AI
vLLM Surges as Open-Source Engine for High-Speed LLM Serving
UC Berkeley-born vLLM surges as an LLM inference platform, achieving high throughput with PagedAttention and continuous batching.
Tags: GPU
Gadgets
No GPU Upgrade Needed as Upscaling and Tweaks Lift FPS
Engadget details how to boost FPS on older GPUs with DLSS, FSR, frame generation and settings optimization without upgrading.
Dev
Valve Revives Decade-Old AMD GPUs for Linux Gaming
Valve's Kristof moves GCN 1.0/1.1 to AMDGPU, fixing display and power management to improve Linux gaming.
Dev
Mesa Starts Shiva Common Vulkan Framework, Targets Results in Two Years
Development begins on Shiva, a shared base for Vulkan drivers in Mesa, aiming to ease the growing Vulkan support burden.
Dev
What Is CUDA? How It Works and How It Compares to ROCm, ZLUDA
Explaining CUDA's architecture and toolkit, its differences from AMD's ROCm and the ZLUDA compatibility layer, and how to choose.
Gadgets
NVIDIA GPUs Pack 10-40 RISC-V Cores for Control Tasks
Each NVIDIA GPU includes 10-40 RISC-V cores for non-graphics control. Explaining the transition from Falcon.
Dev
Bend Arrives: A Fast Language That Blocks AI Mistakes with Proofs
Bend blocks AI mistakes via proof checking, with C-like speed and GPU parallelization.
Gadgets
eGPUs Work but Have Major Limitations; Buying a PC Is the Smarter Choice
eGPUs can boost laptops but face bandwidth, cost, and compatibility issues. We break down the trade-offs based on measured data.
Dev
Intel Introduces PISA, Enables LEO by Default on Nova Lake
Intel releases compute stack 26.35.39758.10 with LEO default on Nova Lake S and initial PISA support.
Dev
Mesa 26.2.3 Released with Support for AMD GFX1171
Mesa 26.2.3 released with AMD GFX1171 support and bug fixes for RadeonSI, RADV, Intel ANV and more.
Dev
Running CUDA on AMD GPUs Demonstrated on RX 9060 XT
An attempt to run CUDA assets on AMD GPUs in Windows by linking ZLUDA and HIP has emerged. Demonstrated on the RX 9060 XT.
Gadgets
Modified RTX 5090 96GB Appears on Alibaba for $3,888
A Chinese firm lists a modified RTX 5090 with 96GB VRAM on Alibaba for $3,888, tripling memory with a custom PCB.
Gadgets
Moore Threads Falls 20% on Lock-up Expiry, Domestic GPUs Face Headwinds
Moore Threads plunged 20% on its lock-up expiry day, highlighting a correction in domestic GPU stocks and earnings structure challenges.
AI
Nvidia Expects 70% Growth Next Year, Huang Explains
CEO Huang expects 70% revenue growth next fiscal year, citing demand visibility and supply-chain insight.
Gadgets
DLSS vs. FSR vs. XeSS Compared! Differences Explained in the 2026 Edition
Comparing DLSS, FSR, and XeSS mechanisms, image quality, and supported GPUs for 2026, with selection guides by use case.
Dev
AMD Expands Rust Across Entire GPU Stack, Forms Elite Team
AMD forms elite team to expand Rust across its entire GPU stack, from firmware to drivers and compilers, as NVIDIA makes similar moves.
Gadgets
DLSS Unlocked Brings DLSS 5 and Multi-Frame Generation to Older RTX Cards
Integrated mod DLSS Unlocked enables DLSS 5 and multi-frame generation on older RTX cards, aiming for up to 6x boost with FSR 3.1.
Gadgets
4K Is Optimal for Gaming: Why You Shouldn't Choose 5K
Although 5K is 77% sharper than 4K, its benefits for gaming are limited. We explain why 4K is optimal considering GPU load and price.
Gadgets
RTX 5090 Surges to $5,000 as All Blackwell Models See Price Hikes
RTX 5090 hits $5,000 in North America, 2.5x its MSRP and over double in Europe. All Blackwell models average 60% price surge.
Gadgets
RTX 5090 Sold in a Bundle with Eight Motherboards
A Taiwanese e-commerce site sells RTX 5090 bundles with multiple motherboards, CPUs, and mid-range GPUs. The structural changes in the PC DIY market are highlighted by memory component price hikes driven by AI demand.
Gadgets
HBF Opens the TB Era for GPU Memory, Surpassing HBM Through NAND Stacking
High-Bandwidth Flash (HBF) developed by Sandisk and SK Hynix stacks 16 layers of NAND flash, achieving 512GB capacity and 1.6TB/s bandwidth. While expected as an alternative to HBM, it still faces fundamental challenges in write endurance and latency.
Dev
Linux 7.3 to Provide Stable Support for Intel Nova Lake S Integrated GPU
In the Linux 7.3 kernel cycle, next-generation Intel Nova Lake S processors' integrated graphics will be treated as stable. Previously requiring the force_probe parameter, it will now be enabled by default.
Dev
Intel Arc Alchemist to Enable HuC Media Acceleration via Xe Driver
A patch has been submitted to add HuC media acceleration to the Xe kernel driver for Intel Arc A-Series (Alchemist), expanding GPU encoding/decoding capabilities beyond the i915 driver.
Gadgets
MSI and Colorful Raise RTX 50-Series Prices by Up to 59% in China
MSI and Colorful have revised distributor prices for Nvidia RTX 50-series GPUs in China, with some models seeing a price hike of over 75% compared to MSRP.
AI
AMD ROCm.AI Aims to Surpass CUDA in AI Code Optimization
AMD unveils its new AI-optimized platform "ROCm.AI," leveraging frontier models for automatic tuning to break through CUDA's dominance in GPU programming.
Dev
AMD ROCm Shifts to a Six-Week Release Cycle
At the "Advancing AI" event, AMD announced a shift in the ROCm release cycle from quarterly to every six weeks, with the next release set for August, featuring new AI software.
Gadgets
The Environmental Impact of GPUs: Evaluating Their True Value Amid Rising AI Demand
The environmental consequences of GPUs driving the AI boom are under scrutiny, including power and water consumption, and e-waste issues.
AI
Running Qwen3.6-27B on a Home Server: Complete Guide to Building a Local LLM
Ryō Igarashi's detailed report on Zenn covers everything from choosing the Qwen3.6-27B-FP8 model to calculating required VRAM, selecting GPUs, and setting up an inference server.
Gadgets
Unannounced RTX 50 Super GPUs Listed with TGP on Seasonic Calculator
NVIDIA's RTX 5080 Super, RTX 5070 Ti Super, and RTX 5070 Super were spotted on Seasonic's PSU calculator with 10–17% higher TGPs than existing models. No official announcement yet from NVIDIA.
AI
Meta to Sell AI Computing Resources Amid Stalled Model Development, Shifts Focus to GPU Business
Meta is considering a new initiative, "Meta Compute," to open its AI computing resources to external clients, marking a strategic pivot to monetize GPUs.
Gadgets
RTX 3060 12GB Relaunched at $329.99: MSI Ventus 2X OC
The MSI GeForce RTX 3060 Ventus 2X 12GB OC is back on sale for $329.99 at Newegg. Its 12GB VRAM is ideal for modern gaming and running local AI models, making it a notable restock in the GPU-starved market.
Dev
Nova Lake S Adds 7 GPU IDs to Linux 7.3
PCI device IDs for Intel Nova Lake S’s integrated Xe3P GPU have been added to the Linux 7.3 kernel. Seven IDs are now recognized, and new support includes resolving HuC firmware dependencies.
Gadgets
AMD RX 7900 XTX Engineering Sample Found in Second-Hand Market
A red PCB engineering sample of the RX 7900 XTX was discovered in the second-hand market. Equipped with debug connectors like JTAG and I2C, its specs resemble the RX 7900 GRE, offering a rare glimpse into the development process of RDNA 3.
Dev
MAX Models Now Support Apple Silicon GPUs, Covering All Generations Through M5
Modular expands MAX framework support for Apple silicon GPUs in its latest release. Models for text and image generation now run from M1 to M5.
Gadgets
Cooling an RTX 3060 with an Ice Maker: A DIY Innovation That Hits 22°C
YouTuber TrashBench successfully modified a countertop ice maker to cool an RTX 3060, reducing gaming temperatures by up to 62% and achieving a remarkable 22°C while running Cyberpunk 2077. The project involved repurposing a beer fridge thermostat and battling water leaks.
Gadgets
Orbital Data Centers Hit Physical Barriers in Cooling and Cost
At NVIDIA's GTC, space computing was declared, but physics and cost pose major hurdles. A detailed analysis of technical challenges, including heat dissipation in space and radiation effects.