Benchmarking Arm in the Data Center

In our newsletter this week (drop us a line if you would like to subscribe for free) we linked to a blog post from software vendor Red Panda. They evaluated Amazon’s Graviton CPU in comparison to Intel’s latest Xeon Skylake. As long-time proponents of Arm in the data center, we found their result really interesting.

We were not familiar with Red Panda prior to this post. They provide tools for using ‘streaming’ databases like the open source Kafka. In their words:

Redpanda is a developer-centric modern streaming data platform that is API-compatible with Apache Kafka

These are people who care a lot about CPU performance and have the ability to measure such things, it can have a big impact on their business. This means they can provide a neutral analysis, without influence from vendors, unlike almost every other analysis out there.

When AWS first launched Graviton, we noted that the chip seemed to be built more as a proof of a concept than for its commercial prospects. CPUs are hard to design, and AWS would not be the first company to take this dipping-their-toe-in-the-water approach. That analysis seems to have been proven out, as we do not think Graviton version 1 ever went into production. But then AWS launched version 2, with a version 3 in the works. We believe Red Panda was evaluating instances running on version 2 silicon.

Their blog post is full of analysis, looking at all sorts of aspects of chip and system performance. The simple summary of all that is Graviton is roughly on par with Skylake, ahead on a couple of performance metrics, not too far behind on others, but largely in line. However, when Red Panda took the step of comparing performance relative to power consumption (and thus cost) they came to the conclusion that Graviton is about 20% better. That is to say comparable performance at 80% of the power cost. Looking back at other Arm server CPUs, we have seen very similar roadmaps – going from “this works” to “this is good enough and cheaper”. If that pattern holds up, we believe the next version of Graviton will likely surpass comparable Intel parts in terms of raw performance, with the power savings an added bonus.

Obviously, there are some caveats. First, Intel Skylake is not the best x86 CPU out there, that title rests with AMD. Second, if you dig into the details here, Skylake still has better performance per thread, but Graviton seems to have better performance per core. Set aside the complexity of these labels, this means that there will still be many applications where Skylake wins. Most software is not written for the widespread parallelism that multiple cores provides. So Intel is still going to perform better for most average customers (they hyperscalers are a different story). That being said, it is not unreasonable to think that Arm based CPUs could someday overtake x86 on these dimensions as well. These parts definitely seem to be advancing at a faster pace than x86. Predicting the future on this depends a lot on what will happen to Intel, so it is too hard to call at this stage.

All that being said, we see this analysis as a very strong endorsement that Arm is more than viable in the data center, and could very well overtake not only Intel but all of x86 in the not too distant future. If (when) that happens, and Arm CPU performance overtakes x86, it would provoke a major shift in the economics of data center silicon.

Leave a Reply