TLDR Hardware 2026-09-02
Ternus teases iPhone launch 📱, World models for spatial intelligence 🌐, BLE satellites 🛰️
CXMT reportedly begins risk production of HBM3E, targeting mass production in 2027 (4 minute read)
CXMT has reportedly started risk production of HBM3E, a generation behind the HBM4 that Micron, Samsung, and SK hynix are already mass producing, but a meaningful jump for a company that only recently began mass-producing standard DRAM at competitive nodes. HBM is substantially harder to manufacture than ordinary DRAM, requiring high-yield die stacking, through-silicon vias, fine interconnects, thermal management, and packaging on top of competitive dies themselves, so reaching even a risk-production milestone signals China's memory ecosystem advancing past commodity DRAM.
New CEO John Ternus teases "huge launch next week" in first staff memo (2 minute read)
John Ternus, Apple's former SVP of Hardware Engineering, officially took over as CEO from Tim Cook, and used his first staff memo to hype the company's September 9 iPhone event, rumored to unveil Apple's first foldable iPhone alongside a broader Siri overhaul positioning it against ChatGPT and Claude. Ternus, a hardware-first executive taking the helm at a pivotal product moment, also teased a longer pipeline beyond next week, with camera-equipped AirPods, smart glasses, and a touchscreen laptop all reportedly in development.
🔨
Engineering and Applications
Qualcomm unveils Dragonwing Q-2390 and IQ-2390 to push AI into cheaper edge devices (3 minute read)
Qualcomm unveiled two new Dragonwing processors aimed at bringing on-device AI, vision, and connectivity to product categories that have historically struggled with cost, power, and size constraints. The Q-2390 targets consumer and commercial gear like retail kiosks, smart appliances, home robots, and fitness equipment, integrating application compute, on-device AI, camera and display support, and LTE connectivity onto a single chip.
Seraphim Space Invests $30M in Satellite Bluetooth Firm Hubble Network (3 minute read)
Seraphim Space Investment Trust committed $30 million to Hubble Network, a company operating a low Earth orbit satellite constellation that connects directly to off-the-shelf Bluetooth Low Energy (BLE) chips. This architecture enables global connectivity and asset monitoring without requiring cellular towers, ground gateways, or specialized satellite modems. For builders, direct-to-satellite Bluetooth offers an ultra-low-power, cost-effective telemetry solution for tracking and managing distributed field hardware across remote regions.
Atlas: A World Model for Spatial Intelligence (5 minute read)
Atlas is an omni world model and multimodal autoregressive diffusion transformer designed to generate, reconstruct, and simulate interactive 3D environments from text, images, and video. The model natively handles spatial and camera context to synthesize novel views, reconstruct scenes into 3D Gaussian splats, and model dynamic scene interactions over time. Atlas unlocks scalable "Real-to-Sim" workflows by creating realistic synthetic sensor streams (RGB-D) and interactive physical environments from simple video captures, significantly reducing the cost and complexity of training navigation and manipulation policies.
Hydra-0: Action Flow for Generalist World Modeling and Control (7 minute read)
Hydra-0 is a generalist world model that translates physical robot commands into pixel motion trajectories termed "action flow". Operating as a hybrid simulator, it pairs physics engines for robot kinematics with a learned generative video model to predict complex scene reactions like deformable materials and fluid interactions. This shared visual abstraction eliminates the need for custom coordinate remapping across different embodiments, making cross-platform policy training and sim-to-real evaluation significantly more efficient.
Nvidia lays out what a RISC-V CPU needs to run CUDA and join NVLink Fusion (5 minute read)
At Hot Chips 2026, Nvidia detailed the requirements a RISC-V CPU must meet to become a full host processor for its CUDA and NVLink Fusion platforms, alongside its existing x86 and Arm support. For CUDA, Nvidia points to RVA23 profile compliance and the RISC-V Boot and Runtime Services spec to guarantee software binary compatibility across different RISC-V chips, plus hardware-level PCIe I/O coherence and peer-to-peer GPU communication so data can move directly between host and device memory without unnecessary cache flushes or extra copies.
Autonomous-UAV: Sim-First ROS 2 Autonomy Stack for Indoor Person-Following Drone (5 minute read)
This open-source repository provides a simulation-first ROS 2 autonomy stack for an indoor person-following drone, utilizing Gazebo simulation and an ArduPilot Software-in-the-Loop (SITL) digital twin before physical deployment. The architecture integrates vision-based perception models for person detection with flight-control pipelines to handle autonomous tracking and indoor navigation. This project offers a clear reference implementation for designing, simulating, and validating vision-in-the-loop flight control workflows before testing on physical airframe hardware.
The latest in robotics, semiconductors and hardware engineering
Join 500,000 readers for
one daily email