Cody Ho and Niklas built a fully OpenGL ES 3.0 compliant GPU driver for Apple's M4 Mac Mini and MacBook Neo in about one month by reverse engineering the AGX firmware ABI and user-space components using a hypervisor. The driver achieves high performance, running Minecraft at 200fps and supporting WebGL in browsers, and involved clean-room development techniques without accessing Apple binaries.
A browser-based drone video editor that lets users draw property lines, automatically track boundaries using GPU-powered CoTracker, correct tracking with editable keyframes, and export styled videos with animated labels—keeping all processing local except for the tracking computation.
The author describes running a local coding agent using Oh My Pi with Qwen3.8-27B on two RTX 3090s. Key optimizations include adjusting thinking budgets, token limits, and subagent concurrency to achieve practical inference speeds. Local setups offer privacy and cost predictability but require careful tuning and accept slower inference compared to hosted frontier models like Claude or GPT.
OrthoRoute is a GPU-accelerated PCB autorouter plugin for KiCad that uses Manhattan lattice routing and the PathFinder algorithm to handle high-density boards with thousands of nets. Developed to route a complex 17,600-pad backplane that existing autorouters couldn't handle efficiently, it leverages KiCad's new IPC plugin system for improved performance and flexibility compared to traditional Python-based routers.
Neurogrid is a community-owned cloud platform that runs language models on underutilized GPUs from distributed users worldwide, offering affordable inference by leveraging existing consumer hardware instead of centralized data centers.
Nvidia introduced the RTX Pro 5500 Blackwell workstation GPU with 84 GB of GDDR7 memory and advanced Tensor Cores for professional AI workloads including LLM inference, generative AI, and computer vision tasks.
ImageArm is a native macOS app that compresses images up to 80% without visible quality loss using multiple competing compression tools and GPU Metal acceleration. It offers drag-and-drop functionality, headless mode for scripts, Finder integration, and is free and open source.
Nvidia's RTX 5090 graphics card has vanished from official online retail channels in the US, with third-party sellers now charging $6,500–$9,500, up from $4,299 in June. Inventory depletion is driven by AI server demand, though scams and inflated pricing plague the secondary market, with Micro Center remaining a safer in-person alternative.
Fly.exe is a neural simulator that runs a complete male fruit fly brain connectome (165,122 neurons and 25.5 million synaptic connections) in a physics-based virtual environment on a single GPU. Twelve simulated flies with embodied neural circuits navigate a virtual arena with food and obstacles, demonstrating goal-directed behavior driven purely by their connectome-based neural networks without explicit steering rules.
A technical handbook explaining FlashAttention, an optimization technique that accelerates transformer attention mechanisms through tiling, online softmax, and recomputation without approximating the mathematical function. The key insight is that wall-clock speed depends on GPU memory traffic rather than FLOP count alone, achieved by reducing expensive reads and writes to high-bandwidth memory.
Nvidia announced the RTX Pro 5500 Blackwell workstation GPU with 84 GB of GDDR7 memory and advanced AI capabilities. The card is designed for professional AI workloads including LLM inference, AI agents, generative AI, and computer vision applications.
In September 2026, NVIDIA announced CUDA Rust, enabling native GPU kernel programming in Rust compiled directly to PTX. The language supports two programming models: SIMT (traditional thread-based) and Tile (newer, architecture-agnostic), with cuda-oxide providing a custom rustc backend for compilation. NVIDIA is expanding Rust adoption across its AI systems stack including drivers and serving infrastructure.
NVIDIA CUDA Tile support has been successfully ported to Java through TornadoVM, enabling tile kernels to be chained with JIT SIMT kernels in a single task graph. The implementation includes a new tornado-api tile package with working JVM fallbacks, C++ tile kernels compiled to CUDA 13 cubins using tensor cores, and comprehensive compiler integration for tile operations with support for compute capability 8.0 and toolkit 13.3.
Pragtical 3.12, a lightweight Lua-scriptable code editor, introduces a modular rendering system with an optional SDL-based GPU backend while maintaining its 50MB RAM and 10MB install footprint. The GPU backend targets Vulkan, Direct3D 12, and Metal for cross-platform support, offering benefits for high-DPI displays and large files, though it trades some memory overhead for efficiency gains. The release also adds trust prompts for project modules to prevent arbitrary Lua execution.
A comprehensive total cost of ownership analysis compares local LLMs against cloud APIs across three usage tiers for 2026, accounting for hardware, electricity, labor, and depreciation costs beyond token pricing. The article models 12-month and 36-month scenarios using mid-2025 hardware prices and published API rates from OpenAI, Anthropic, and Google to identify break-even points for solo developers, startups, and engineering teams.
A review of the LattePanda Sigma, an industrial single-board computer featuring an Intel Core i5-1340P processor, 32 GB LPDDR5 memory, and extensive I/O options including Thunderbolt 4 and dual 2.5GbE ports. The author tests its hardware capabilities and explores GPU inference using the Xe iGPU with llama.cpp.
MOLT is a thermally aware, memory-efficient QLoRA fine-tuning tool for consumer NVIDIA GPUs on Windows, enabling local language model fine-tuning with hardware telemetry and thermal controls. Version 0.12.0 includes experimental optimizations for low-overhead update attribution and thermal pacing, with installation via a single PowerShell command and support for Windows 10/11 with Python 3.12.
ZLUDA is a drop-in replacement for CUDA that enables unmodified CUDA applications to run on non-NVIDIA GPUs with near-native performance, providing an alternative to NVIDIA's proprietary GPU computing platform.
A browser-based tool that compares your GPU performance against data from the Steam Hardware Survey, allowing PC gamers to see how their graphics card ranks globally. The tool runs locally without uploading any data.
Beam is an open-source serverless runtime for AI workloads offering fast cold starts, GPU support, autoscaling, and a Pythonic interface for deploying containerized applications with minimal infrastructure overhead.