The Future of AI is Modular: Why the SiFive-NVIDIA Milestone Matters

The incredible potential of AI is currently limited by a major bottleneck: data movement. In today's systems, the processing speed of the GPU is often held back by the interconnects and the efficiency of how data moves between the CPU, the accelerator, and the rest of the system. For the most popular NVIDIA-based systems the high-performance specialized NVLink interconnect technology has recently been expanded to include custom CPUs and accelerators. Announcements have come from Intel and Arm as they join the NVLink Fusion partner program, but their solutions still limit data center customers’ choices to proprietary ISAs with limited flexibility. Until now, the benefits of open standards and the customizability of RISC-V were not an option. We are thrilled to be able to change that.

Last week, SiFive and NVIDIA announced a major milestone: we are working together to integrate NVIDIA NVLink™ Fusion into SiFive’s high-performance RISC-V data center solutions. This collaboration brings the flexibility of SiFive RISC-V to the world's most popular AI ecosystem, offering customers choices that were previously impossible under restrictive ISAs.

This isn't just about adding another name to a partner list; it is about bringing true architectural freedom to the AI data center.

Data Movement Challenges Call for Customization

In this rapidly evolving landscape of AI infrastructure, the traditional separation between processing units is becoming a problem. As models grow in complexity, data movement has emerged as the critical bottleneck, often consuming more time and energy than the computation itself. To solve this, system architects require super-tight integration between the CPU, the GPU/AI accelerator, and the rest of the system components.

NVIDIA today provides fully integrated solutions that tightly integrate their custom CPUs with their GPUs (for example, their recent announcement of the Vera CPU coupled with the Rubin GPU). However, some customers want to pair their own specialized, custom CPU architectures with NVIDIA’s world-class GPUs to meet specific workload requirements or power constraints.

Recognizing this demand for flexibility, NVIDIA has broadened NVLink Fusion to allow select third-party chipmakers to integrate their own designs directly with NVIDIA’s ecosystem via high-bandwidth, low-latency links. For the first time, customers can pair NVIDIA GPUs and other products with their own custom CPUs using SiFive’s RISC-V based IP. This partnership directly answers many of our customers' requirements, providing the ability to combine NVIDIA’s acceleration with the custom architectural freedom that only SiFive RISC-V IP can provide.

Why This Partnership Changes the Game

SiFive represents a uniquely compelling choice. By leveraging RISC-V open standards, SiFive allows architects to build a highly tailored, "right-sized" CPU that can be co-designed with the NVLink interface. RISC-V is rapidly gaining share across every market segment, from the smallest sensors to the largest supercomputers. In the data center specifically, companies globally are now increasingly turning to RISC-V to power their future CPU needs. The reason is simple: companies want to make their hardware and software investments around a compute architecture and business model that doesn’t lock them in or restrict what they might do in the future. The open standard of RISC-V is enabling more and faster innovation across the ecosystem.

SiFive is the leading provider of RISC-V IP with billions of customer chips shipped based on our designs. The SiFive P870-D is the 3rd generation of our out-of-order CPU design. It is the leading RISC-V solution for servers and is now in silicon with our lead customers. We are continuing to invest in our high-performance CPU roadmap and will be working with NVIDIA over multiple generations of our technology.

Lastly, because SiFive provides the architectural blueprints as custom IP, rather than a finished chip design, this collaboration will enable NVLink Fusion across multiple end customers, each building unique products optimized for their specific niche in the AI market.

The Future is All About Flexibility

Ultimately, this announcement is about expanding customer choice. AI compute is not a "one-size-fits-all" problem.

We know our customers solve these challenges in different ways:

• Custom Accelerators: Some customers want to build their own AI accelerators. For them, SiFive provides a range of options based on our Intelligence products, either acting as the control and assist units to their own accelerator designs, or by building a scalable SiFive XM Accelerator.
• NVIDIA Ecosystem: For customers who want to leverage NVIDIA’s market-leading GPUs, combining NVLink Fusion with SiFive’s RISC-V CPUs enables them to build a more efficient, integrated system than was previously possible with legacy ISAs.

The world is turning to RISC-V for its flexibility, power density, and open standard platform. SiFive is proud to pioneer this integration, bringing the industry’s most advanced interconnect to the world’s most flexible instruction set. Together, we are moving past the era of restrictive ISAs and into a future defined by unprecedented architectural freedom.