In a surprising move at Computex 2024, NVIDIA’s CEO Jensen Huang revealed the company’s next-generation GPU architecture, codenamed Rubin.
This announcement, initially nestled within an AI-focused keynote, signifies a strategic acceleration in NVIDIA’s product development, positioning the company to release new GPU products annually.
Named after Vera Rubin, an American astronomer renowned for her groundbreaking research on dark matter and galaxy rotation rates, the Rubin architecture symbolizes NVIDIA’s commitment to pushing technological boundaries.
Rubin’s pioneering work on galaxy rotation rates reshaped our understanding of the universe’s composition, making her a fitting namesake for NVIDIA’s forward-thinking technology.
NVIDIA’s Fast-Paced Innovation: From Blackwell to Rubin
This rapid pace of innovation follows NVIDIA’s recent unveiling of its Blackwell platform. While the first iteration of Blackwell GPUs (B100/B200) is set to grace data centers later this year, NVIDIA isn’t resting on its laurels.
A supercharged Blackwell Ultra variant is in the pipeline, featuring enhanced memory capabilities with 12Hi memory stacks across 8 sites, a substantial upgrade from the 8Hi stacks in current products. This chip is expected to launch in 2025, showcasing NVIDIA’s commitment to continually enhancing its current product lines.
But the real star of the show is the upcoming Rubin GPU series. The NVIDIA Rubin R100 GPUs, part of the R-series lineup, are slated for mass production in the fourth quarter of 2025. High-end systems like DGX and HGX solutions are expected to follow in the first half of 2026.
NVIDIA confirms that Rubin GPUs will harness the power of HBM4 memory, with initial models featuring 8 HBM4 stacks. An even more powerful Rubin Ultra, boasting 12 HBM4 stacks, is planned for 2027, promising a substantial boost in memory capabilities and overall performance.
Cutting-Edge Tech: HBM4 and Advanced Packaging
Technically, NVIDIA’s Rubin R100 GPUs are expected to use a 4x reticle design, a step up from Blackwell’s 3.3x. They’ll be manufactured using TSMC’s CoWoS-L packaging technology on the N3 process node.
TSMC’s ambitious plans for up to 5.5x reticle size chips by 2026, featuring a 100x100mm substrate, could allow for up to 12 HBM sites. While TSMC is also planning an SoIC design with over 8x reticle size in a 120x120mm package, a 4x reticle size for Rubin GPUs seems more realistic in the near term.
Powering these GPUs will be next-generation HBM4 DRAM. NVIDIA currently uses HBM3E memory for its B100 GPUs and plans to refresh these chips with HBM4 variants when they become widely available in late 2025, aligning perfectly with the R100 GPUs’ mass production schedule. Both Samsung and SK Hynix have plans to develop HBM4 in 2025, offering up to 16-Hi stacks.
Vera CPU: NVIDIA’s Latest ARM Processor
NVIDIA isn’t just focusing on GPUs. They’re also upgrading their CPU line with the introduction of the Vera CPU, named as part of the new Vera Rubin board.
This development continues NVIDIA’s superchip series, following the Grace Hopper and Grace Blackwell superchips. While specific details about Vera’s architecture remain under wraps, it’s expected to integrate with NVIDIA’s latest hardware innovations.
The GR200 Superchip module will house two R100 GPUs alongside an upgraded Grace CPU based on TSMC’s 3nm process. This is a significant step up from the current Grace CPU, built on TSMC’s 5nm process with 72 cores (144 in the Grace Superchip solution). The name “Vera” for the next-gen ARM CPU solution is a thoughtful touch, harmonizing with the Rubin architecture’s namesake.
NVIDIA’s future hardware will also benefit from enhanced connectivity. The new NVLink 6 switches promise data transfer speeds up to 3600 GB/s, while the CX9 SuperNIC will offer network speeds up to 1600 Gb/s. These advancements will significantly boost connectivity and throughput for complex AI and data center tasks.
Focus on Power Efficiency and NVIDIA’s Accelerated Tech Roadmap
One of NVIDIA’s primary focuses with the Rubin R100 GPUs is power efficiency. Acknowledging the growing power demands of data center chips, NVIDIA is committed to providing substantial improvements in this area while simultaneously enhancing the AI capabilities of its chips.
Although the R100 GPUs are still on the horizon, with an unveiling likely at next year’s GTC, NVIDIA’s roadmap reveals a company that’s not just keeping pace with the industry but actively accelerating it.
From the Rubin GPU architecture to the Vera CPU and advanced networking solutions, NVIDIA is laying the groundwork for significant leaps in AI and data center technology. As these innovations materialize, they promise to reshape our computational landscape, much like Vera Rubin’s work reshaped our understanding of the universe.