RISC-V chip designer SiFive will integrate Nvidia's proprietary NVLink interconnect into its CPU cores, marking the latest defection from the open UALink standard that aims to challenge Nvidia's datacenter dominance.
The open ISA movement suffered a significant setback this week as SiFive, the premier RISC-V chip design company, announced it will integrate Nvidia's proprietary NVLink Fusion technology into its processor IP. This move represents a broader industry consolidation around Nvidia's interconnect strategy and casts serious doubt on the viability of UALink, the open alternative backed by AMD, Intel, Amazon, and other industry heavyweights.
What Happened
SiFive designs CPU cores based on the RISC-V instruction set architecture, licensing them to chip manufacturers similar to Arm's business model. The company's processors span from edge devices to datacenter applications. With this announcement, SiFive will now offer NVLink Fusion compatibility in its high-performance compute subsystems, allowing customers to build custom CPUs that can seamlessly connect to Nvidia's GPU ecosystem.
NVLink Fusion, which Nvidia opened to partners last year, enables heterogeneous computing systems where CPUs and GPUs share memory and operate as a unified accelerator. The interconnect delivers up to 3.6 TB/s of chip-to-chip bandwidth, far exceeding traditional PCIe connections. Until recently, Nvidia kept this technology exclusive to its own products.

The Legal and Competitive Landscape
This announcement follows a pattern of industry defections to Nvidia's proprietary standard. Intel, Arm, Fujitsu, and Qualcomm have all embraced NVLink Fusion. Intel plans client systems using the technology to connect its CPU chiplets to Nvidia GPU dies. Qualcomm acquired Ventana to pursue RISC-V alternatives on Arm architecture.
The contrast with UALink is stark. UALink was conceived as an open, royalty-free alternative to NVLink, backed by the UALink Consortium including AMD, Intel, Amazon, Meta, Microsoft, and Arm. The goal was to prevent vendor lock-in and create a standardized way for different vendors' chips to communicate in AI clusters.
However, UALink faces mounting challenges:
Technical delays: AMD's MI455X-based Helios system, one of the first UALink implementations, required tunneling the protocol over standard Ethernet due to a lack of dedicated UALink switches.
Competing standards: Broadcom, initially a UALink supporter, is now pushing its own Scale Up Ethernet (SUE) fabric, fragmenting the open interconnect ecosystem.
Industry momentum: With SiFive and other major players choosing NVLink Fusion, UALink risks becoming a standard without sufficient vendor support.
Impact on Users and Companies
For datacenter operators: The choice between open standards and proprietary performance becomes more difficult. NVLink Fusion offers proven, high-performance interconnects but locks customers into Nvidia's ecosystem. UALink promises vendor neutrality but lacks mature implementations.
For chip designers: Companies licensing SiFive's RISC-V cores can now build CPUs that integrate directly with Nvidia's AI infrastructure. This reduces development time but may limit future flexibility if they want to switch GPU vendors.
For the open hardware movement: RISC-V was meant to provide an open alternative to proprietary ISAs like x86 and ARM. By embracing Nvidia's proprietary interconnect, SiFive risks undermining the open philosophy that attracted many RISC-V supporters.
What Changes
SiFive's customers can now design custom CPUs using SiFive cores that include native NVLink Fusion support. According to SiFive CEO Patrick Little, the company has "a number of datacenter design licenses and several have taped out," though specific customers remain undisclosed due to standard IP licensing practices.
The broader industry shift toward heterogeneous, co-designed systems means open CPU architectures increasingly require advanced interconnects to remain relevant in AI datacenter computing. This creates tension between the open ISA philosophy and the practical need for high-performance GPU integration.
The Technology Behind the Decision
NVLink Fusion operates through two primary configurations:
- Partner CPUs + Nvidia GPUs: Standard configuration where third-party CPUs connect to Nvidia GPUs
- Nvidia CPUs + Partner XPUs: Allows customers to connect their custom AI accelerators to Nvidia's Grace or Vera CPUs
SiFive is focusing on the first configuration, enabling customers to build custom CPUs that pair with Nvidia's AI infrastructure. The company claims this delivers "exceptional efficiency at datacenter scale."
The interconnect's 3.6 TB/s bandwidth is crucial for AI workloads where massive datasets must move between CPUs and GPUs without bottlenecks. Traditional PCIe 5.0 offers only 128 GB/s bidirectional bandwidth, making NVLink essential for cutting-edge AI training.
Industry Implications
This development suggests that in the race to build AI supercomputers, performance and ecosystem integration trump open standards. Nvidia's strategy of opening NVLink Fusion while maintaining control over the core technology appears to be working.
For UALink to survive, the consortium must deliver production-ready switches and demonstrate clear advantages over both NVLink Fusion and Broadcom's SUE. The standard's backers need to show that open governance can match Nvidia's execution speed.
The situation mirrors earlier format wars like VHS vs. Betamax or HD-DVD vs. Blu-ray, where technical superiority often loses to market momentum and ecosystem support.
Looking Ahead
SiFive's announcement signals that the window for an open interconnect standard may be closing. If major CPU IP vendors continue choosing NVLink Fusion, UALink could become a niche standard for specific use cases rather than a universal alternative.
This consolidation around proprietary technology raises questions about long-term innovation and vendor lock-in in the AI hardware space. While customers gain access to proven, high-performance solutions, they lose the ability to mix and match components from different vendors freely.
The next year will be critical for UALink's viability. AMD's Helios system launch will serve as a real-world test of whether the standard can deliver on its promises. Meanwhile, Nvidia's continued expansion of NVLink Fusion partnerships suggests the company is winning the interconnect war before UALink fully enters the market.
For now, SiFive customers have a clear path to Nvidia's ecosystem, but the cost may be the open hardware future that RISC-V was meant to enable.

Comments
Please log in or register to join the discussion