Nvidia has unveiled its next-generation AI platform dubbed Rubin, a successor to Blackwell that will launch in 2026. The company will also launch a new CPU platform dubbed Vera the same year, Nvidia CEO Jensen Huang said in a keynote address at National Taiwan University. Huang didn’t share the specs of Rubin the chip will likely be made with TSMC’s 3-nanometer (nm) node and chip-on-wafer-on-substrate (CoWoS)-L process. Blackwell launching this year is being manufactured using TSMC’s 4nm node. CoWoS-L doesn’t use an interposer and instead connects the chips through a local silicon interconnect __ a technology similar to Intel’s embedded multi-die bridge and Samsung’s I-Cube-E. Rubin is likely to use 12-stack HBM4 as its memory. SK Hynix is planning to start mass production of the Gen 6 HBM in 2025. Meanwhile, CEO Huang also announced Blackwell Ultra and Rubin Ultra, which will launch in 2025 and 2027, respectively. The Nvidia chief said GPUs will have a one-year life cycle post-Rubin. This means HBM suppliers Samsung and SK Hynix will also accelerate the life cycle of their HBM chips. SK Hynix last month said HBMs in the past were developed with a 2-year life cycle but this has changed to a year recently.
|