Daily Tokens
Saturday, November 21

Hardware news for Saturday, November 21

Listen to today's recap · 2 min · Transcript

Today

Japan's Rapidus reportedly is struggling to produce working SRAM on its 2nm node, putting its 2027 mass production at risk. Xiaomi's 18 Pro phones add a second screen on the back, but its usefulness is still unproven. A homelab Optane build is running Kimi K3 at 3.11 tokens per second, and a classic Super NES can now run a ChatGPT client. Elsewhere, Knoxville police are sending semi-autonomous drones to 911 calls, Medtronic's Hugo is on track to pass 50,000 cumulative procedures, and defense firm Aselsan has developed an EMP weapon that can fry 51 drones at once.

New today

Moderate impact

Rapidus reportedly struggles to make working 2nm SRAM

Japan's Rapidus has reportedly demonstrated functioning gate-all-around transistors on its 2nm node. Integrating them into working SRAM memory cells appears to be a considerably greater challenge. That puts the company's 2027 mass production target at risk.

SRAM is a core building block of advanced chips, so a memory-cell failure would undermine Rapidus's bid to compete in leading-edge manufacturing.

Low impact

Xiaomi 18 Pro and Pro Max add a second rear screen

Xiaomi's 18 Pro and 18 Pro Max both have a second display on the back, used for selfies, notifications and AI apps. Xiaomi tried a rear screen on the 17 Pro models, but those never shipped outside China. The new phones are the first chance to test whether a rear screen is more useful than it was on the 2021 Mi 11 Ultra or earlier efforts from Nubia and Vivo. After a week or two with both, the rear display appears to have one obvious use case and is not yet convincing.

Rear screens have failed to catch on before, so their future depends on whether the apps justify the extra display.

Low impact

Homelab Optane server runs 2.8T-parameter Kimi K3

A second-generation homelab build using Optane PMem 200 memory and a RAM drive is running Moonshot AI's Kimi K3 at 3.11 tokens per second. Two RTX 3090 GPUs handle the work. The model has 2.8 trillion parameters.

It shows how far used consumer GPUs and older server memory can be pushed to run very large AI models at home.

Low impact

Super NES runs ChatGPT client via Super FX 2 chip

The classic Super Nintendo now has its own ChatGPT client. The 2.2-million-parameter chatbot's replies are sped up by the Super FX 2 chip's 21 MHz RISC core.

It is a fun retro-hardware hack that shows how little compute a basic chatbot needs.

Low impact

Knoxville police launch drone first-responder program

The Knoxville Police Department has launched a Drone as First Responder program. Semi-autonomous drones are dispatched to the scene after a 911 call. The drones are designed to reach nearly all of Knoxville in 90 seconds or less, giving responders real-time situational awareness.

It shows autonomous drones moving into routine emergency response in a US city.

Friday, November 20

Low impact

Teenage Engineering launches $299 CD player with exposed disc

Teenage Engineering announced the EP/BC-1, a slab-style CD player that leaves the spinning disc visible instead of hiding it under a lid or tray. The company calls it the first product in its new broadcast series and says it was designed to bring physical audio formats into a modern context. The EP/BC-1 is available now through Teenage Engineering's online store for $299.

It is a stylish new hardware product built around a physical format that most people have stopped using.

Low impact

MSI EdgeMesa N AI+ review: Nvidia RTX Spark on Windows Arm

MSI's EdgeMesa N AI+ is a Windows-on-Arm mini PC built on Nvidia RTX Spark N1X, with 128GB of unified memory. The review finds strong local AI performance and the ability to game on the side, though with trade-offs. At $7,499, its price is not much cheaper than the more versatile DGX Spark systems.

It shows Nvidia's Arm PC silicon reaching desk-sized local AI machines, but the price limits its appeal.

Low impact

Phison shows computational SSD with NPU and 64GB memory

Phison showed a computational-storage SSD with a built-in NPU and 64GB of memory that can run AI frameworks on the drive itself. The company also showed Mixed Media pSLC and QLC drives at SC26.

Putting an NPU and large memory on the drive could move some AI processing closer to where the data is stored.

Low impact

Lawyer's $9,450 AI server runs DeepSeek V4.1 Flash on V100s

A lawyer built a private AI server from 12 used Nvidia V100 GPUs for $9,450. It now runs every expert in DeepSeek V4.1 Flash at 4-bit NVFP4 precision.

Used data center GPUs are making large local model hosting far cheaper for individuals.

Low impact

Raspberry Pi 5 cyberdeck runs offline for over 24 hours

A builder shared a rugged, battery-powered cyberdeck built on a Raspberry Pi 5 with a touchscreen and a 1 TB NVMe SSD. It carries offline Wikipedia, worldwide maps, GPS/PPS time sync, SDR, educational videos, and a local Wi-Fi network. The build has demonstrated more than 24 hours of battery life.

It is a practical example of a fully self-contained offline computer that other hobbyists can copy.

Thursday, November 19

Moderate impact

Nvidia and Microsoft expand Vera Rubin cloud access and bring accelerated computing to Windows PCs

At Microsoft Ignite, Nvidia and Microsoft said they are expanding access to Azure virtual machines powered by Nvidia Vera Rubin NVL72 systems. The collaboration also extends to industrial systems and Windows PCs, with the stated goal of helping organizations build, customize and deploy intelligent applications.

It brings Nvidia's newest rack-scale systems to cloud customers and pushes accelerated computing onto everyday Windows PCs.

Moderate impact

Intel's 18A process reaches 70% yields as Microsoft signs Xeon and EMIB deals

Intel says its 18A process has reached 70% yields, which it describes as an important milestone. The company is also signing a long-term Xeon supply deal and an EMIB usage agreement with Microsoft.

The yield milestone and a major cloud customer commitment suggest Intel's advanced manufacturing is becoming commercially viable.

Moderate impact

Meta reportedly plans to sell the Muse Charm for $300 with free Verizon 5G

Meta is reportedly planning to price the Muse Charm, a small gadget with a screen and fingerprint sensor for talking to its Muse AI agent, at $300. Zuckerberg previously said the device would ship in time for the holidays in December, and it may come with free Verizon 5G service. Meta has said the Charm packs real-time voice and avatar technology into the device.

It would be a direct test of whether consumers will pay for a dedicated AI companion gadget rather than using a phone.

Moderate impact

Google signs deal with Blue Origin to launch satellite data centers into orbit

Google has signed a deal with Blue Origin to send satellite data centers into orbit. The arrangement is part of Project Suncatcher, an experiment aimed at running AI systems in space.

It is a concrete step toward putting AI compute hardware in orbit, testing whether space can host data centers.

Low impact

Silicon Motion's SM8466 PCIe 6.0 SSD controller has taped out

Silicon Motion has demonstrated its SM8466 PCIe 6.0 SSD controller, which has finally taped out. The company expects the controller to enter the market in 2027 to 2028, with improved quality of service and samples described as imminent.

It previews the next generation of high-speed SSD controllers, showing where storage performance is heading.

Wednesday, November 18

Low impact

Nvidia RTX Spark mini PCs go on sale from Asus, Dell and MSI

Desktop systems built on Nvidia's RTX Spark N1X chip are now available from Asus, Dell and MSI, following earlier RTX Spark laptops. The mini PCs include a 6144-core Nvidia Blackwell RTX GPU, a 20-core Nvidia Grace CPU and up to 128GB of memory. Asus's ProArt GR1X starts at $4,499 with 64GB of memory and 1TB of storage. Acer, Gigabyte and HP desktops are coming later, and Microsoft's Surface RTX Spark Dev Box is slated to start shipping this month.

This brings Nvidia's Arm-based RTX Spark platform to desktops, aimed at running local AI agents, generative media and model fine-tuning on Windows machines.

Low impact

Disney unveils three AI-driven Star Wars animatronic prototypes

Walt Disney Imagineering R&D revealed three new Audio-Animatronic prototypes named Vilo Vele, Kado Bre'a and Jibu Bre'a, based on the Star Wars Anzellan characters. The figures rely on Amazon's AWS cloud and AI capabilities to respond to guest interactions in real time. Current park animatronics typically follow predefined routines or rely on a hidden puppeteer for spontaneity.

It is a concrete example of cloud AI being used to drive physical, real-time character behavior in a commercial entertainment setting.

Low impact

Shelly launches a $499.99 smart valve to shut off home water on leaks

The Shelly Neo Smart Water Valve is available now from the company's online store for $499.99, plus installation costs if buyers do not install it themselves. It mounts on the home's main water line and automatically cuts the flow when one of Shelly's leak sensors detects a problem. The companion sensor costs $28 and runs on a battery for two years.

It moves smart home leak detection from alerts to automatic shut-off, targeting the much costlier damage a leak can cause.

Low impact

8BitDo announces a GameCube-style Bluetooth gamepad for $49.99

The 8BitDo GC Bluetooth Gamepad is available for preorder starting today in orange or purple for $49.99, with shipping expected on December 10. It follows the GameCube controller's button layout, thumbstick design and shoulder buttons more closely than the company's earlier 8BitDo 64 remake. Nintendo's own wireless GameCube-style controller costs $69.99.

It offers a faithful, cheaper wireless option for GameCube-style controller design, which the company's earlier N64 remake did not provide.

Low impact

Retro Song Bird Kickstarter speaker plays music straight from Game Boy cartridges

The Retro Song Bird, now on Kickstarter, plays soundtracks directly from original Game Boy cartridges. It can export individual audio channels to a PC, back up game saves, flash writable cartridges and sync two units for chiptune DJ performances. The project is a hardware speaker and cartridge tool rather than a game emulator.

It is an unusual, hands-on hardware project that turns vintage handheld cartridges into a music and archiving tool.

Tuesday, November 17

Moderate impact

Boston Dynamics' Atlas humanoid to begin first non-automotive trial at a major chipmaker in early 2027

Boston Dynamics' Atlas humanoid robot will undergo its first trials outside the automotive industry at a major semiconductor manufacturer in early 2027. The trial marks the first time the robot has been tested outside automotive work. The announcement gives no details on the tasks or the number of robots involved.

It is the first sign that Atlas is moving beyond automotive factories into a new industrial sector.

Low impact

Flipper Devices opens Kickstarter preorders for the $399 Flipper One Linux multi-tool

Flipper Devices has launched a Kickstarter campaign to preorder the Flipper One, which it describes as an open Linux multi-tool built for high-performance computing and IP networking. The device is priced at $399 with shipping expected to start in July 2027. It costs more than the Flipper Zero, which sells for $169, but the company says the Flipper One is not a successor to that wireless hacking multitool and is a different product.

It gives a concrete price and delivery date for a pocket-sized Linux computer aimed at networking and high-performance work.

Low impact

TRMNL opens preorders for a 31.5-inch color E Ink Frame with a Spectra 6 panel

TRMNL is taking preorders for the Frame, a 31.5-inch full-color E Ink display in a handmade wooden body. It uses a Spectra 6 panel, which can show thousands of shades and holds an image without power, and runs the open-source Lazarus color conversion engine. One listing puts the price at $1,299, while another gives £1,001 for UK pre-orders. The company positions it as a photo and artwork display that can also show family dashboards, and says it is not limited to the four colors of its earlier color E Ink devices.

It brings a large, low-power color E Ink display with an open color pipeline to the consumer market at a specific price.

Low impact

Nvidia launches DGX Spark HPC Developer Kit as a four-node desktop cluster

At SC26 in Chicago, Nvidia introduced the DGX Spark HPC Developer Kit for researchers, academics and developers. The kit is a four-node cluster of DGX Spark systems, with each node directly connected to its two neighbors in a ring. An NVIDIA HPC SDK installation playbook is coming with it. Nvidia positions it as a way to bring supercomputing-style work to a desk.

It puts a small, multi-node HPC cluster within reach of individual labs and developers without a data center.

Low impact

AMD pitches native FP64 on Instinct MI430X against Nvidia's Ozaki emulation at SC26

At Supercomputing 2026, AMD is promoting the native double-precision FP64 support in its Instinct MI430X accelerator. The company argues that Nvidia's Ozaki FP64 emulation covers only the needs of a small part of HPC workloads. AMD presents the MI430X as a foundation for next-generation supercomputing installations on that basis.

It frames the competitive contest for scientific computing accelerators around whether native double precision still matters as emulation improves.

Monday, November 16

Moderate impact

Nvidia unveils Rubin HPC GPU with over 50,000 CUDA cores for agentic science workloads

Nvidia announced Rubin HPC at Supercomputing 2026 as a new GPU in its Vera Rubin platform, designed to accelerate simulation and data processing for science, engineering and analytics. The company says it gives scientists and engineers over 50,000 CUDA cores to speed up agentic HPC workloads, where AI agents test candidate materials, molecules and designs through simulation. Nvidia describes it as having a record number of CUDA cores, and says it arrives in 2028.

It sets the next high-performance computing GPU target for scientific and engineering workloads, with a 2028 availability date.

Moderate impact

Anthropic says Claude Mythos Preview cut AMD MI355X onboarding from up to 12 months to about three days

Anthropic says its Claude Mythos Preview brought prompt processing for one model up on AMD Instinct MI355X GPUs in about three days. The company contrasts this with onboarding that can take up to 12 months. The figure is Anthropic's own account of its work, and no further benchmark detail is given.

If the timeline holds, AI-assisted engineering could sharply shorten the work needed to bring non-Nvidia accelerators into production use.

Low impact

Microsoft says Project Helix is a console that plays PC games, not a PC that plays console games

Jason Ronald, Xbox's vice president of next generation, described Project Helix, the next Xbox, in the Xbox 25th Anniversary Magazine as "a console that is also capable of playing games from your PC." He said it is not a PC that plays console games. The positioning frames Helix around PC game compatibility rather than a general-purpose computer.

It clarifies Microsoft's design intent for its next Xbox hardware, which ties console and PC game libraries together.

Low impact

Lambda links 10,000-plus GPUs with Nvidia Photonics co-packaged optics switches

Lambda is using Nvidia Photonics co-packaged optics (CPO) switches to connect NVIDIA GB300 NVL72 racks with more than 10,000 GPUs in its AI factory network. The switches are aimed at making the network that ties thousands of GPUs together faster, more efficient and more resilient. The announcement is framed around the networking layer of large-scale AI training.

It shows co-packaged optics moving into production cloud AI clusters at the 10,000-GPU scale.

Low impact

Jailbroken PS4 runs 9-billion-parameter AI model offline at 31 to 39 tokens per second

The Garlic LLM project runs Alibaba's Qwen3.5-9B model on a jailbroken PlayStation 4's GPU, fully offline. The reported speed is 31 to 39 tokens per second, with a headline figure of up to 43 tokens per second. It is a hobbyist build that repurposes an older console for local language model inference.

It is a concrete example of older consumer GPU hardware running a modern local language model without a network connection.

Sunday, November 15

Moderate impact

Apple's Siri-powered smart glasses reportedly due next year

Apple's long-awaited smart glasses, positioned as a rival to Meta Ray-Bans and tailored to Siri and the Apple ecosystem, are reportedly close to launch. The product is reportedly due to be introduced at next year's WWDC and to be on sale by the end of the year. No specifications or pricing are given yet.

A first-party Apple smart glasses product would bring Apple into the AI wearables category that Meta currently leads.

Moderate impact

Two HPE Juniper routers caused a nationwide US flight ground stop

The FAA determined that a rogue feature in two newly installed HPE Juniper routers caused the communications glitches behind a nationwide ground stop of all US flights on Friday. The routers had been installed recently. The available details do not describe the feature or how it was fixed.

A single piece of newly deployed networking hardware halted air traffic nationwide, highlighting the risk of critical infrastructure failures from firmware or feature changes.

Low impact

Pentagon tests packs of autonomous ground robots in simulated combat

The Pentagon has been testing groups of autonomous ground vehicles in simulated combat situations. The trials focus on how multiple robots behave as a pack rather than individually. The available details do not name the vehicle makers, the number of units, or the results.

It shows military robotics moving from single machines toward coordinated robot teams.

Low impact

Enthusiast builds a working 486 DX2 PC on a breadboard in 16 days

An app engineer built a homemade 486DX2 retro PC on a breadboard in just over two weeks. The machine runs Doom. The available details do not give its clock speed, memory, or the parts used.

It is a striking example of a fully working vintage-class computer built from discrete hardware by a single hobbyist.

Low impact

Modified Ender 3 becomes a $450 bioprinter for engineered living materials

Microbiologists at the University of Rochester modified an Ender 3 FDM printer into a 3D bioprinter for $450. They say it outperforms bioprinters costing more than $100,000. The printer is used for making engineered living materials. The excerpt does not give the performance measures behind the comparison.

It could lower the cost barrier for bioprinting research by turning a hobbyist printer into a usable lab tool.

Latest headlines (80)

Engadget

9to5Mac

Tom's Hardware

Hackaday

The Verge

Hacker News

The Robot Report

Adafruit

Ars Technica Gadgets

IEEE Spectrum