re:Invent Amazon has designed its own 64-bit Arm server processors, dubbed Graviton, and is right now renting them out on AWS.
Just in time for its annual re:Invent conference in Las Vegas, the internet titan today revealed its A1 family of EC2 instances, ranging from the a1.medium with 1 vCPU, 2GB of RAM, and up to 3.5Gbps EBS and 10Gbps network bandwidth, which costs $0.0255 per hour on-demand, all the way up to a1.4xlarge with 16 vCPUs and 32GB of RAM, costing $0.408 per hour.
This can work out cheaper than, say, comparable T instances in terms of CPU; specifications differ between instance types, though. You can compare and contrast for yourself here. Amazon claims A1 instances are up to 45 per cent cheaper than their x86 virtual machines, depending on the configuration.
In any case, you shouldn't just look at the raw price: you ought to test your software stack on the A1 VMs for performance and compatibility before committing to the architecture, due to the differences between Arm and x86.
Did some basic benchmarking of a1.4xlarge vs c5.4xlarge using Phoronix C-Ray benchmark. Render took 108s on c5, 112s on a1. c5 is $0.68/hr, a1 is $0.408/hr. Looks like it really is quite the bargain. #reinvent— rbranson: EVP, Thoughts (@rbranson) November 27, 2018
The Graviton CPUs are AArch64, run Amazon Linux 2, Red Hat Enterprise Linux 7.6, or Ubuntu, and are available in Amazon's US East (North Virginia and Ohio), US West (Oregon), and Europe (Ireland) regions in various forms from on-demand to dedicated hosting. In terms of applications, think scale-out services, web hosting, and so on.
If you're running stuff written in portable scripting or interpreted languages, you can probably just transfer this code to A1 virtual machines and give it a whirl. If not, then crack out that Armv8-A toolchain, grab pre-built packages from your preferred Linux distro, or try pre-built binaries, depending on availability.
"If you didn’t look very closely, you might not at first even notice that you were using an Arm server. That was by design," gushed Red Hat's Jon Masters who worked on getting RHEL supported by A1 instances. "I’ve been playing with Amazon’s A1 instances for a while and I’m blown away by the quality of the engineering and the capabilities of the team."
If you're wondering where this all came from, well, back in 2014, we did warn you Amazon was mulling using Arm chips for cloud compute resources. A major step in that quest was gobbling up Annapurna Labs for a few hundred million bucks in 2015. That biz is an Arm system-on-chip designer and licensee, and initially concentrated on Internet of Things projects. Then AWS revealed it was using Annapurna-designed custom chips dubbed Nitro in its cloud. These run networking and storage tasks offloaded by EC2 hypervisors, allowing x86-based instance-hosting machines to focus on running customer workloads while the Nitro ASICs shifted packets and data.
Now the chip design team has used that experience to craft and deploy Graviton Arm processors in the AWS cloud so that they run customer code on general-purpose virtual machines. As you'd expect from the highly secretive Amazon, there ain't much more in the way of technical details available right now. We've poked AWS's PR people for more info.
In the meantime, looking at
/proc/cpuinfo on an A1 instance, the CPU ID number is 0xD08, which suggests the Graviton is based on Arm's 2015-era Cortex-A72 blueprints. Also we note that one vCPU maps to one physical core. The 16 vCPU instances are arranged in four quad-core clusters with 2MB of shared L2 cache per cluster, and 32KB of L1 data cache, and 48KB of L1 instruction cache, per core.
Last month Arm claimed a million Arm-powered data center servers would ship in 2018, and most of them converged systems, with us predicting Amazon would make up a big chunk of that number. And Amazon is also offering AMD Epyc processors in its EC2 cloud. Meanwhile, Microsoft Azure wants to be at least 50-per-cent Arm powered.
All of which puts further pressure on Intel. Chipzilla's x86 CPUs make up more than 90 per cent of the world's data center compute market, according to analysts. It appears Epyc, Graviton, Cavium's ThunderX2, and other competitors, are, here and there, chipping away at that Intel market share. Pun intended. ®