NVIDIA's Next-Gen Rubin GPU Is Reportedly 6 Months Ahead Of Schedule
The site doesn't cite any specific sources, as usual, but it claims that the Rubin platform is some half a year ahead of schedule, and that NVIDIA may launch it early in cooperation with its partners, including TSMC and Taiwan's King Yuan Electronics Company (KYEC), which offers wafer testing and packaging services.
Even if Rubin really is ahead of schedule, NVIDIA may not elect to launch it as soon as possible if the companion "Vera" CPUs are not available. Alternatively, it may behoove NVIDIA to boot Rubin out the door ASAP, as the company is facing fierce competition from longtime rival AMD, whose MI300 accelerators actually offer superior generalized compute performance to NVIDIA's current-generation Blackwell GPUs.
Of course, NVIDIA's focus is on AI, and even if another company were to outpace NVIDIA on AI processing hardware, it’d still have to overcome the titanic challenge of producing a software stack as well-supported as NVIDIA's CUDA. In that context, the company's decision to get out ahead of GPU compute as a trend with CUDA way back in the GeForce GTX 200 days feels incredibly prescient.
United Daily News quotes Morgan Stanley analyst Jia-hong Zhan when saying that Rubin will be built on TSMC's 3nm process, use optical packaging, and that the Rubin Ultra product will be twice the size of Blackwell GB200, making use of four separate chips (where Blackwell bonds two.) We already know that it's going to use no less than twelve stacks of HBM4 memory. Truly a monstrous processor.
