It’s only been a few weeks since Amazon announced that it would offer AMD Epyc servers as an option for Amazon Web Services (AWS). Now the company is announcing a new type of hardware platform for its just-created A1 instances — the Amazon Graviton processor.
The company writes:
Today we are launching EC2 instances powered by Arm-based AWS Graviton Processors. Built around Arm cores and making extensive use of custom-built silicon, the A1 instances are optimized for performance and cost. They are a great fit for scale-out workloads where you can share the load across a group of smaller instances. This includes containerized microservices, web servers, development environments, and caching fleets.
Instances that use scripting code can move their applications to A1 without a rewrite, but if your code runs natively you’ll need to rebuild it for an A1 instance.
The Graviton CPU
Here’s what we know about the Graviton to date. It’s based on the Cortex-A72, with a maximum clock speed of 2.3GHz. AWS VP James Hamilton writes:
These new instances feature up to 45% lower costs and will join the 170 different instance types supported by AWS, ranging from the Intel-based z1d instances which deliver a sustained all core frequency of 4.0 GHz, a 12 TB memory instance, the F1 instance family with up to 8 Field Programmable Gate Arrays, P3 instances with NVIDIA Tesla V100 GPUs, and the new M5a and R5a instances with AMD EPYC Processors. No other cloud offering even comes close.
The new AWS-designed, Arm-based A1 instances are available in 5 different instances types from 1 core with 2 GiB of memory up to 16 cores with 32 GiB of memory.
It’s not clear exactly what kind of custom work Amazon did on the CPU. According to the Register, the 16 vCPUs that make up each SoC are arranged in clusters of four, with a 2MB L2 cache shared between each quad. Each individual core has a 32KB L1 data cache and 48KB L1 instruction cache, which is a standard Cortex-A72 configuration. One vCPU maps to one physical CPU core. The Register reports that overall performance is quite varied in benchmarks. In some cases, 16 Graviton cores were unable to match even 5 cores of a Xeon E5-2697v4, a Broadwell-class CPU.
Almost an AMD Win
It’s easy to forget, but once upon a time, AMD was making a huge bet on ARM CPUs, not x86. When AMD announced its K12 CPU back in 2015, it chose to hand the “K” moniker — a label previously reserved for x86 chips — to a future ARM core. K12, we were told, would share extensive resources and development strategy with Zen. There were plans for an ambitious joint x86-ARM platform, dubbed Project Skybridge. According to The Register, Amazon and AMD were working closely together in 2015, until “AMD failed at meeting all the performance milestones Amazon set out.”
It’s not clear what milestones, precisely, AMD failed to meet — but we could pick from several. The company’s Project Skybridge was abruptly canceled (we speculated at the time that GF’s manufacturing woes could have been part of the problem), and its Cortex-A57 CPU, the Opteron A1100, was announced in 2014 but didn’t actually ship until 2016. If we had to guess, we’d guess the problems were with the A1100. Back then, AMD explained that part of the reason the A1100 was held back was that the infrastructure for ARM server deployments wasn’t as solid as it needed to be and that more work had to be done bringing the software stack up to snuff. That may well have been entirely true — ARM servers have taken years longer to come to market than anyone originally expected — but Amazon apparently wasn’t willing to wait. The firm dropped AMD and bought Annapurna Labs to perform its design work.
This may also explain why AMD shelved K12. Lisa Su has focused on semi-custom parts, and serving as a dedicated provider for Amazon’s ARM business could definitely have fit that target. In this telling, either Skybridge, Seattle, or both were trial balloons to demonstrate that AMD could bring ARM IP to market before launching a new custom part based on some of the same architectural building blocks as its x86 CPU. When AMD lost the Amazon opportunity, it may have moved to shelve its ARM design — and that, in turn, may have been part of why Jim Keller left the company when he did. All of this is speculation, to be clear — it just fits the timeline reasonably well.
Regardless of how things turned out with AMD, this is the first time we’ve seen ARM servers deployed in cloud instances and a significant step forward for the ecosystem for that reason alone.
Now Read: Amazon Launches Alexa App in Windows Store, Amazon May Split HQ2: One in Virginia, One Maybe in New York, and Amazon Tried to Sell ICE Its Faulty Facial Recognition Tech