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 introduced the RTX PRO 5500 Blackwell graphics card with 84 GB of GDDR7 memory and 21,760 CUDA cores, matching the RTX 5090's core count but offering 2.6x more VRAM. Designed for workstation and enterprise deployment, it supports agentic AI, physical simulation, and graphics workloads with a 600W TDP and memory bandwidth of 1398 GB/s.
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.
Kache v0.21.0 expands caching for C/C++ compilation, adding support for whole-program links, preprocessor output, Windows clang-cl flags used by Firefox, and single-source CUDA nvcc compilation. New features include machine-level configuration defaults, trust domain isolation for persistent stores, and improved telemetry.
Kairo is a research workbench for LLM inference routing on RTX 5090 GPUs that measures workloads, validates results, and selectively enables optimizations like CUDA Graphs only when proven beneficial. The system uses fail-closed policies, promoting only exact measured workload buckets into production runtime decisions while avoiding premature global defaults.
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.
Neve is a deep learning framework implemented in 1400 lines, demonstrating neural network training on CIFAR-10 with GPU acceleration. The project includes ResNet implementations and comparisons with PyTorch, running on CUDA 12.3 compatible GPUs via Docker.
A Windows project enables CUDA-targeted applications to run on AMD GPUs using ZLUDA and ROCm/HIP compatibility layers. The stack has been validated on AMD Radeon RX 9060 XT and successfully runs LibTorch workloads including neural-network inference and reinforcement-learning training, though compatibility depends on which CUDA APIs and libraries each application uses.
soup-connectome is a portable runtime for executing the fruit-fly connectome across CPU, CUDA, WebGPU, and WASM using sparse graph streaming and deterministic LIF neuron semantics. The project converts MaleCNS connectome data (211,577 neurons, 24.7M edges) into portable .scx artifacts and demonstrates reproducible execution across multiple backends with explicit device planning.
BioNeMo Inference Runtime (BioIR) is NVIDIA's GPU-accelerated library for protein structure prediction, converting FASTA/MSA inputs to PDB/mmCIF outputs using AlphaFold-lineage models. It requires Linux with NVIDIA GPU, Python 3.12, CUDA driver 580+, and is available via PyPI with precompiled kernels for supported architectures including H100, A100, and GB200.
Cohere introduces a serving engine for North Mini Code using a decode megakernel, achieving 1.25–1.41× speedup over vLLM on H100 GPUs by consolidating multiple kernel launches into a single persistent kernel that runs the entire forward pass, eliminating GPU idle time during autoregressive decoding.
A C++ pointer cannot distinguish host from device memory, creating a type system vulnerability where GPU data accessed from the host compiles but crashes at runtime. Vx fixes this by encoding memory space directly in the type signature, making unsafe memory accesses a compile-time error instead.
A FreeBSD user successfully ran generative AI workloads using Linuxulator and an external eGPU by leveraging FreeBSD's Linux binary compatibility layer, installing NVIDIA drivers, Rocky Linux 9.7 in a chroot environment, Python 3.10, and CUDA libraries to enable GPU acceleration for local AI tools.
In September 2026, NVIDIA announced CUDA Rust, enabling developers to write GPU kernels natively in Rust compiled to PTX. The announcement includes two programming tracks—SIMT and Tile—with cuda-oxide providing a custom rustc codegen backend that routes Rust code through MIR and LLVM to GPU compilation, while Tile is recommended for architecture-agnostic code.
AutoUVM is an automated prefetching framework that optimizes GPU memory management for large language models operating under memory oversubscription. It bridges deep learning frameworks and NVIDIA UVM by enabling fine-grained, tensor-level prefetching without requiring code modifications, achieving up to 3.1x speedup over baseline UVM and significantly reducing page faults.
An open-source protocol for migrating Large Language Model KV-Cache states across datacenters, achieving 81.73ms latency and reducing GPU compute overhead by 95% compared to standard re-computation methods. The zero-copy transport eliminates prompt re-computation during session handoffs, enabling faster Time-To-First-Token latency.