The patent, published last week, describes an opt-in system, so this isn't something Microsoft would include as a background Windows task, and it could result in cash payments.

Microsoft is seeking to patent technology that uses idle gaming PCs to train AI models in exchange for some form of digital currency.

Filed in 2025 by Microsoft Technology Licensing and published last week, the patent describes distributing AI training tasks in a system similar to Folding@Home, in which a global network of powerful gaming PCs could use their idle GPU cycles for AI training, PC Games Hardware reports.

The patent, titled Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization (US-20260349820-A1), describes an opt-in system, so this isn't something Microsoft would include as a background Windows task, at least not in its currently described form. Whenever that system ran idle, it could automatically be looped in to completing portions of larger AI training workloads. The system would then track completed work and issue credits that could be redeemed for use with the resulting AI model, free software, or even cash.

The patent describes a responsive system that tracks utilization of key components like memory, CPU, and GPU, as well as temperatures and power draw, and searches for an ideal time when the GPU is not in use and, in particular, is running cool. The last thing Microsoft wants is its distributed AI training system setting anyone's PC on fire. Although RTX 50 GPU owners may need to be more careful there.

Users will be able to manually start or stop jobs, too, so it's not like the system would prevent you from gaming while it churns away on AI training. You'd also be able to set a usage limit to prevent overuse of energy or hardware. Although GPUs don't necessarily degrade with heavier use like some electronics, I'm not sure whether AI training would be covered by your warranty.

 

Sorry if this has been posted before but couldn't found it in my searches. (It's from September 25)

The new systemd-report collects system facts and runtime metrics into unified, timestamped JSON reports for monitoring and fleet management.

systemd is getting a new reporting subsystem designed to collect both static system information and dynamic runtime metrics in one place.

Called systemd-report, the new tool gathers data from multiple metric provider services via a generic Varlink interface and combines it into a single timestamped JSON report. The work targets system and fleet management, where administrators need a consistent view of many machines.

The collected information is split into two groups. The first covers relatively static system “facts,” such as CPU architecture, hostname, kernel version, OS identification, installed RAM, CPU count, virtualization technology, TPM information, and other hardware or platform details.

The second covers changing runtime metrics like CPU and memory use, load averages, disk I/O, swap usage, system pressure, cgroup statistics, network interface states, systemd unit states, restart counters, queued jobs, and high-priority journal messages.

According to Lennart Poettering, the underlying metrics service was initially contributed by Meta with a focus on dynamic runtime information. The same interface was later extended to handle more static system facts as well, something Amutable is using for areas such as operating system deployment management.

Once gathered, systemd-report compiles all of this information into a consistent snapshot representing the state of the machine at a particular point in time.

The subsystem is designed for remote fleet management. Reports can be uploaded directly to an HTTPS server using PUT requests, letting organizations send system information to a central control plane. The setup includes periodic uploads via a systemd timer and reports triggered by events like system boot, reboot, or shutdown.

For security, reports can be cryptographically signed before being sent, and systemd currently provides three signing methods.

The simplest uses a software-generated asymmetric key stored locally. Systems with TPM hardware can produce a TPM quote covering platform measurements and the report data. A third option supports confidential-computing environments by generating a CPU TSM quote that ties the report to measurements of the secure virtual machine.

This makes systemd-report different from traditional monitoring tools like Prometheus and its node_exporter. While both expose similar information, systemd obtains data directly from lower-level OS components instead of relying on a separate scraping model.

More importantly, systemd-report focuses on complete, signed, timestamped snapshots rather than independent streams of metrics. Poettering notes that the two approaches are not mutually exclusive, and the new infrastructure could eventually provide the basis for integration between them.

The architecture is also intentionally extensible. Third-party projects can provide additional metrics or signing mechanisms simply by exposing the appropriate Varlink socket. These providers can be written in any programming language and only need to implement a small interface.

In other words, systemd-report is not just another command for displaying system information. It is shaping up as a broader reporting framework built directly into systemd, combining system facts, runtime metrics, secure signing, and remote reporting in a format suitable for managing everything from individual Linux machines to large server fleets.

For additional details, see Poettering’s announcement.

[–] 4 points 3 days ago (7 children)

I see, didn't know about dimming zones. If it's related to HDR, probably none of the devices we have have that though.

The best screens for energy saving are e-ink screens but they are way too expensive for PC monitors. I hope they get cheaper in time, I would love to replace my vertical monitor with an e-ink one.

  • source
  • parent
  • context
  •  

    Alternative Imgur version for the main image

    Previously I was using Samsung N150P (netbook) as my home server and it served me around 6 years. >> https://programming.dev/post/708584

    Though after upgrading to Debian 13, additional 90 packages were installed just with the upgrade. I also had a little adventure with it :) >> https://programming.dev/post/708488

    Anyway, this new mini PC I got is from 2016 and surprisingly it has 8GB DDR3 RAM (I'm rich now). It's generally in good condition. Even though it's Celeron N3000, this one doesn't have hyperthreading so only 2 cores. Still it's a real upgrade over Atom N450 CPU.

    Since last year, I was planning to get a Raspberry Pi and moving from there. However thanks to AI overlords it got so expensive I had to bail from that fantasy. I got this for 40$ and with 8GB RAM I think it was a good deal.

    I was already planning to change my home server to Alpine for a while but I just couldn't touch a perfectly fine system because lazy. So this time I went with Alpine and I already love how small it is. I just installed a couple monitoring tools and packages like sudo along with Pi-hole so currently it has quite few packages.

    Don't worry though, I'm planning to use that RAM, with time. :)

    view more: next ›