Nvidia Unveils GPU Roadmap Featuring Rubin HBM4 Memory

Nvidia has recently provided a glimpse into its upcoming chip roadmap, setting the stage for the unveiling of its Blackwell GPU architecture at the Hot Chips conference in the US next week. The company's plans include the introduction of Blackwell Ultra, a version featuring 288Gbytes of HBM3e memory. Additionally, Nvidia is gearing up to launch a new CPU named Vera in 2026, followed by a new GPU architecture called Rubin in 2025, which will utilize next-generation HBM4 memory. The Rubin architecture will also have a variant with enhanced memory capacity known as Rubin Ultra.

"In Blackwell Ultra, we will be increasing the compute capabilities, although the exact details are not being disclosed at this time," stated Dave Salvator, director of accelerating computing at Nvidia. The company is focusing on enhancing both memory and compute performance to meet the demands of evolving technologies and applications.

Looking ahead to 2026, Nvidia is preparing to introduce the Ruben architecture, with plans for a subsequent release of Ruben Ultra. On the CPU front, the transition from the current Grace generation to the Vera CPU architecture is expected to take place in 2026. Nvidia emphasizes the importance of network products in its data center platform strategy, highlighting the interconnected nature of its hardware components.

The JEDEC Solid State Technology Association is nearing the finalization of the HBM4 standard for high-performance memory. HBM4 is poised to offer increased channel count per stack compared to HBM3, with a larger physical footprint. The standard is designed to ensure compatibility with both HBM3 and HBM4, enabling a single controller to work with either memory type. Various configurations will necessitate different interposers to accommodate the diverse footprints.

Nvidia's senior engineers are set to showcase the latest advancements powering the Blackwell platform, along with research on liquid cooling for data centers and AI agents for chip design. The GB200 NVL72 liquid-cooled rack system, connecting 72 Blackwell GPUs and 36 Grace CPUs, represents a significant step towards optimizing performance and energy efficiency in data center environments.