The biggest and perhaps unexpected announcement – but thinking about it not too carefully – of Google Cloud Next 2024 is called Google Axion. The house of Mountain View, once synonymous with web services, is now much more and from today also CPU designer for their own datacenter.
Google Axion: what do we know?
The answer to the question above is “not everything”. This is an architecture-based processor ARMprecisely i Core Neoverse V2. The Neoverse V3 were recently announced, but development times are long and for this reason Google based its project on cores from an older generation.
With this commitment to design, which follows that on TPUs or other accelerators for YouTube’s infrastructure such as the Video Processing Unit, Google puts itself on the same level as rival companies in the cloud world and beyond. Rumors about its plans had circulated in February 2023 and it should not be forgotten that in 2021, three years ago, Google hired a certain Uri Frank, an Intel veteran with proven ability and genius in designing CPUs for servers.
The first to embark on the venture was Amazon Web Services (AWS), with the first Graviton processor over five years ago: at the end of last year we talked about the Graviton4, a server processor with 96 Neoverse V2 cores. Microsoft, in the same period, announced its first server CPU, called Cobalt 100, a 128-core solution based on Neoverse N2 cores.
It is unclear how many cores Google Axion integrates, but it is rumored to be a monolithic chip, at least that’s what it seems from the few shots. Thomas Kurian, CEO of Google Cloud, limited himself to declaring that Axion is already in production, which does not mean that TSMC or whoever is making it for it, but that already in use in internal data centers, although probably in limited numbers: The processor will then be accessible to external customers throughout the year via Google Cloud instances.
As regards the performance, Google said that the excellent ARM core base combined with its optimizations allowed it to achieve up to 30% better performance than the fastest “general purpose” ARM CPU-based instances available. Not only that, Google is sure it can tear up x86 CPUs offering up to 50% higher performance and up to 60% better energy efficiency.
Google’s CPU: a problem and an opportunity for Intel
The arrival of Axion, if we add it to that of Graviton and Cobalt, This doesn’t seem to be a good sign for incumbent CPU designers like Intel and AMD. Until now, Google had to necessarily turn to them to provide its instances and equip its data centers, but in the future that need will be increasingly less pressing. Mind you, we are not saying it will happen to us for sure, or tomorrow morning, but it could happen within a few years.
In a similar scenario Intel’s decision to open up to the production of chips for other companies takes on a certain importance, in direct competition with TSMC and other companies in the sector. Intel Foundry is bidding to produce processors just like Google’s Axion with the motto “If you can’t sell ’em, fab ’em“, i.e. if “you can’t sell them, produce them”.
Intel will produce a 64-core ARM processor
Intel’s business is in its infancy, between investments in new Fabs in Europe and the United States and the development of new production processes, but the plan launched by CEO Pat Gelsinger, not entirely clear in the past, is taking on ever greater consistency observing the movements of the great giants of the web and artificial intelligence. From here to its success, obviously, there is a long way to go, but if there is anyone who can do it, it is Intel. One wonders what AMD thinks and, in the worst case scenario, whether there is already a possible “plan B”…
An expected novelty? The direction of the market seems clear
We’ve been seeing ARM CPUs trying to conquer the server market for a while now. Marvell, Qualcomm and other names have tried several times to create a market with CPUs designed for the server world, failing. The problems they faced were mainly related to performance, significantly lower than their x86 counterparts, and the absence of a robust software ecosystem. Both of these elements no longer exist, given the developments of recent years.
What is really paving the way for ARM processors in the server world, however, is the fact that they are mostly designed and used by the companies themselves: hyperscalers. These are CPUs created specifically to meet the specific needs of large cloud operators, who have the resources and the will to “make everything they need at home”. In fact, if we exclude Ampere and its Altra processors, there is no real market for ARM CPUs in the server world: AWS has its Graviton, Microsoft has Cobalt, and now Google has Axion. All processors not available on the market and limited to the data centers of the three hyperscalers.
The only exception, which somewhat proves the rule, is NVIDIA’s Grace, which is used by all major cloud providers. However, it is a unique product, as it aims at the HPC world and, more specifically, at the AI field, thanks to the coupling between CPU and GPU. Therefore not a competitor of hyperscaler CPUs, but a processor Sui generis used in very specific areas.
Google’s move is not surprising, therefore, also because the company has been producing CPUs for several years: these are the Tensors that it uses on its smartphones, but the skills necessary for their design are very similar to those needed to design Axion. The market trend is to produce chips with many cores and high power consumption: in the cloud world there is often a need to run applications that do not require enormous computing power, and therefore it is more convenient for the service provider to offer a low-power CPU and with many cores so as to be able to host many clients on a single server.
Google Axion therefore does not represent a particular innovation, but confirms a trend that has already become evident over the last few years. It will be interesting to see whether Ampere can survive the fact that its major customers are, effectively, leaving it behind.
