A new Entrepreneur analysis compares three local AI hardware options for developers: NVIDIA's DGX Spark, built on 128GB unified memory and the Grace Blackwell GB10 chip, offers the most mature CUDA-based software compatibility but has slower memory bandwidth at 273 GB/s. The piece frames local AI machines as tools for data privacy, cost control and predictable performance rather than replacements for cloud-based large-scale training.
entrepreneur.com
· 2026-10-23
A Hugging Face contributor and Arch Linux package maintainer built a bootable Arch Linux ARM setup for the ASUS ProArt P16 H7607BA-SE033W, a laptop with 128GB of RAM, 6,144 CUDA cores and a 2TB SSD that ships with official Windows-only support. The build incorporates NVIDIA's Ubuntu kernel work, custom ACPI overrides, a patched hid-asus driver for the keyboard, and official NVIDIA and ASUS firmware, with 7-Zip performance matching Windows and the OS correctly reporting the full 116GiB unified memory pool to CUDA.
josie.lol
· 2026-10-19
Stealthium published research on Rank-One Model Editing (ROME) attacks, a technique where malicious Python code alters a model's weights directly in GPU memory without touching disk files. The company ran a ROME attack in its lab and showed its runtime sensor catching the tampering by linking CUDA kernel launches to the Python code that triggered them, alongside ongoing work to trace weight changes back to their originating call stack.
stealthium.io
· 2026-10-19
rGPU is a newly shown project that lets a PyTorch program send tensor operations over TCP to execute on a remote NVIDIA GPU while the application itself keeps running locally. It offers two integration routes: a simple PyTorch 'rgpu' device for programs that opt in, and a CUDA shim that intercepts existing Linux CUDA binaries using libcuda, CUDA Runtime, cuBLAS, cuBLASLt, and cuDNN. Setup involves installing via pip and using an SSH tunnel through the rgpu-run command to connect to the GPU host.
github.com
· 2026-10-07
A developer released TurboGPT, an open-source (MIT-licensed) CUDA C++ implementation for training tiny byte-level GPT models. The project's Windows build instructions and CLI usage are documented, and a sample run on a dataset called hn1g reportedly reached 2.52435 bits-per-byte after 1.5 billion training tokens.
github.com
· 2026-09-29
A Show HN post introduces a tool that automatically generates, tests, benchmarks and refines CUDA GPU kernels from a workload description. It uses an LLM agent via LangGraph to propose code and launch-configuration changes, a C++/NVRTC harness to compile and run candidates, and Python to validate outputs against NumPy references, keeping only the fastest kernel that passes all correctness checks. The project was developed on Windows with an RTX 3060 Laptop GPU and requires Python 3.12+, CMake, a C++17 compiler and an OpenAI API key.
github.com
· 2026-09-25
Bend is a newly introduced programming language that compiles to native code, runs on multiple CPU cores or GPUs without manual threading, and includes a built-in proof checker similar to Lean and Rocq. Developers can write rules in a LAWS.bend file, and Bend will refuse to accept any AI-generated code change that violates those rules, checking each change in about a second. Its creators demonstrate this by showing an AI agent attempt to break a game's win condition, which Bend blocks until the AI produces a compliant fix.
bend-lang.com
· 2026-09-17
Nvidia announced it is extending CUDA to support GPU kernels written natively in Rust, compiled directly to PTX rather than wrapped around another language. The move covers both the existing SIMT model used in CUDA C++ and Python, and the newer Tile programming model, with development continuing into 2027.
developer.nvidia.com
· 2026-09-16
Nvidia has introduced the RTX Pro 5500 Blackwell Workstation Edition, a professional graphics card built on the same GB202 die as the GeForce RTX 5090 and offering similar CUDA core counts and performance. The key difference is memory: the RTX Pro 5500 carries 84GB of GDDR7, 2.6 times more than the RTX 5090, positioning it between the existing RTX Pro 5000 and RTX Pro 6000 in Nvidia's professional lineup.
tomshardware.com
· 2026-09-14
An independent developer known as Speedstu built a PowerShell-based toolkit that combines ZLUDA with AMD's HIP/ROCm SDK, allowing CUDA-only applications to run on AMD GPUs in Windows without dual-booting or virtual machines. The setup successfully mapped CUDA's driver API and libraries like cuBLAS, cuSPARSE, and cuFFT to AMD equivalents, and was tested by training a small reinforcement-learning model on a Radeon RX 9060 XT using unmodified CUDA code.
tomshardware.com
· 2026-09-14
Developers behind the ZLUDA project have released a reproducible stack combining ZLUDA v6-preview.69 with AMD's HIP SDK 6.4, allowing CUDA-targeted Windows applications, including LibTorch-based machine learning workloads, to run on AMD hardware. The setup has been validated specifically on the Radeon RX 9060 XT (gfx1200), where core CUDA libraries like cuBLAS and cuFFT passed checks and a PPO reinforcement learning network completed training over 65,536 timesteps.
github.com
· 2026-09-13
A technical blog post walks through building a custom attention kernel for Nvidia's Blackwell B200 GPU entirely from scratch, using CUDA and some PTX code. Through a 14-step progression of optimizations illustrated with 60 diagrams, the kernel reaches 94.4% of FlashAttention-4's performance on 4K, 8K, and 16K sequence shapes, and is demonstrated by plugging it into a video-generation model.
iaroslavelistratov.github.io
· 2026-09-02
GoKawiil