The future will be built on ARM

ARM CEO Rene Haas outlined why he sees his company in a strong position during his Computex keynote.

The development of new processors is becoming increasingly expensive and complex, while resources such as developer capacity and energy are becoming scarce. This tension was the central theme of the keynote delivered by ARM CEO Rene Haas at the Computex computer fair in Taipei. However, he did not see himself being hindered by these opposing trends; instead, he saw a tremendous potential. After all, the demand for computing power is growing inexorably, not only in data centers and client devices but also in the automotive sector. He provided specific figures for the latter: The envisioned but yet-to-be-realized level-5 autonomous driving, for example, would require 4000 trillion operations per second (TOPS) – per vehicle, of course.

He also presented solutions for the obstacles in the way. ARM has been focused on developing energy-efficient computing cores since its inception. This has been true for the first Apple Newton as well as in recent times, during which the sales of ARM processors have multiplied. According to Haas’ statements, 50 billion ARM chips were installed by 2013, and by 2021, this number had quadrupled to 200 billion due to the boom in smart devices (ranging from smartphones and tablets to watches and speakers). Furthermore, it is expected to surpass the 250 billion mark this year.

Significant Volume Growth

The rapidly growing numbers across all device classes likely explain why ARM is currently engaged in a legal dispute with one of its largest customers, Qualcomm. It is rumored that ARM intends to change its business model and no longer charge fees per chip, but rather base them on the total price of the device equipped with the chip. However, Haas did not mention a word about this legal dispute on stage.

In terms of development capacity, ARM aims to provide stronger support to its customers. In previous systems-on-a-chip (SoCs), the respective licensees had to incorporate the cores into their designs and handle all integration and tape-out obstacles themselves. However, SoCs are becoming obsolete, and the transition from monolithic chips to clusters composed of multiple chiplets presents new opportunities. ARM plans to deliver not only the pure intellectual property (IP) but also complete subsystems that are tailored to the latest manufacturing processes and only need to be integrated into a chiplet design. SoC developers can then concentrate on the other components of the cluster.

As an example, he mentioned the recently launched chip by Nvidia, referred to as the “Superchip,” named Grace Hopper. The CPU component consists of two Neoverse-V2 blocks, each with 72 cores, supplied by ARM. This allowed Nvidia to focus on the remaining system design. The delivery of ready-made subsystems did not cost any jobs but rather ensured that Nvidia developers could concentrate on their own tasks, resulting in faster completion of the overall package. It is worth noting that this close collaboration may be partly due to Nvidia’s previous attempt to acquire ARM, which likely coincided with the development timeline of Grace Hopper.

Speaking of development time, it can be significantly reduced by leveraging ARM’s services. Haas mentioned a project that he couldn’t discuss publicly yet, which took only thirteen months from kick-off to tape-out — an impressive timeframe regardless of the chip’s unknown complexity.

And lastly, Haas does not believe that the demand for increased performance will decrease in the near future; on the contrary, it persists despite all challenges. Artificial intelligence simply craves it, and he sees a self-fueling interplay between training and application of models. Larger models require more client computing power to be able to use them locally. Once certain tasks can run on relatively underpowered systems like smartphones, there is a rapid demand for even better or better-trained models, which in turn require more performance in data centers.