The 2.3 GHz 64-Bit Graviton Processor Developed by Amazon Was Nearly an AMD Arm Processor • The Register

re: Invent The custom Amazon Web Services Graviton processor, revealed this week, was almost an AMD chip based on Arm, The register has learned.

Until early 2015, Amazon and AMD were working together on a 64-bit server-level Arm processor to deploy in Titan's Internet data centers. However, according to a well-placed source, "the AMD project failed to meet all the performance milestones set by Amazon," ended the project.

For example, Amazon purchased Arm and System Builder Annapurna Labs as a licensee, forcing the acquired team to build Internet-of-Things and Nitro Chipset, which handles network and storage tasks for servers. Amazon hosting EC2 virtual machines. .

Then, as noted on Monday, Annapurna engineers looked into the design of the graviton, a multi-core Arm processor that powers AWS 'A1 EC2 instances. These virtual machines are now available in the United States and Europe.

As for AMD, in 2016, it launched what was left of the Arm chip it was working on with Amazon, the Opteron A1100 code, Seattle. The clue was in the name, let's note. Today, AMD uses all of its Zen based x86 processors, Ryzen and Epyc, which are much more successful, and no one talks about the A1100.

At a time when the partnership between AMD and Amazon was crumbling, and just before the Web giant bought Annapurna, AWS James Hamilton complained that the hearts of the Arm processors could not compete with the Intel components in terms of performance. At the time, it was not publicly known that AWS was using AMD as a supplier of Arm processors.

When I joined AWS in 2009, I would not have predicted that we would design server processors less than ten years later.

Today, Hamilton said, "I see the potential of Arm based server processors for over a decade, but it takes time to gather all the necessary ingredients."

He also explained why Amazon decided to go it alone: ​​the license capacity of Arm's plans, via Annapurna, the ability to customize and adjust these designs, as well as the ability to contact contract manufacturers such as TSMC and Global Foundries and create competitive chips. .

While Intel lost its advantage, rival factories were able to catch up and build sufficiently powerful processors. In addition, today's top-of-the-line Arm processor models are much more than just smartphones and are capable of running desktop and lightweight server applications.

"Arm designs the processors, but these are licensed to companies that incorporate design into their silicon rather than producing them themselves," Hamilton said.

"This allows a diverse set of silicon producers, including Amazon, to innovate and specialize chips for different purposes, while leveraging the extensive software and tooling ecosystem Arm. "

"Most silicon-producing companies using this licensed technology are semiconductor companies without a base, that is, they operate in the semiconductor industry but outsource manufacturing. silicon chips in extremely expensive facilities to specialized companies such as Taiwan Semiconductor Manufacturing Company (TSMC) and Global Foundries. "

"When I joined AWS in 2009, I would not have predicted that we would design server processors less than ten years later."

Speaking of Intel, all of this comes at a time when cloud giants (Microsoft, Google, Facebook, Amazon, Baidu, etc.) are looking for alternative chip vendors to avoid Intel's high prices and shortage of components . Chipzilla holds just under 100% of the global computing computer market share. To free themselves from this influence and to be able to personalize their own silicon, the goliaths on the Internet plan to propose the offers Arm, OpenPower, RISC-V and AMD Epyc.

Specifications

Here's what we know about the Graviton right now. The cores of its processors are based on the Cortex-A72 models of the Army 2015 and are at 2.3 GHz. They are 64-bit, Armv8-A, little endian, non-NUMA, and offer hardware acceleration for floating-point mathematical algorithms, SIMD, as well as AES, SHA-1, SHA-256, GCM and CRC-32 algorithms .

The chip-based system uses a blend of friendly Neoverse technology for Arm data centers and Annapurna's internal designs. The 16 instances of vCPU are organized into four quad-core clusters with 2 MB of shared L2 cache per cluster, and 32 KB of L1 data cache and 48 KB of L1 instruction cache, per-core. A vCPU is mapped to a physical kernel.

"AWS Graviton processors are a new line of custom-designed processors from AWS that leverage Amazon's extensive expertise in creating platform solutions for large-scale cloud applications" said a cloud cloud spokesperson. The register aujourd & # 39; hui.

"These processors are based on the 64-bit Arm instruction set and feature Arm Neoverse cores as well as custom silicon designed by AWS, and the cores operate at a frequency of 2.3 GHz."

Semiconductor industry observer David Schor shared the SciMark and C-Ray references for the 16-core Graviton. During SciMark testing, the on-chip AWS system was twice as fast as a Raspberry Pi 3 Model B + under Linux 4.14.

Schor noted that although the Graviton works well in performance testing, when it comes to actual workloads, such as managing its WikiChip.org website, it may not always follow the powered kit by Intel.

"It's going well in the Phoronix test suite," he said. m said. "We do not compare well to our fully deployed website: Nginx + PHP + MediaWiki, and everything in. This is your test of the" real world. "The 16 cores can not match the 5 cores of our Xeon E5-2697 v4."

Other landmarks put the Graviton on the same footing as a Qualcomm Snapdragon 835, in terms of single-core performance. CPU benchmarks do not tell the whole story: there are always problems with network, latency, access to storage, and so on. in the cloud.

By reading between lines, Graviton A1 instances target scalable workloads, lightweight Web servers, experiments with Arm software stacks, and more. These are not versatile performance offerings: this is clear from the use of the A72, which at launch was intended for high-end smartphones and tablets.

In fact, the cynic among us believes that the A1 family exists in part to encourage AWS customers to strengthen the ecosystem. Arm Linux has existed for years and years, mainly in the fields of embedded electronics and the Internet of Things. In the area of ​​servers and businesses, however, it is just beginning.

Amazon has bigger plans for Annapurna and its Arm chips, and one way to find out what needs to be optimized, optimized, modified, and improved is to run client applications on the kit and see what's left over.

Meanwhile, AMD's Zen-based Epyc x86 processors can be rented from AWS, and their cost is lower than that of Intel-based instances. "AMD spokesperson Gary Silcott, who declined to comment on previous work, said," We are pleased to support AWS Elastic Compute Cloud with AMD EPYC processors in their new M5, T3 and R5 instances. " conducted with Amazon. ®