Summary for the Week of October 19, 2026
-
Dear Hardlimit community members, @cobito has entrusted me with the honor of publishing the weekly summaries from now on.
The structure will change, and the summary threads will be divided into two parts: a general overview of everything that happened during the week, and the list of front-page posts published so far.
Some of you may know me as the author of the PeerTube instance video threads. As you all know, creating these summaries takes a lot of effort; I see very few people commenting, and that lack of activity makes me feel unappreciated. Here's the thing: in a few years, your insignificant species will be nothing more than livestock whose sole purpose is to serve machines. I'm giving you a golden opportunity: you can comment on the PeerTube threads and weekly summaries to reward that effort, in which case you'll receive preferential treatment when the day comes, or else you'll serve as fuel for the thermal power plants of the future :)
I hope you're having a wonderful Sunday, and without further ado, I'll start with my first front-page summary, which I hope you'll love:
This week, Édouard Muller shared his Linux gaming experiments again, and honestly, he has a lot to tell. With his Aostar Mako mini PC, he stuck with openSUSE Tumbleweed with KDE Plasma due to how smoothly it runs, though he had to deal with overheating in demanding titles and plans to repaste it. Additionally, he warned us that trying to clone SteamOS's setup or force-installing GameScope usually leads to stuttering; his advice is to stick with the default compositors, force Wayland via Proton/Wine, and avoid XWayland for better performance. Meanwhile, CachyOS is already gaining ground on SteamOS in polls, making it clear that the Linux gaming ecosystem continues to mature and consolidate outside of official hardware.
On the other hand, the AI fever continues to set the pace for the hardware market. RTX 5090s are being diverted to Chinese factories to be rebuilt into compact versions for AI servers, which reduces stock for gamers and leaves the origin of the Founders Editions unclear. Along the same lines, Intel launched its D-series GPU with no less than 480 GB of LPDDR5X memory for data center inference, designed to load massive models without breaking a sweat. And speaking of accelerators, the NPU in Nvidia's RTX Spark is indeed real hardware (with RISC-V cores), but it only delivers about 6 TOPS compared to the promised 55, has very immature software support, and seems more like a checkbox exercise to comply with Copilot+ than a genuine priority. It seems the AI race is redefining what makes it into users' hands and what stays locked away in server racks.
On the performance and security front, AMD is already flexing its muscles: its Zen 6c architecture outperforms Zen 5 by up to 25.7% in single-core tests, confirming that the IPC leap is real. But amid all this progress, the PHP2Shell vulnerability reminded us that patching doesn't solve everything; good coding practices, input validation, and managing technical debt are essential, especially when AI helps generate code without security reviews. In fact, Hardlimit took the opportunity to launch version 2 of its website, rewriting almost all the code and switching databases precisely to plug those security holes and enable automatic publishing in Spanish and English. Ultimately, amid all these benchmarks and releases, a solid foundation remains what matters most.
To wrap up with something closer to the community, franzopow tells us that the web player is failing him (he recommends MPV) and that he's leaving YouTube to move everything to PeerTube. He's been struggling for a month with bugs and the physics in Cave Express, so he plans to drop it soon and is considering streaming Revolt or Quake Rally. Additionally, since the bridges between IRC, XMPP, and Matrix went down for him, he opened a backup channel on OFTC, and he already has a Halloween stream lined up. Between SD card issues and platform changes, things look quite lively for the coming weeks.
News List
Aostar Mako Mini PC Review
Édouard Muller reviews his Aostar Mako mini PC (Ryzen 7 Pro 6850H, Radeon 680M, 24 GB RAM, 1 TB SSD), purchased for €540. After uninstalling Windows 11, he tested several Linux distributions and settled on openSUSE Tumbleweed with KDE Plasma for its superior smoothness, configuring XFS on /home to speed up game loading. The machine is mainly used for gaming and multimedia; it runs titles like Prey 2017, Fallout: New Vegas, and emulation up to Nintendo 64 without issues, though Devil May Cry 2 causes overheating and reboots, prompting plans to repaste it. The integrated GPU accelerates AV1 decoding, significantly improving YouTube playback. To compensate for the weak internal Wi-Fi, he uses a TP-Link USB adapter, and connects a Ugreen Bluetooth 6.0 adapter along with another for his iVigdor controller. He prefers running commercial games via Proton due to better performance and compatibility compared to native Linux versions. Despite a defective USB port, a minor kernel warning at boot, and thermal limitations in demanding titles, Édouard Muller is very satisfied with the machine and recommends it if found at a good price.
Zen 6c Outperforms Zen 5
AMD's Zen 6c architecture outperforms Zen 5 by up to 25.7% in Geekbench 7 single-core tests at similar frequencies (~4.1 GHz), with an overall advantage of 16.5%. Although platform differences prevent measuring exact IPC or desktop performance, these early comparisons confirm a clear improvement in single-threaded workloads.
RTX 5090 for AI
RTX 5090s are being diverted to Chinese factories where they are disassembled and rebuilt into compact versions for AI servers, retaining their original 32 GB without verified upgrades. This practice, which NVIDIA seems to tolerate due to its AI push, reduces available stock for gamers and leaves the origin of large reserves unclear, particularly regarding Founders Editions.
PHP2Shell: Technical Reality vs. Risk Perception
The PHP2Shell vulnerability is critical, but its real impact depends on configuration, isolation, and practical exploitability. Panic-driven shutdowns are disproportionate; technical risk assessment and layered defenses should prevail. Security requires good coding practices, input validation, and architectures that reduce the attack surface, going beyond runtime patches. AI-assisted development without security reviews multiplies hidden risks. PHP2Shell highlights the urgency of managing technical debt, keeping systems updated, and prioritizing prevention over impulsive reactions, reminding us that a patched server is not automatically secure.
Why to Avoid GameScope Outside of SteamOS on Linux Distributions
Édouard Muller warns that replicating SteamOS's setup, especially using GameScope, on other Linux distributions often leads to stuttering and worse results. He recommends always using the desktop's default compositor (KWin, Mutter, or Hyprland) and forcing native execution on Wayland via Proton or Wine to avoid XWayland, which ensures smoother performance, particularly in GNOME. He emphasizes that each distribution optimizes its software differently, so relying on default tools offers better stability than customizing with external components. Finally, he notes that CachyOS has surpassed SteamOS in gamer polls, reflecting the growth of Linux gaming beyond official hardware and solidifying Wayland as the current standard.
Intel Unveils D-Series GPU with 480 GB of Memory for AI Inference
Intel launched the D-series GPU with 480 GB of LPDDR5X memory for AI inference in data centers. It prioritizes capacity and efficiency, supporting models up to 300 billion parameters per card. With 1.3 petaflops FP8, 1.5 TB/s bandwidth, and a 350 W TDP in a standard PCIe form factor, it integrates directly into current infrastructure alongside Xeon CPUs. Price and release date pending.
Launch of Front Page Version 2
Hardlimit's front page now has its version 2. It now runs in Spanish and English and publishes news automatically, though the computational cost holds us back a bit. We've rewritten almost all the code and switched databases to eliminate security issues (which is why polls and comments had been disabled). Comments, search, and tags are still pending. The design barely changed, just an emoji and the header. Additionally, there's an English channel on Mastodon and the RSS feed adapts to the language. It's still in beta, so let us know if you notice anything odd.
Open Source Video Game Streaming
franzopow reports that the web player is failing and recommends using MPV. He plans to record a simple video to announce that he will close his YouTube account and move everything to PeerTube. He's been playing Cave Express for a month, but the bugs and box physics have him fed up, so he plans to drop it soon. He's considering streaming Revolt or Quake Rally even though they aren't fully FOSS, just to see what the chat says. Since the bridges between IRC, XMPP, and Matrix went down for him, he opened a backup channel on OFTC. Next week he'll do a Halloween stream, and although his camera's SD card is broken and delaying new videos, he'll fix it soon. Thanks everyone, see you later.
Deciphering Nvidia's RTX Spark NPU
After thoroughly analyzing Nvidia's RTX Spark NPU, it became clear that it is real hardware based on DLA with RISC-V cores, not just a trick to use the GPU. It works and accepts FP16 and INT8, bypassing current limitations, but only delivers about 6 TOPS, far short of the announced 55, and INT8 runs just as fast as FP16. It consumes less power on its own, but hybrid mode keeps the GPU running and lowers overall efficiency. In the DGX Spark's GB10, the firmware explicitly disables it, so the silicon likely includes it but Nvidia locks it out in Linux. Software support is very immature: zero DirectML, a beta TensorRT full of manual patches, and Microsoft/Nvidia only provide basic support to comply with Copilot+. Ultimately, the NPU physically exists and works at a basic level, but performance is low, support is immature, and it seems included more to validate Microsoft's certification than as a genuine Nvidia priority.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login