Nvidia is reportedly collaborating with Taiwan Semiconductor Manufacturing Company (TSMC) to develop next-generation GPUs using advanced chiplet technology. This partnership is anticipated to be crucial for Nvidia’s upcoming “Rubin” architecture, the rumored successor to the current Blackwell generation.
This move towards a chiplet-based design represents a significant shift from traditional monolithic GPU structures. Chiplets offer several advantages, including improved performance, scalability, and cost-effectiveness. By assembling multiple smaller semiconductor dies into a single package, manufacturers achieve better yields and lower production costs.
This approach is gaining traction in the semiconductor industry, especially as chip designs become more complex and traditional scaling methods encounter limitations. By leveraging TSMC’s advanced manufacturing processes, Nvidia can potentially boost the power efficiency and processing capabilities of its future GPUs, making them ideal for demanding applications like AI, data centers, and high-performance computing.
It’s rumored that Nvidia’s Rubin GPUs will be manufactured using TSMC’s cutting-edge N3P process node. This optimized version of TSMC’s 3nm technology provides improved performance, power efficiency, and transistor density compared to previous generations. The N3P node is specifically designed to maximize the benefits of chiplet architectures, enabling Nvidia to push the boundaries of GPU performance while maintaining energy efficiency.
To further optimize performance, Nvidia will utilize TSMC’s advanced packaging technologies, including System-on-Integrated Chip (SoIC). This technology allows for vertical stacking of chips, improving power efficiency and reducing latency between chiplets within the GPU. AMD has already employed this design in its 3D V-Cache CPUs.
TSMC plans to significantly expand its SoIC capacity by the end of 2025. Nvidia’s upcoming Rubin lineup is expected to utilize an SoIC design and HBM4 memory. The Vera Rubin NVL144 platform is rumored to feature a Rubin GPU with two reticle-sized dies, delivering up to 50 PFLOPs of FP4 performance and 288 GB of HBM4 memory. The higher-end NVL576 model is expected to incorporate a Rubin Ultra GPU with four reticle-sized dies, achieving 100 PFLOPs of FP4 performance and boasting 1TB of HBM4e memory across 16 HBM stacks.
Nvidia’s adoption of chiplet technology aligns with a broader industry trend. Leading semiconductor companies like AMD and Intel have already incorporated similar designs into their processors. The modularity of chiplets allows manufacturers to combine different processing units, optimizing performance for specific workloads. With the increasing demand for powerful hardware driven by AI and high-performance computing, the collaboration between TSMC and Nvidia is poised to deliver groundbreaking advancements in GPU design.
This partnership holds significant promise for the future of GPUs. As the demand for high-performance computing continues to grow, the innovative solutions arising from this collaboration will likely shape the landscape of the semiconductor industry.
