NVIDIA Expected To Mass Produce Next-Gen R100 “Rubin” GPUs In Q4 2025: TSMC N3, 8 HBM4 Stacks, 3nm Grace CPU & Focus on Power Efficiency

NVIDIA is expected to mass produce its next-generation Rubin R100 GPUs with HBM4 memory on the TSMC 3nm node by Q4 2025.

NVIDIA’s Next-Gen Rubin R100 GPUs To Focus On Power Efficiency While Increasing AI Performance, To Use HBM4 Memory & TSMC 3nm Node

The new information comes from TF International Securities analyst, Mich-Chi Kuo, who states that NVIDIA has laid the groundwork for its next-generation Rubin R100 GPUs named after Vera Rubin, an American astronomer who made significant contributions to the understanding of dark matter in the universe while also pioneering work on galaxy rotation rate.

Kuo states that the NVIDIA Rubin R100 GPUs will be part of the R-series lineup and are expected to be mass-produced in the fourth quarter of 2025 while systems such as DGX and HGX solutions are expected to be mass-produced in the first half of 2026. NVIDIA recently unveiled its next-generation Blackwell B100 GPUs which feature a monumental increase in AI performance and are the first proper chiplet design from the company that first laid its foundation in the Ampere GPU.

Four years ago, we split GA100 into two halves that communicate through an interconnect. It was a big move – and yet barely anyone noticed, thanks to amazing work from CUDA and the GPU team.

Today, that work comes to fruition with the Blackwell launch. Two dies. One awesome GPU. https://t.co/XuaUQPskkM pic.twitter.com/svRKhwPYEn

— Bryan Catanzaro (@ctnzr) March 18, 2024

It is expected that NVIDIA’s Rubin R100 GPUs will use a 4x reticle design (versus 3.3x of Blackwell) and will be made using the TSMC CoWoS-L packaging technology on the N3 process node. TSMC recently laid out plans for up to 5.5x reticle size chips by 2026 which would feature a 100x100mm substrate and allow for up to 12 HBM sites versus 8 HBM sites on current 80x80mm packages.

The semiconductor company also plans to move to a new SoIC design which will feature a greater than 8x reticle size in a 120x120mm package configuration. These are still being planned out so we can more realistically expect somewhere between 4x reticle size for Rubin GPUs.

Other information mentioned states that NVIDIA will be utilizing the next-generation HBM4 DRAM to power its R100 GPUs. The company currently leverages the fastest HBM3E memory for its B100 GPUs and is expected to refresh these chips with HBM4 variants when the memory solution gets widely mass-produced in late 2025. This will be about the same time when R100 GPUs are expected to enter mass production. HBM4. Both Samsung and SK Hynix have revealed plans to commence development of the next-gen memory solution in 2025 with up to 16-Hi stacks.

NVIDIA is also set to upgrade its Grace CPU for the GR200 Superchip module which will house two R100 GPUs and an upgraded Grace CPU based on TSMC’s 3nm process. Currently, the Grace CPU is built on TSMC’s 5nm process node and packs 72 cores for a total of 144 cores on the Grace Superchip solution.

One of the biggest focuses for NVIDIA with its next-gen Rubin R100 GPUs will be power efficiency. The company is aware of the growing power needs of its data center chips and it will provide significant improvements in this department while increasing the AI capabilities of its chips. The R100 GPUs are still far away and we shouldn’t expect them to be unveiled until next year’s GTC but if this information is correct, then NVIDIA definitely has lots of exciting developments ahead for the AI and Data Center segment.

NVIDIA Data Center / AI GPU Roadmap