NVIDIA's "Blackwell" series of AI GPUs has just begun to ship, and its next-generation architecture is already on the horizon. Codenamed "Rubin" and named after Vera Rubin, the new architecture will power NVIDIA's future AI GPUs, making a generational leap in performance, but more importantly, with a design focus on reducing power consumption.
This will become especially important because NVIDIA's current architecture is close to kilowatts, so the computing center will not be able to scale indefinitely. Mich-Chi Kuo, an analyst at Tianfeng International Securities, said that Nvidia's first "Rubin"-based AI GPU, the R100, is expected to enter mass production in the fourth quarter of 2025, which means that it may be unveiled and displayed earlier; Some customers may get the chip earlier for evaluation.
According to Mich-Chi Kuo, the R100 is expected to be powered by TSMC's 3nm EUV FinFET process, specifically the TSMC-N3 node. In contrast, the new "Blackwell" B100 uses the TSMC-N4P. It will be a chip-scale GPU with a 4x reticle design (Blackwell has a 3.3x grille) and uses TSMC's CoWoS-L package like the B100.
The chip is expected to be one of the first customers of HBM4 stacked memory, with 8 stacks, the height of which is not yet known. The Grace Ruben GR200 CPU+GPU combo is likely to feature a new "Grace" CPU made on the 3nm node, most likely with optical shrinkage technology to reduce power consumption.
The Q4 2025 volume roadmap target means that customers will start receiving the chips in early 2026.