Nvidia seems intent on speeding up its roadmap. Until now, the company had announced an annual update of its roadmap, but rumors seem to contradict this highly ambitious tempo. Some information now seems to indicate that Rubin GPUs and Vera CPUs should be available for sampling as early as September, with production already underway. Why is this? Some (scurrilous) commentators suggest that Blackwell isn’t an unqualified success, and that their performance isn’t good enough. Others, more credible, point to Jensen Huang’s fear of losing his lead to the Chinese, and his consequent desire to accelerate to knock out the competition.
With its Rubin GPUs, Nvidia intends to make its mark with a major new leap in power for AI processing, while paying particular attention to efficiency. At the same time, Nvidia will also be launching the replacement for its Grace CPUs with the new Vera range. This new SoC will be based on next-generation ARM cores and should offer a significant improvement. Obviously, this Rubin/Vera duo will be at the heart of the new solutions that Nvidia will be pushing.
Nvidia wants to knock out the competition in AI
With Rubin, Nvidia will introduce several new processes and methods to its GPU design. Firstly, Rubin will be designed using the chiplet method, a first for the brand. Secondly, the use of next-generation HBM4 chips should bring a significant improvement over the current HBM3E standard. Finally, on the etching side, Nvidia will switch to TMSC’s 3nm (N3P) process, taking advantage of the Taiwanese company’s CoWoS-L packaging, which could make the chip denser while improving its energy efficiency. Over the past few weeks, Jensen Huang has been touting TSMC’s superiority in this field
However, the release of this duo does not mean that everything will go smoothly. First of all, TSMC is a bit of a bottleneck, and secondly, integrating the CPU/GPU pair into a computing system doesn’t happen with a few snaps of the fingers. The arrival of the GPU and CPU could therefore be confirmed in the form of samples at the end of the summer, but integration into complete, functional systems could still take many months.
