Exclusive Amazon Web Services' customized Graviton processor, revealed this week, was very nearly an Arm-based chip from AMD, The Register has learned.
Up until 2015, Amazon and AMD were working together on a 64-bit Arm server-grade processor to deploy in the internet titan's data centers. However, the project fell apart when, according to one well-placed source today, "AMD failed at meeting all the performance milestones Amazon set out."
In the end, Amazon went out and bought Arm licensee and system-on-chip designer Annapurna Labs, putting the acquired team to work designing Internet-of-Things gateways and its Nitro chipset, which handles networking and storage tasks for Amazon servers hosting EC2 virtual machines.
Next, as reported on Monday, the Annapurna engineers turned their hands to designing the Graviton, a multi-core Arm processor that powers AWS's A1 EC2 instances. These virtual machines are available now in the US and Europe.
As for AMD, in 2016 it launched what remained of the Arm chip it was working on with Amazon, the Opteron A1100 codenamed Seattle. The clue was in the name, we note. Today, AMD is all in with its much more successful Zen-based x86 processors, Ryzen and Epyc, and no one talks about the A1100. Well, except maybe Softiron.
Around the time the AMD-Amazon Arm partnership was collapsing, and just before the web giant bought Annapurna, AWS veep James Hamilton complained that Arm CPU cores couldn't match rival Intel parts in terms of performance. It wasn't known publicly at the time that AWS was tapping AMD as an Arm processor supplier.
When I joined AWS in 2009, I wouldn’t have predicted we would be designing server processors less than a decade later.
Today, Hamilton said, "I’ve seen the potential for Arm-based server processors for more than a decade, but it takes time for all the right ingredients to come together."
He also spelled out why Amazon decided to go it alone: the ability to license Arm blueprints, via Annapurna, the ability to customize and tweak those designs, and the ability to go to contract manufacturers like TSMC and Global Foundries, and get competitive chips made.
As Intel has lost its edge, rival factories have been able to catch up and fabricate good-enough processors. Also, today's high-end Arm CPU blueprints are much more than smartphone brains, and are capable of running desktop and light server applications.
"Arm does the processor design, but they license the processor to companies that integrate the design in their silicon rather than actually producing the processor themselves," said Hamilton.
"This enables a diverse set of silicon producers, including Amazon, to innovate and specialize chips for different purposes, while taking advantage of the extensive Arm software and tooling ecosystem.
"Most of companies that are producing silicon that license Arm technology are fabless semiconductor companies, which is to say they are in the semiconductor business but outsource the manufacturing of silicon chips in massively expensive facilities to specialized companies like Taiwan Semiconductor Manufacturing Company (TSMC) and Global Foundries."
He added: "When I joined AWS in 2009, I wouldn’t have predicted we would be designing server processors less than a decade later."
Speaking of Intel, this all comes at a time when cloud giants – Microsoft, Google, Facebook, Amazon, Baidu, and so on – are looking at alternative chip suppliers to avoid Intel's high prices and component shortages. Chipzilla owns just under 100 per cent of the world's data center compute market share. To break free from this grasp, and be able to customize their own silicon, internet goliaths are considering Arm, OpenPower, RISC-V and AMD Epyc offerings.
Here's what we know about the Graviton right now. Its CPU cores are based on Arm's 2015-era Cortex-A72 designs, and are clocked at 2.3GHz. They are 64-bit, Armv8-A, little endian, non-NUMA, and feature hardware acceleration for floating-point math, SIMD, plus AES, SHA-1, SHA-256, GCM, and CRC-32 algorithms.
The system-on-chips use a mix of Arm's data-center-friendly Neoverse technology, and Annapurna's in-house designs. 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. One vCPU maps to one physical core.
"AWS Graviton processors are a new line of processors that are custom designed by AWS utilizing Amazon’s extensive expertise in building platform solutions for cloud applications running at scale," a spokesperson for the cloud titan told The Register today.
"These processors are based on the 64-bit Arm instruction set and feature Arm Neoverse cores as well as custom silicon designed by AWS. The cores operate at a frequency of 2.3 GHz."
Semiconductor industry watcher David Schor shared SciMark and C-Ray benchmarks for the 16-core Graviton. In the SciMark testing, the AWS system-on-chip was twice as fast as a Raspberry Pi 3 Model B+ on Linux 4.14.
Schor noted that while the Graviton performs well in benchmarks, when it comes to real-world workloads, such as running his WikiChip.org website, it can't always keep up with Intel-powered kit.
"It does well on the Phoronix Test Suite," he said. "It does poorly benchmarking our website fully deployed on it: Nginx + PHP + MediaWiki, and everything else involved. This is your 'real world' test. All 16 cores can't match even 5 cores of our Xeon E5-2697 v4."
A rumble in Amazon's jungle: AWS now rents out homegrown 64-bit Arm server processorsREAD MORE
Other benchmarks put the Graviton on the same footing as a Qualcomm Snapdragon 835, in terms of single-core performance. CPU benchmarks don't tell the whole story: there's always networking, latency, storage access, and so on, to worry about in the cloud.
Reading between the lines, the Graviton A1 instances are aimed at scale-out workloads, light web serving, experimentation with Arm software stacks, and similar. They are not general-purpose performance offerings: that much is obvious from the use of the A72, which at launch was aimed at high-end smartphones and tablets.
In fact, the cynical among us reckon the A1 family exists partly to encourage AWS customers to firm up the ecosystem. Arm Linux has existed for years and years, mainly in the worlds of embedded electronics and the Internet-of-Things. In the server and enterprise space, though, it's only now emerging from infancy.
Amazon has greater plans for Annapurna and its Arm chips, and one way to find out what needs to be optimized, tuned, tweaked, and improved is to run customer applications on the kit and see what sticks.
Meanwhile, AMD's Zen-based x86 Epyc processors can be rented out from AWS, and they work out cheaper than Intel-based instances. "We’re thrilled to be supporting AWS Elastic Compute Cloud with AMD EPYC processors in their new M5, T3 and R5 instances," AMD spokesperson Gary Silcott told us, declining to comment on any previous work with Amazon. ®