Nvidia recently provided a glimpse into its upcoming chip roadmap, offering insights into its Blackwell GPU architecture and future plans at the Hot Chips conference in the US. The company is set to introduce the Blackwell Ultra version featuring 288Gbytes of HBM3e memory, alongside a new CPU named Vera in 2026 and a cutting-edge GPU architecture called Reuben in 2025. The Reuben architecture will utilize next-generation HBM4 memory, with a high-memory variant known as Reuben Ultra also in the pipeline.
According to Dave Salvator, director of accelerating computing at Nvidia, the Blackwell Ultra will not only enhance memory capacity but also boost compute capabilities. While specific details are yet to be disclosed, Nvidia is gearing up to deliver significant advancements in the Blackwell Ultra model.
The JEDEC Solid State Technology Association is nearing the finalization of the HBM4 standard, which promises high-performance memory solutions. HBM4 is expected to double the channel count per stack compared to HBM3, with a larger physical footprint. The standard aims to ensure compatibility by enabling a single controller to work seamlessly with both HBM3 and HBM4 configurations, albeit with varying interposer requirements.
Nvidia's senior engineers are poised to showcase the latest innovations powering the Blackwell platform, including 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, exemplifies Nvidia's commitment to cutting-edge technology.
Furthermore, Ajay Tirumala and Raymond Wong, directors of architecture at Nvidia, will offer a comprehensive overview of the platform, highlighting how these advancements synergize to set a new standard for AI and accelerated computing performance while enhancing energy efficiency. Ali Heydari, director of data center cooling and infrastructure at Nvidia, will delve into various designs for hybrid-cooled data centers, showcasing innovative solutions to optimize energy consumption and operational costs.