Arm: Credit where credit is due

We have been fairly downbeat about Arm lately. This was not entirely intentional, as we have been discussing a lot of things around Arm like RISC V, Amazon’s plans in the data center, and the Soap Opera that was Arm China. We are also a bit confused about what they are doing at the corporate level – their seeming fear of going public, and the latest lawsuit against Qualcomm. But we have not written much about their successes – especially in the data center where they now have incredible momentum that merits attention.

A few weeks ago, we were on Ben Thompson’s podcast where we lamented the fact that Arm needs to do more to develop its data center software ecosystem, and wondered why they were not hosting developer conferences in San Francisco. So now we feel obligated to say – Arm is hosting a developer conference in San Francisco October 26-27. If this space is of any interest to you, you should attend. This will be the first conference we have attended since December 2019, drop us a line if you can make it.

And now is a really good time for Arm to be talking about the data center. All seven of the so-called “Super 7” hyperscalers (Ali, Amazon, Baidu, Facebook, Google, Microsoft and Tencent) are now running Arm CPUs in their data center. This runs the gamut from internally designed chips like Ali’s YiTian and AWS’s Graviton, to privately-held Ampere’s CPUs at Google and Microsoft.

As a result, we are starting to see the very real performance gains that these CPUs offer. Some of these hyperscalers are talking about 30%, 40% and even 50% improvements over x86 – in terms of total cost of ownership. Others are talking about Arm CPUs allowing 25% increase in server density in their data centers. This lines up with our comments about AWS’s CPU ambitions – data centers are constrained by their electricity budgets and Arm servers require less power which means more CPUs can fit in the same footprint.

And these are not experimental deployments, these are production systems. AWS has said that 48 out of its top 50 customers run workloads on Graviton. Combine this with the all the real-world evaluations we have linked to in our recent newsletters (which you can sign up for by clicking this link or on the side of this site), and there is a clear pattern emerging. Arm CPUs have reached a critical turning point in adoption. From where we sit it looks very clear that Arm CPUs have all the performance they need to overtake x86 in the data center. This is a major turning point for the industry.

Of course, we have always cautioned that the key barrier to adoption for Arm in the data center is all the software that needs to be optimized to run on it. But even here, it looks increasingly like we have reached that tipping point. Most, if not all, of the hyperscalers have ported their software stacks over, and this includes the many open source projects they contribute to. Is everything optimized for Arm? No, there will always be a long tail of code subject to inertia, but more and more the major software companies are taking Arm optimization seriously as they are seeing the performance gains that it makes possible.

We can quibble how long this will take to spread, and what will be the ultimate split between x86 and Arm, but it is hard to argue that Arm has now, finally become a real force in the data center.

Photo by SpaceX on Unsplash

Leave a Reply