This Hardware archive tracks the practical side of local AI infrastructure: GPUs, NPUs, mini PCs, edge accelerators, memory bandwidth, and power efficiency tradeoffs that directly impact LLM inference quality. We prioritize benchmark-backed updates and deployment notes useful for real build decisions, from compact home labs to enterprise pilot clusters. Use this stream to compare total cost of ownership, thermal constraints, and model-fit scenarios across current devices, then deepen with our hardware pillar guide and connected LLM coverage.
The open-source Nova driver, written in Rust, continues its integration path into the Linux 7.0 kernel. The goal is to provide support for NVIDIA's Turing GPUs. In parallel, work continues on the Arm Mali Tyr drivers.
Updates in the Intel Xe open-source graphics driver for Linux, in preparation for kernel 7.0. Key changes include disabling GuC Power DCC for Panther Lake and improvements in GPU temperature sensor reporting.
A user has built a local computing cluster based on four Lenovo P620 workstations, each equipped with two NVIDIA RTX 3090 GPUs, for a total of 192GB of VRAM. The configuration, interconnected via a 10Gbit network (awaiting a 100Gbit upgrade), is intended for running large language models (LLMs) locally, leveraging frameworks such as vLLM and Ray.
An analysis of using the Dell Pro Max workstation equipped with a GB10 GPU to solve complex reporting tasks. The original article reports a practical experience that led to a return on investment (ROI) within a 12-month period, focusing on real-world application.
Libcamera 0.7 has been released, a software library for image signal processors (ISPs) and embedded cameras on Linux. The key update is initial support for GPU acceleration within the software ISP (SoftISP), aiming for improved performance compared to CPU-only processing.
Qilimanjaro Quantum Tech has announced EduQit, a modular quantum computing kit designed to enable hands-on training, experimental learning, and early-stage research. EduQit enables universities and research institutions to work directly with a physical quantum computing system, addressing a gap in quantum education that often relies on theory and simulations.
Libgcrypt 1.12 has been released, the low-level library used by GnuPG and other software for cryptography. This release introduces optimizations for AMD Zen 5 CPUs via VAES and AVX-512, promising a performance increase of up to 2x for AES operations.
Timur Kristóf of Valve's Linux graphics team addressed issues in the open-source AMDGPU driver. This allows older AMD GCN 1.0 and 1.1 GPUs to transition to using AMDGPU by default instead of the Radeon driver. New patches overcome limitations of GCN 1.1 APUs, enabling AMDGPU DC support for Kaveri and related products.
NVIDIA-VAAPI-Driver 0.0.15 was released. This VA-API driver, built atop NVIDIA's NVDEC interface, enables video acceleration for NVIDIA GPUs with the Firefox web browser on Linux, supporting VA-API but not NVIDIA's NVDEC.
Cooler Master showcased in China a new 360x360mm AIO (All-in-One) liquid cooler designed for high-performance workstations with CPUs up to 2000W. This cooling system is intended for users who need to dissipate high thermal power.
Moltbot announces progress that could rehabilitate Apple's Mac mini in the development of artificial intelligence applications, overcoming traditional limitations related to the CUDA ecosystem.
Alibaba's T-Head Semiconductor has developed a new AI chip aiming to compete with the performance of Nvidia's A800 and A100 GPUs. This move could intensify competition in the AI hardware market, potentially offering new options for inference and training workloads.
The AI server market is shifting towards rack-level architectures, with Delta and Liteon emerging as key suppliers. This trend reflects the increasing demand for higher compute density and improved thermal management capabilities in AI data centers.
A user has built a high-performance AI server using consumer-grade components, achieving 768GB of memory between RAM and VRAM. The configuration, based on a Threadripper Pro and multiple GPUs, demonstrates how a relatively contained budget can compete with more expensive enterprise solutions, pushing the limits of local AI inference.
The upcoming release of GNOME 50, expected in distributions like Ubuntu 26.04 LTS and Fedora Workstation 44, will feature improved discrete GPU detection within the GNOME Shell. This effort, which has been two years in the making, has finally been merged.
New research suggests thermodynamic computing could drastically reduce the energy consumption of AI in image generation. Prototypes show promise, but the challenge of creating competitive hardware is significant.
The Mesa NVK open-source Vulkan driver has temporarily disabled support for larger memory pages due to a discovered issue. This feature, implemented in the Nouveau kernel driver in Linux 6.19, aimed to improve performance through compression and the use of larger page sizes.
Micron's Fab 10B, now under construction in Singapore, promises to more than double the company's 3D NAND output from the region when it is fully operational in about 10 years.
When budgets are tight for RAM, storage, or GPUs, replacing your PC case can refresh your system. A new case offers modern features and an updated look, representing a lasting investment for future builds.
Phison's Pascari SSDs have set a new reliability benchmark by powering the world's first lunar data center. This milestone demonstrates the ability of SSDs to operate in extreme environments, opening new frontiers for data processing in space.