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.
Linux enthusiasts have made AMD's BC-250 mining accelerator, based on a cut-down PlayStation 5 GPU (Oberon, RDNA 2), work with FSR 4 upscaling. The new fork BC250 FSR4 4.0.0-rc9 is packaged as a DLL for Windows x64 that can be injected into games with FSR support or via OptiScaler.
Nvidia unveiled energy-efficiency optimizations for Vera Rubin platforms: DSX MaxLPS delivers up to 40% more tokens per megawatt. Lambda's 19-node cluster gained 24% more tokens on the same power budget, rising from 4 to 5 million tokens per second.
Insider Kepler_L2 claims AMD and Nvidia have delayed next-generation graphics cards to 2028. The only exception will be the AT2 GPU on RDNA 5 architecture with 40 compute units, launching in 2027 to compete with RTX 5080.
SnowX unveiled the SnowPod, a compact SLM printer measuring 360×360×660 mm for printing with 316L stainless steel. It features a 40 μm laser spot, 30–60 μm layers, an 80×80×100 mm build area, closed powder handling, a NitroPod nitrogen generator, and the free SnowPath slicer.
In Zhengzhou, China's largest tunnel boring machine was completed: diameter 16.7 m, length about 125 m, mass 5,291 tons. The machine is designed for pressures up to 8 bar and can handle steep curves and slopes; it will be used for underwater and urban tunnels.
Analysts note that 6G may disappoint as operators lean toward deploying the new standard on the same spectrum, same equipment, and same vendors used for 5G. Iowa's cornfields have become a testing ground for spectrum.
SanDisk introduced the Optimus GX PRO 8100 and Optimus GX 7100 NVMe SSDs in India. The flagship 1TB model delivers read speeds up to 14,900 MB/s and write speeds up to 11,000 MB/s, with the lineup aimed at gamers among others.
CoreWeave launched the first multi-rack NVIDIA Vera Rubin NVL72 cluster among AI clouds, linking hundreds of Rubin GPUs into one system for agentic AI. It also added cross-region write acceleration and an archive tier to CoreWeave AI Object Storage.
Nvidia CEO Jensen Huang said on the All-In podcast that China will develop its own advanced lithography systems by 2030, calling it 'just a matter of time.' China's SMEE currently mass-produces only 193-nm ArF dry scanners for 90-nm processes, and immersion scanners for 28 nm have not entered mass production.
Huawei announced the Kirin 9050 Pro chipset, claiming it is completely free of US components, and is already using it in the foldable Mate XT. According to internal Huawei research, transistor density increased by 55% and power consumption in some tasks dropped by 66%, but the data has not been independently verified.
Tektronix released the EA-ELR 21000 platform for testing high-voltage DC architectures of AI data centers, supporting ±400 V and 800 V with regenerative efficiency up to 95%. The EA-ELR 21000-80 4U HS module handles 1000 V, 80 A, and 30 kW, with systems scaling to 1.92 MW across eight racks.
SoftBank, KDDI, NTT Docomo and Rakuten Mobile signed a memorandum to jointly study ways to expand mmWave 5G coverage in the 28 GHz band. The operators will consider shared base stations, relays and possible joint equipment access while keeping their own deployments.
Ouster released the Rev8 OS1 Max lidar with native color imaging that detects power lines from up to 200 meters. The device is designed for aviation and infrastructure tasks where color detail is important.
According to insider Jaykihn, Intel is preparing a budget Raptor Lake Refresh chip with 8 P-cores and 8 E-cores to replace the Core i5-14600K, which has only 6 P-cores. It is expected to support DDR5 and DDR4 and cost around $249.
Hanwha Aerospace demonstrated its Mohave STOL drone at a military drone exhibition in Pocheon: the aircraft landed on a narrow tank road and took off again. Takeoff took less than 200m, landing about 100m. The drone has been developed jointly with GA-ASI since 2025 and carries up to 1 ton of payload.
US companies AEVEX Aerospace and Divergent Technologies announced a strategic partnership combining AEVEX's drone expertise with the DAPS digital manufacturing platform based on AI, additive technologies and robotic assembly. The goal is to cut development and certification times for autonomous systems and scale production quickly.
Modix introduced the Generation 5 platform for large-format BIG-series 3D printers based on Duet 3 with distributed CANBUS electronics and capacity for more than 20 axes and drivers. It enables automatic Bondtech INDX tool changing, a compact pellet extruder and 45-degree angled printing. Upgrades are available to existing printer owners, with a demo at Formnext 2026.
Insider Kepler_L2 said the RDNA 5 architecture will get its own neural rendering feature similar to Nvidia's DLSS 5. He said AMD may limit the feature to the new GPU generation, with release expected no earlier than mid-2027.
The Falta Special Economic Zone committee approved SiCSem's project to build a silicon carbide integrated device manufacturing plant near Bhubaneswar. The investment is 34.06 billion rupees (about $355 million).
D-Robotics (a Horizon Robotics spin-off) and startup GigaAI agreed to port the open embodied model GigaBrain-0.7 to the Sunrise S600 edge platform. The partners will handle conversion, quantization and optimization of the model to run directly on robots rather than in the cloud.
Samsung Electro-Mechanics and Qualcomm are jointly developing an organic bridge for 2.1D chip packaging. The technology aims to improve interconnects between dies in advanced AI and HPC products.
BizLink has begun small-volume shipments of co-packaged optics (CPO). According to chairman Roger Liang, mass deployment is held back by fiber array unit (FAU) assembly, thermal management and system reliability, not production capacity.
Schneider Electric introduced a line of prefabricated power modules and skids rated at 2, 2.25 and 2.5 MW for hyperscale and AI data centers. The modules based on Galaxy VXL UPS and reference designs with Nvidia include power and liquid cooling management for the GB300 NVL72 platform. Production is set up in Barcelona, available immediately in Europe and by year-end in North America.
The owner of a PNY GeForce RTX 5090 in a DigitalStorm prebuilt PC found a hole in the 16-pin power connector after smelling burning during a game of Crypt of the NecroDancer. PNY denied the warranty claim, citing use of a non-original cable; a local repair was estimated at about $180.
Growing demand for optical interconnects in AI servers and the transition of co-packaged optics (CPO) to mass production are shifting silicon photonics testing from component characterization to wafer yield protection and early defect detection.
Rising AI infrastructure demand is pushing Taiwan's copper-clad laminate (CCL) makers to adopt Nvidia's high-frequency M9-grade materials. AI servers are moving to 52-layer PCBs, making CCL upgrades critical for system performance.
GMKtec announced the compact EVO-X5 Pro workstation with AMD Ryzen AI Max+ Pro 495, up to 192GB unified memory (up to 160GB for graphics). It claims offline inference of models up to 300B parameters; release on September 28, pricing not announced.
