Linux kernel 7.2 is set to integrate Cache Aware Scheduling support, a long-awaited feature designed to optimize performance on processors equipped with multiple last-level caches. This integration, expected before the merge window opens, marks a significant step for Linux system efficiency, particularly for workloads demanding careful hardware resource management in on-premise environments.
Team Group has reached a $1.1 million settlement in a false advertising lawsuit. The dispute concerns T-Force Xtreem ARGB DDR4-3600 CL14 memory modules, whose advertised speeds required BIOS tweaks and overclocking settings to be achieved. This incident raises questions about the transparency of hardware specifications, a crucial aspect for decision-makers evaluating on-premise infrastructures for AI/LLM workloads.
SpacemiT has released the first benchmarks of its K3 SoC, featuring X100 RISC-V cores and RVA23 compliance. This platform, also available in a Pico-ITX form factor, represents one of the first readily available RVA23 RISC-V solutions for use with operating systems like Ubuntu 26.04 LTS. The introduction of RISC-V hardware like the K3 is relevant for companies exploring on-premise alternatives for AI workloads, offering new options for data control and sovereignty.
AMD is set to launch its Ryzen AI Halo PC, a desktop system featuring 128GB of system memory and priced at $3999. This configuration aims to support local AI and LLM workloads, offering a solution for enterprises seeking greater data control and reduced latency compared to cloud options.
Neurosoft Bioelectronics has secured $7.5 million in an oversubscribed seed funding round, bringing its total funding to over $20 million. The Swiss company develops scalable, minimally invasive brain-computer interfaces (BCIs) using proprietary soft electrodes. Its long-term goal is to build a neural data platform supporting cortical foundation models, aiming to enhance BCI performance. The new funds will advance clinical milestones and US commercialization.
An in-depth analysis of Qwen3.6 Large Language Models performance on an RTX 5080 16GB GPU reveals surprising results. The benchmark, focused on on-premise deployment scenarios, highlights how the 35B MoE model achieves 56 tokens/second with a 128k context. Contrary to expectations, the Multi-Token Prediction (MTP) feature in llama.cpp does not offer significant advantages in this scenario, suggesting different optimizations to maximize efficiency on local hardware.
Alibaba has unveiled the Zhenwu M890 AI processor, a multi-year silicon roadmap, and the new Qwen 3.7-Max LLM. This strategic move aims to build an integrated AI stack, focusing on AI agents and technological sovereignty, thereby reducing reliance on external vendors. The massive investment and existing deployment experience highlight a long-term commitment to the sector.
Alibaba T-Head, the semiconductor division of the Chinese tech giant, is intensifying its commitment to developing dedicated artificial intelligence infrastructure. The introduction of the Zhenwu M890 marks a significant step in this direction, aiming to enhance computing capabilities for AI workloads. This move highlights the increasing importance of proprietary hardware solutions to manage the computational demands of LLMs and generative AI, with implications for on-premise and hybrid deployment strategies.
ASML is set to deliver its first High-NA EUV lithography systems within months. This technology is crucial for producing next-generation chips, essential for advancing artificial intelligence and Large Language Models. However, concerns about the costs associated with this innovation raise questions about the TCO for on-premise AI infrastructures.
Alibaba's T-Head chip unit has revealed the Zhenwu M890, a new GPU-class AI chip. This development marks a significant step in China's strategy to establish domestic hardware alternatives to NVIDIA's products, amidst increasing export controls. The company stated that the chip is already in scaled mass production.
At Tech Forum 2026, liquid cooling emerged as an indispensable solution for AI servers. The increasing power density of AI hardware stacks, essential for Large Language Models, is driving the adoption of more efficient thermal dissipation technologies. Dr. Clyde Chu of META Green Cooling Technology highlighted this transition, emphasizing the importance of addressing thermal challenges in on-premise deployments.
Automakers are increasingly collaborating with Taiwan to accelerate the development of advanced vehicle electronics. This strategic move reflects the growing demand for on-board artificial intelligence computing capabilities, ADAS systems, and infotainment. The choice of Taiwan underscores the island's crucial role in the global semiconductor supply chain, which is fundamental for innovation and the deployment of on-premise and edge AI solutions.
A user recently expressed plans to upgrade their VRAM from 32GB to 48GB for local LLM workloads. This move highlights the critical importance of video memory for on-premise Large Language Model deployments, where hardware capacity is a key limiting factor. The article explores the trade-offs and technical considerations for managing LLMs on self-hosted infrastructures, analyzing how VRAM influences model selection and performance.
The introduction of the Nvidia H200 GPU, a critical component for demanding AI workloads, faces a complex geopolitical landscape, particularly export restrictions to China. This situation raises significant questions for companies planning on-premise Large Language Model (LLM) deployments, affecting hardware availability, costs, and data sovereignty strategies.
The surging demand for high-performance memory, driven by the expansion of artificial intelligence, is creating a global supply squeeze. According to Korean media, this dynamic is weakening the bargaining power of companies like Apple, which face higher costs and reduced availability for crucial components. This phenomenon highlights the infrastructural and supply chain challenges companies must navigate to support the development and deployment of AI solutions.
Google announced its new "audio glasses" at IO 2026, smart glasses designed for voice interaction. Users can issue verbal commands to access Google's ecosystem of apps and services, including the Gemini platform. This initiative marks an expansion of conversational AI into wearable devices, highlighting the trend of bringing advanced AI capabilities directly to end-users, albeit with a cloud-based backend.
A recent leak has unveiled details about the PCB of Intel's Xe3P data center GPU, codenamed "Crescent Island." The board will integrate 160GB of LPDDR5X memory, distributed across 20 x 8GB modules, with an estimated bandwidth between 704 and 760 GB/s. This strategic choice by Intel aims to mitigate current HBM memory shortages, offering an alternative solution for on-premise LLM workloads.
AMD has released details of its new EPYC 8005 processor series, based on the Zen 5 architecture. Featuring configurations of up to 84 cores and a 225W TDP, this line represents a significant evolution for servers, offering a balance between core density and energy efficiency. The series is set to bolster capabilities for on-premise deployments, supporting intensive workloads and addressing data sovereignty and TCO requirements.
AMD has unveiled further details on its EPYC 8005 "Sorano" processor series, succeeding the EPYC 8004 "Siena". With SKUs ranging from 8 to 84 cores, these new chips are crucial for companies planning on-premise AI workload deployments, offering greater control and TCO optimization. The availability of full specifications now allows for in-depth evaluation for self-hosted architectures.
Amprius Technologies, a specialist in high-energy-density silicon anode batteries, has announced a partnership with Matternet, a certified drone delivery company. The collaboration involves Amprius supplying its SiCore® lithium-ion cells for Matternet's autonomous delivery drones, aiming to significantly extend their operational range. This advancement is crucial for the efficiency of edge systems and distributed AI.