When Arm Meets RISC-V: SiPearl, Semidynamics to Co-Develop Sovereign AI Platform

Why It Matters

The partnership gives Europe a home‑grown AI hardware stack, reducing reliance on U.S. and Chinese suppliers and positioning the region to compete in high‑value inference markets. Its open, rack‑scale design could accelerate adoption of European‑sourced AI infrastructure in cloud and enterprise data centers.

Key Takeaways

• SiPearl's Rhea2 CPU uses DDR memory, dropping HBM for simplicity.
• Semidynamics' RISC‑V accelerators are programmable via Linux, GCC, LLVM.
• Platform targets rack‑scale AI inference with full CPU‑accelerator memory coherence.
• Chiplet integration will use UCIe interconnect, enabling mix‑and‑match designs.
• Design follows Open Compute Project specs for data‑center compatibility.

Pulse Analysis

Europe’s AI hardware ambitions have long lagged behind the United States and China, where integrated CPU‑GPU stacks dominate data‑center deployments. By combining SiPearl’s Arm Neoverse‑V2‑based Rhea2 processor with Semidynamics’ RISC‑V inference ASICs, the partnership creates a uniquely European stack that sidesteps traditional supply‑chain dependencies. The decision to forgo high‑bandwidth memory on the CPU in favor of DDR simplifies coherence with the accelerators, a critical factor for large‑scale inference workloads that demand low latency and predictable power consumption.

Technical depth distinguishes the platform: Rhea2 handles orchestration while the RISC‑V accelerators deliver programmable tensor performance, all tied together with UCIe chiplet interconnects. Full CPU‑accelerator memory coherence, a rare feature in heterogeneous systems, promises a smoother programming model and reduced software overhead. Future generations will push integration to the chiplet level, echoing industry moves by AMD and Intel, but with an open‑source ethos that leverages Linux, GCC, and LLVM rather than proprietary stacks. This approach aligns with the Open Compute Project, ensuring the hardware can slot into existing rack infrastructure without custom engineering.

From a market perspective, the initiative could reshape Europe’s AI supply chain by offering a cost‑effective, open alternative to Nvidia’s CUDA‑centric ecosystem. By targeting inference—particularly large language model serving and emerging agentic AI workloads—the platform addresses the segment where efficiency and total cost of ownership matter most. If the rack‑scale solution delivers performance‑per‑watt comparable to leading global players, European cloud providers and enterprises may adopt it to meet data‑sovereignty mandates while staying competitive on price and scalability.