For the past 11 years, the phrase “ARM server” has been very nearly an oxymoron. There have been a handful of products and prior entries, but none of them have achieved the kind of performance or pricing that would position them to make a serious dent in the x86 market. Small start-ups like Calxeda and larger, established firms like AMD and Qualcomm all lined up to take a crack at ARM servers and either went bankrupt (Calxeda) or shelved their plans to lanuch parts (AMD, Qualcomm).
More recent start-ups like Ampere and Nuvia have declared their intent to compete in servers — Ampere initially led with a design it bought from AppliedMicro before beginning work on its own architecture, while Nuvia is too new for us to know much anything about its CPU core. Amazon also launched its own Graviton CPU back in 2018, but the CPU wasn’t all that great a deal. Graviton2 looks to be a different story.
The Graviton2 is based on ARM’s Neoverse N1, with a slightly lower clock than ARM specified (2.5GHz). Each Graviton2 core offers 128KB of L1 (64K i-cache, 64K d-cache), and a 1MB L2. The entire 64-core CPU is backed up with a 32MB L3. The Neoverse N1 is a derivative of the Cortex-A76 and Amazon is using ARM’s CMN-600 mesh interconnect. There are 8 memory channels supporting DDR4-3200.
Graviton2 is a completely different product than Graviton, according to Anandtech, which recently reviewed it against competing solutions from Intel and AMD. Graviton2 is being offered in cloud server instances at a somewhat lower price than competing servers from AMD or Intel, and it’s offering 64 full CPU cores, where both the AMD and Intel systems offer 32C/64T implementations. Anand compared the 64C/64T CPU against an AMD Epyc 7571 (first-generation Epyc, 32C/64T, 2.5GHz base, 2.9GHz boost) and an Intel Xeon Platinum 8259CL. This system was a dual-socket system with 16 CPU cores active in each socket and a 2.9GHz base / 3.2GHz boost clock. Prices were $2.464 per the Graviton2, $2.752 an hour for the first-gen Epyc, and $3.808 for a Cascade Lake Xeon.
Anandtech ran the server instances through a series of memory, latency, and SPEC 2006 and 2017 CPU tests. While this obviously doesn’t account for every workload that could be run on Graviton2 servers, the combined set of tests offers an approximation of overall workloads. The full results are well worth reading if you want to understand some of the subtleties of CPU design trade-offs — Graviton has some serious trouble with workloads that require lots of memory and cache bandwidth due to competition between the CPU threads, and its scaling falls off rapidly in these sorts of tests:
While the per-thread scaling is weak in some workloads compared with what you might expect to see from Intel or AMD, Amazon has two tricks up its sleeve. First, remember, Graviton2 has 64 full CPU cores, while AMD and Intel are both limited to 32C/64T. The reason Anandtech compared in this fashion is that these are the x86 systems priced most closely against the Graviton2. It may not be a fair comparison in terms of core count, but Amazon is factually offering 64-core Graviton2 instances for less money than 32-core / 64-thread Intel and AMD systems.
The end of result of this is that the Graviton2 beats the Epyc 7571’s performance in virtually every workload and mostly beats Intel’s Cascade Lake as well. But that’s not Amazon’s only advantage. The company is offering Graviton2 at an excellent price compared with its competition.
These are extremely strong results for Graviton2 — remember, lower is better in these graphs, as it reflects cost per SPEC run. Amazon has delivered a 40 percent improvement in performance per dollar. With that said, these results don’t include any data for AMD’s 7nm Epyc CPUs, which offer stronger overall performance than first-generation Epyc. That CPU is not currently available for comparison in an equivalent Amazon cloud instance and Anandtech wasn’t able to test it. Without knowing how Amazon will price cloud instances that use Rome, it’s impossible to know how that comparison will play out.
According to Anandtech, “In terms of value, the Graviton2 seemingly ends up with top grades and puts the competition to shame.” After more than a decade of promises and a handful of very modest product launches, Graviton2 could be the first in a number of ARM chips to challenge dominance in the overall server industry — at least in terms of what cloud hyperscalers are deploying.
- Ampere Altra ARM CPUs Launch with Up to 80 Cores to Challenge Xeon, Epyc
- Nuvia Will Challenge Intel and AMD For Hyperscaler CPU Sockets Using Custom ARM Design
- Nvidia Supercharges ARM HPC Deployments