CUDA Expands to RISC-V: What It Means for GPU Compute and AI Workloads

NVIDIA has long been the go-to choice for high-performance GPU compute, especially in machine learning applications. Historically, CUDA, NVIDIA's parallel computing platform and API model, supported only x86-64 and aarch64 CPUs. However, at Hot Chips 2026, NVIDIA announced plans to extend CUDA support to RISC-V. This move is significant for both the RISC-V ecosystem and GPU compute practitioners.

The Requirements for RISC-V in CUDA

To ensure that RISC-V CPUs can seamlessly integrate with CUDA, NVIDIA has outlined a set of requirements. These requirements are designed to ensure that RISC-V platforms meet the performance and reliability standards necessary for server-grade applications.

• RVA23 CPU: The baseline requirement is an RVA23 CPU, which is a high-performance variant of the RISC-V architecture.
• Server SoC and Platform Specifications: Adherence to RISC-V’s server SoC and server platform specifications is mandatory. These specs include:
  • RAS (Reliability, Availability, and Serviceability): Features that ensure the system can operate reliably in a server environment.
  • Specialized Security Processor: A dedicated processor for handling security tasks, enhancing overall system security.
  • Baseline Features: Other essential features like memory protection, virtualization support, and more.

NVIDIA's requirements go beyond these specifications to address specific software challenges:

• Vector Extensions: Support for vector extensions is crucial. Predication in vector operations helps avoid branches, which can improve performance by reducing the overhead of conditional statements.
• ACPI (Advanced Configuration and Power Interface): Initially a challenge, ACPI support for RISC-V was added in 2025 by the UEFI forum. The RISC-V BRS (Boot and Runtime Services) specification includes ACPI, which is essential for hardware discovery and management.
• PCIe Coherency: Ensuring memory ordering consistency between the CPU and GPU is critical. PCIe coherency helps avoid issues where data written by the CPU but sitting in a cache can lead to incorrect DMA operations.

Under the Hood

NVIDIA's decision to support RISC-V is driven by the growing adoption of RISC-V in server environments. Here’s a deeper look at how these requirements impact both hardware and software:

• Performance Enhancements: Vector extensions, in particular, are key for performance. Predication allows for more efficient vector operations by eliminating unnecessary branches, which can significantly speed up compute-intensive tasks like machine learning.
• Software Portability: ACPI support is crucial for ensuring that CUDA software can discover and manage hardware resources effectively. This is especially important for power management, thermal control, and performance tuning.
• Memory Coherency: PCIe coherency ensures that data consistency is maintained between the CPU and GPU, which is essential for reliable and high-performance compute operations.

What to Watch

The extension of CUDA support to RISC-V opens up new possibilities for developers and system architects. Here are a few key points to watch as this development unfolds:

• Ecosystem Growth: As more RISC-V platforms meet NVIDIA’s requirements, we can expect to see a growing ecosystem of compatible hardware and software.
• Performance Benchmarks: Early benchmarks will be crucial in understanding how well RISC-V-based systems perform with CUDA. This will help developers make informed decisions about adopting this technology.
• Developer Tools: NVIDIA is likely to release updated developer tools and documentation to support RISC-V development. Keeping an eye on these resources will be essential for early adopters.

In the broader context, this move aligns with trends in computational memory and AI infrastructure. For instance, Apple’s application of generative AI algorithms to accelerate custom chip design highlights the importance of efficient compute platforms. Similarly, the 7 Pillars of Agentic AI emphasize the need for autonomous, collaborative, and aligned intelligent agents, which can benefit from the performance gains offered by RISC-V and CUDA integration.

As NVIDIA continues to push the boundaries of GPU computing, the inclusion of RISC-V is a significant step towards a more diverse and robust compute ecosystem.