Preparing your page.
The news will be here in a moment.
IN A CONNECTED WORLD
Multiple perspectives. One place.
Your curiosity, your daily feed.
Tina Huang shows which AI models run locally on each hardware class, from Arduino and ESP32 through Raspberry Pi, iPhone, MacBook and home servers up to an RTX 4090 and eight H100s, all in live demos.
The Forlinx OK1126Bx-C squeezes 3 TOPS of edge AI inference, a 4K video pipeline and 0.007W sleep draw onto a 120x75mm board around the Rockchip RV1126B, with a Raspberry Pi-compatible header for fast prototyping.
The iPhone 18 Pro's A20 Pro chip hit 66-80 FPS at 1080p in emulated GTA 5, matching a desktop GTX 1650 with 1,180% better performance per watt. 2nm WMCM packaging and a bigger vapor chamber broke the mobile thermal wall.
Beijing is preparing to let Alibaba and ByteDance buy Nvidia's 84 GB RTX Pro 5500 workstation card. The signal comes from The Information via TNW and remains unverified by Reuters; the 6,000-dollar card sits in the gray zone of US export controls.
ASUS unveiled the ExpertCenter Pro ET900N G3 workstation with NVIDIA's GB300 Grace Blackwell Ultra superchip: 748 GB of unified memory, 20 PFLOPS, a 1,600-watt titanium PSU and closed-loop liquid cooling. The 27-kilo tower brings trillion-parameter models to the desk.
Experimental virtio-nvgpu keeps one RTX 3060 with the host while opening graphics access to four Linux guests at once; the card gave 102.9 fps alone and 103.69 fps combined when split four ways.
As Micron reports Q4 on September 30, the real question goes beyond HBM: is agentic AI turning CPU memory and enterprise storage into critical infrastructure too?
IonQ will install its 256-qubit Superion 256 system at Nvidia's accelerated quantum research center in Boston, becoming the center's first quantum processor; the race to fuse classical GPUs and CPUs with quantum processors into one giant system has officially reached hardware.
CXMT said on September 20, 2026 that its fifth-generation DRAM platform entered mass production, with an 11.95-nanometer half-pitch and fifty percent more dies per wafer feeding 24Gb mobile and desktop chips.
Before buying the $12,000 Mac Studio M5 Ultra, the presenter rents a 96GB RTX Pro 6000 for $2 an hour and puts an open-weights model through three real jobs. Two tests end in draws, the frontier model wins the trap question, and the decision rule sharpens: rent first, measure, then buy.
In 2026 Alphabet said it would begin delivering TPU accelerator systems into customer-owned data centers and opened parts of its software stack to run on other hardware. The move turns a defensive advantage into a standard, and the industry's real bottleneck is neither design nor capital but high-bandwidth memory and grid power.
Lon.TV bought the entry 16 GB / 256 GB M6 Mac mini with his own money and put the $899 base model to the test: despite a steep hike over the M4, its 2.7 W idle, 69.3 Speedometer score, instant 4K Final Cut effects and 45-50 fps gaming show the base config is surprisingly capable for everyday and creative work — except for heavy local AI and storage.
In an August press release Apple cleared two thresholds at once: it put the 2-nanometer M6 inside the $899 Mac mini and pushed the 3-nanometer M5 Ultra to the top with UltraFusion that makes four dies behave like one processor.
With M3 Ultra and new M5 Ultra side by side, I tested Apple’s claims of up to 4.3× AI, 2× storage and 1.2 TB/s bandwidth on identical models and stacks; results largely confirm marketing, but the felt gain depends on prompt length.
Tech YES City connects $5,000-plus GPUs to Fed policy and a GPU-collateral debt boom: M2 is rising while QT is over, hyperscalers pour over $700 billion into silicon, and gamers foot the bill.
On the All-In Podcast stage, Naveen Rao unveiled Unconventional AI's oscillator-based dynamical computer: a first physical prototype that draws roughly 500 nanojoules per image, pitched as a thousand-fold efficiency answer to the 12-gigawatt energy wall implied by Google's 3.2 quadrillion tokens per month.
In Dave2D's hands-on, the 2026 Mac Mini with M6 and M5 Pro cuts GPU pressure from ~75% to ~60% on a real-time 4K editing rig, while the M5 Ultra Mac Studio pushes unified memory to 512 GB and pools up to 768 GB over Thunderbolt 5.
Unveiled in 2015 with claims of thousand-fold speed and endurance gains, 3D XPoint tried to turn half a century of phase-change research into Intel's Optane franchise, yet runaway 3D NAND scaling, heavy data-center integration costs and diverging roadmaps from its two parents quietly ended the new-category dream in 2022.
Two years after Qualcomm cancelled the X Elite dev kit, the first real Snapdragon X2 Elite desktop is here; the Asus Ascent Q10 packs 18 cores, 152 GB/s memory and 80 TOPS NPU, pushing Apple M4 family in Geekbench and browser tests, yet performance collapses without the right ARM64 builds.
Inference spending has overtaken training, close to two percent of United States output is flowing to artificial intelligence infrastructure, and rack-scale integrated systems are overtaking part picking. Austin Lyons explains through the prefill versus decode split and the rise of neoclouds where the next trillion-dollar chip company could emerge.
As RAM and storage prices surge, Linus Tech Tips asks whether Blu-ray is truly the cheapest consumer archive per gigabyte; the math favors optics by up to 2×, yet longevity, speed and scale quickly rewrite the total cost.
A Reddit poll on $10,000 home AI rigs, cheap Chinese subscriptions and a used RTX 3090 climbing from $639 to $1,400 all circle the same question: does the 128 GB box in your cart beat a $200-a-month plan? The video measures unified memory, bandwidth and coding agents' hidden token load and ties the answer not to money, but to privacy and the job you actually run.
In Hardware Unboxed's 4,316-word investigation, AMD's RDNA 4 Radeon RX 9050 is dissected: the retail 8GB model offers a 128-bit bus and 288 GB/s with 32MB Infinity Cache, while the OEM-only 4GB variant halves that to a 64-bit bus, 144 GB/s and 16MB cache, hidden inside a $750 CyberPowerPC Gaming Master at Best Buy under the same name.
In a late-2026 Shorts update Paul's Hardware retracts two years of takes: the RTX 5090 that launched in January 2025 at $1,999 as a GeForce gaming card now lists from $7,000 with shady third-party sellers and above $9,500 at retail on Newegg, pricing it like an AI accelerator rather than a gaming purchase.
Among the noise, meaningful connections.
