NVIDIA Unveils Next-Generation AI GPU Architecture and Enhanced CPUs

At the esteemed Computex 2024 event, NVIDIA shared exciting details about their future generations of AI GPU architectures, specifically mentioning the much-anticipated Rubin. This new technology is said to be inspired by the trailblazing contributions of notable astronomer Vera Rubin, who significantly advanced our understanding of dark matter and the dynamics of galaxy rotation.

Rubin and Blackwell: Powering AI's Future

Before the tech community has even fully experienced NVIDIA's forthcoming Blackwell AI GPUs, which include the B100 and B200 models slated for release later this year, NVIDIA CEO Jensen Huang is already generating buzz with the announcement of the Rubin GPU architecture. The Blackwell GPUs, with their substantial enhancements, will boast 12-Hi HBM3E memory stacks across eight sites, in contrast to the previous generation's eight. The release of these more powerful AI GPUs is expected in 2025.

Looking ahead, the Rubin R100 AI GPU will be part of NVIDIA's new R-series. It's set to enter mass production in the final quarter of 2025, with related DGX and HGX AI systems expected in early 2026. Moreover, NVIDIA asserted that the Rubin AI GPUs would be partnered with platforms available in 2026, with a robust Rubin Ultra AI GPU variant lined up for 2027. Further distinguishing these next-gen GPUs, NVIDIA confirmed the adoption of advanced HBM4 memory technology for the Rubin R100.

Technological Advances in Design and Packaging

The anticipated NVIDIA Rubin R100 AI GPUs will showcase a 4x reticle size, made possible through TSMC's state-of-the-art CoWoS-L packaging technology and crafted on the new N3 process node. TSMC's ambition extends to creating chips with up to 5.5x reticle size by 2026, potentially accommodating 12 HBM sites on a new, distinct 100 x 100mm substrate. This compares to the current 8 HBM sites seen on 80 x 80mm packages.

NVIDIA also aims to incorporate TSMC's SoIC design, allowing for packages that exceed even 8x reticle sizes, with aspirations for a 120 x 120mm package configuration. However, analysts currently anticipate about a 4x reticle size for the first Rubin R100 AI GPUs.

Vera CPUs: A Nod to Innovation

NVIDIA's development horizon also includes an advanced Grace CPU for the GR200 Superchip module, pairing it with two R100 AI GPUs and leveraging TSMC's next-gen 3nm process. This is an upgrade from the current 5nm process of the Grace CPUs that feature 72 cores, or 144 cores in total for the Superchip. In a complementary homage, the next-generation Arm-based CPU solution has been coined Vera, echoing the Rubin name and paying tribute to the same scientific pioneer.NVIDIA, Rubin, Blackwell, Vera, AI, GPU, CPU