Nvidia Broadens CUDA Reach with Python 1.0, RISC-V Support, and Open-Source Advances

Nvidia released CUDA 13.3, bringing CUDA Python 1.0 to developers to provide full platform access from Python. [5] This release includes cuda.core 1.0.0, which turns basic vocabulary like devices, streams, and buffers into ordinary Python objects to give GPU libraries a common foundation. [5] The release also packages cuda.bindings 13.3.0, which provides low-level one-to-one bindings to the CUDA C APIs. [5]

Alongside the Python updates, Nvidia introduced the CUDA Core Compute Libraries (CCCL) runtime as a modernized C++ alternative to the traditional CUDA runtime. [3] The CCCL runtime leverages modern C++ features through collections of headers such as <cuda/stream>, <cuda/buffer>, and <cuda/launch>. [3] Both the new CCCL runtime API and the traditional CUDA runtime API are built on top of the CUDA driver API. [3]

At the Hot Chips 2026 conference, Nvidia announced that CUDA will officially support RISC-V CPUs as its third supported host architecture alongside x86 and ARM. [6] Nvidia is partnering with SiFive to demonstrate this integration on an upcoming high-core-count server chip at the conference. [6] The hardware requirements mandate that chips implement the RVA23 CPU profile, comply with the RISC-V Server SoC specification, and support PCIe coherency. [6]

Because of these strict enterprise requirements, current consumer RISC-V hardware like single board computers will not run CUDA. [6] This strategic move targets server systems, acknowledging that China currently accounts for approximately 50 percent of global RISC-V shipments. [6]

The open-source Vulkan driver NVK recently gained experimental support for Nvidia's DLSS upscaling technology on Linux. [4] This capability functions through a Vulkan extension called VK_NVX_binary_import, which allows applications to load pre-baked Nvidia CuBIN files onto the GPU. [4] The NVK driver can only run DLSS where compatible bytecode already exists, as it lacks a method to translate Nvidia's intermediate assembly into the intermediate representation Mesa drivers compile from. [4]

NVK began as a from-scratch Vulkan driver in 2022 and supports Nvidia's Turing and newer architectures. [4] In late 2024, NVK became the first open-source Vulkan driver for Nvidia hardware to pass Khronos conformance. [4]

Nvidia is aggressively expanding CUDA's accessibility across software languages, open-source graphics stacks, and alternative CPU architectures. By supporting RISC-V, Nvidia adds a third host architecture alongside x86 and ARM, positioning itself to serve regions that heavily adopt the open-source instruction set. Providing native Python abstractions through CUDA 13.3 reduces developer reliance on third-party libraries like PyTorch or CuPy for basic GPU memory management. Meanwhile, experimental open-source driver support for DLSS suggests growing flexibility in Nvidia's previously rigid Linux ecosystem. What the sources don't address: Whether Nvidia plans to eventually support consumer-grade RISC-V hardware for CUDA developers outside of the enterprise server market.