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.
SanDisk has relaunched its Optimus SSD line with PCIe 5.0 models in 2TB and 4TB capacities. The new Optimus GX Pro 8100 are available starting at $999 for the 2TB model and $1799 for the 4TB version, representing a 5% price increase over previous models. Older WD Black versions remain a cheaper alternative.
Intel Panther Lake handheld chips are reportedly delayed until Q2 2026. The alleged 'Core G3' series might launch alongside new Arc B360 and Arc B380 iGPUs.
Venture capital firm Benchmark Capital has announced a $225 million investment in Cerebras Systems, a manufacturer of processors dedicated to artificial intelligence. Benchmark has been an investor in Cerebras since 2016, supporting the development of alternative solutions to Nvidia's GPUs.
Microsoft unveiled Maia 200, a chip designed to optimize AI inference costs. The goal is not to compete directly with Nvidia, but to offer a more cost-effective solution for specific workloads. The chip is intended for Microsoft data centers.
GPUs and accelerators use specialized engines for matrix multiplication (GEMM). This article analyzes the precision of accumulators in these engines, revealing that, for hardware efficiency reasons, the effective precision may be lower than expected. A method for verifying accumulator precision using Triton is presented, with results showing how precision can vary depending on the hardware and configuration.
Amazon's new Kindle Scribe Colorsoft is a color e-ink tablet designed for reading, annotating documents, and taking notes. Despite the hefty price tag, it could be a worthwhile investment for those seeking a dedicated device for these activities.
A lucky shopper found an incredible deal at Walmart, purchasing SSDs worth $1,300 for just $210. The haul included WD, Samsung, and PNY drives, offering significant savings on high-performance storage.
New Linux benchmarks examine the performance of Intel's Panther Lake Core Ultra X7 358H CPU with a higher power budget. The tests reveal significant generational improvements, particularly in energy efficiency, and confirm the excellent performance of the Intel Arc B390 GPU with open-source drivers.
AMD continues the development of its LLVM compiler stack for future GPUs. A new target, GFX1170, also identified as RDNA 4m, has been introduced. This update adds to the ongoing work on GFX1250 and GFX13 targets, expanding support for AMD's upcoming graphics architectures.
Qualcomm is making it easier to use Snapdragon X1 Elite on Linux. Previously, necessary firmware files had to be fetched from the Windows 11 on ARM partition. Now, QUPv3 firmware bits have been integrated into the linux-firmware.git repository, greatly simplifying the process for Linux users.
MetaOptics, headquartered in Singapore and maintaining close ties with Taiwan, is developing heat-resistant metalenses for integration into CPUs. This technology could significantly improve the thermal management of processors.
A user reported a significant throughput increase, up to 26 tokens/second, using the Qwen3-Coder-Next-Q4_K_S model with llama.cpp on an RTX 5090. The optimization was achieved by offloading MoE expert tensors to the CPU and quantizing the KV cache.
South Korea is making significant investments in artificial intelligence, supported by a hardware infrastructure powered by over 260,000 Nvidia GPUs. This strategic move aims to position the country as a leader in the AI sector, with a focus on advanced applications and research.
Intel's Xe graphics driver for Linux, starting with kernel 7.0, will re-enable D3cold support for Battlemage GPUs. This feature was disabled due to instability issues in power state transitions. The change will not apply to all systems, excluding specific NUCs.
A Reddit user benchmarked the Strix Halo's iGPU, testing various software configurations with 13 LLM models and 15 different llama.cpp builds. The aim was to evaluate the impact of ROCm, Vulkan, and various compilation options on inference performance.
Tenstorrent announced a reduction in the number of Tensor cores on its Blackhole p150 cards, from 140 to 120, via a firmware update. The company anticipates a 1-2% performance drop for existing users. New cards will ship with 120 Tensor cores.
Western Digital announces the development of 3.5-inch HDDs (Hard Disk Drives) based on HAMR (Heat-Assisted Magnetic Recording) technology with a capacity reaching 140 TB, thanks to the use of 14 platters. This technology promises to significantly increase the storage density of hard drives.
Nvidia is reportedly developing DLSS 4.5, an advanced version of its upscaling technology that could eliminate the need for denoisers in ray tracing. This is thanks to a Transformer model that reconstructs ray-traced reflections more accurately.
First Linux benchmarks of the Intel Arc B390 GPU, integrated in high-end Panther Lake models. The Xe3 graphics card, equipped with 12 Xe cores, promises interesting performance in desktop and mobile environments for graphics and compute workloads.
ASRock has issued a statement regarding a new round of Ryzen 9000 CPU failures. The company says it is actively collaborating with AMD to identify and resolve the cause of the problem, which appears to affect a limited number of units.