,760 CUDA cores, matching the RTX 5090. Memory bandwidth is 1398 GB/s, TDP up to 600 W, and MIG support splits it into two 42 GB instances. Price not announced.
Hardware Unboxed recorded melting of a 12V-2x6 connector on a ROG Astral RTX 5090: only two wires carried the load, and the connector melted on the PSU side. Nvidia already used per-phase power balancing in the RTX 3090 Ti but dropped it starting with the RTX 4090.
RAMXEED introduced the RAMXEED ULTIMATE line of ferroelectric memory, guaranteeing data retention for 10 years under continuous exposure to 125°C. The products target applications in extreme temperature conditions.
IoT Analytics: cellular IoT module shipments grew 15% year-over-year in the first half despite memory shortages. LTE Cat 1 bis modules drove growth, while RedCap and eRedCap are just entering mass production, and NB-IoT continues to decline from its Q1 2024 peak.
At the NVIDIA AI Infra Summit, AgentX benchmark results showed the Vera Rubin NVL72 delivers up to 30 times more agent workloads per megawatt than the GB300 NVL72. Measurements were made on early silicon and cover only coding-agent sessions, raising questions about generalizability.
Infleqtion and Nvidia presented validated quantum error correction code: 98 physical qubits yield 18 logical qubits, about 5.4 physical qubits per logical at 18.4% encoding efficiency. It uses Infleqtion's open-source qLDPC library and Nvidia's CUDA-Q Logical layer.
Developer LevinAI released an experimental tool called NvpwrControl that allows owners of laptops with GeForce RTX 50-series GPUs to raise the GPU power limit. For RTX 5080 and 5090 Laptop, the maximum is raised from 150 to 225 W, for RTX 5070 Ti to 180 W, and for lower models to 140 W.
Dutch startup Axelera released the Europa platform for AI inference in enterprise data centers. The AIPU delivers 629 TOPS at 45 W, and the Server 250p card achieves 4205 tokens/s on Qwen3 8B with efficiency of 32.1 tokens/s per watt.
Arm is expanding its computing platform for the era of AI agents that will run continuously on smartphones, in data centers, cars and robots, perceiving their environment and reasoning about goals. The company introduced new platforms for edge, cloud and robotics.
MLCommons published MLPerf Inference v6.1: 30 participating organizations and 486 results. End-to-End RAG and Edge Agentic Inference tests were added, and results from NVIDIA Vera Rubin NVL72, AMD Instinct MI350P, Intel Arc Pro B70 and Ryzen AI Max+ 395 passed review for the first time.
Apple released the Mac Studio with the M5 Ultra chip — the first Ultra chip with Neural Accelerators in every GPU core. It claims up to 4.3x peak AI performance and up to 1.8x graphics performance versus M3 Ultra, with up to 512 GB memory at 1.2 TB/s bandwidth.
Taiwanese DRAM developer PieceMakers, backed by Nanya, began trading on the Emerging Stock Board on September 16 at a reference price of NT$740, closing its first session up 23.6% at NT$915. The company is betting on custom memory for inference: DRAM stack is attached directly to the processor via hybrid bonding, not through HBM. About 40% of revenue in the first half of 2026 came from custom AI developments, with the first major client expected no earlier than 2027.
