RESEARCH NOTE: AWS Announces New Graviton4 Silicon for Both Scale-Out and Scale-Up Workloads

By Matt Kimball, Patrick Moorhead - November 28, 2023

When AWS first announced instances based on its own Graviton processors in 2018, it accelerated an industry movement toward Arm-based servers for scale-out, cloud-native workloads. This move further stimulated ISV ecosystem support, leading to other cloud providers establishing their own Arm-based instances.

At AWS re:Invent today, the company signaled an update and an expansion to the Graviton family, announcing the Graviton4 along with the R8g family of instances based on this new CPU.

How has AWS expanded the Graviton family, and what does this mean for customers? Further, what does this mean for the silicon industry? I’ll cover this and more in the following sections.

Graviton’s utility has grown over time

AWS, like Microsoft Azure, Google Cloud, and Oracle Cloud, initially used its Arm-based instances to deliver cost-efficient and power-efficient performance. If a customer wanted to run workloads that were not CPU performance-intensive, instances based on the Graviton were an ideal platform. Indeed, it’s a mark of the company’s success with this model that AWS claims more than 50,000 customers use 150 EC2 instances running on over 2 million Graviton CPUs.

As AWS delivered newer Graviton generations, the company increased performance and the type of workloads that could be supported. While Graviton1 was squarely focused on scale-out performance, Graviton2 expanded support to more general-purpose workloads such as MySQL database support. As Graviton3 hit the market, the R7g instances deployed on this chip were able to support some machine learning (ML) and high-performance computing (HPC) applications.

It’s important to note that all of this support came from a single-socket (i.e., one-CPU) server design built on Arm’s Neoverse N-Series platform. That platform is targeted at delivering the best performance for scale-out, cloud-native workloads, which translates into providing the best performance-per-watt capability for cloud service providers.

The expansion of Graviton, along with the introduction of instances driven by Arm-based processors from Ampere at the other CSPs, elicited responses from CPU giants AMD and Intel. In June 2023, AMD announced its 128-core EPYC CPU (codenamed “Bergamo”) to counter the growing Arm threat. In September, Intel announced that its Xeon CPU (codenamed “Sierra Forest”) will ship with up to 288 cores. These announcements are perhaps the strongest validation AWS and other CSPs could hope for with their Arm strategies.

Graviton4 flips the script

Before getting into the specifications of Graviton 4, it’s worth noting that the company made a design decision that is quite impactful. With this new CPU, AWS moved away from the Arm Neoverse N-Series (N2) in favor of the V-Series (V2). This is significant because the V2 architecture was designed as a pure performance play for scale-up, whereas N2 is explicitly designed for scale-out. This doesn’t mean that efficiency is fully surrendered in the name of performance, but rather that the performance-versus-power knob can be dialed further toward performance. Stated more directly, this puts Graviton4 in direct competition with AMD and Intel across all performance bands at AWS.

Graviton4 specs
Graviton4 delivers a richness of cores, memory, and I/O — Source: AWS

Looking at the top-level specifications of Graviton4 (as in the graphic above), it is easy to see how the chip competes with the performance CPUs from the mainstream x86 players. It features a lot of highly performant cores complemented by a big L2 cache, rich memory capacity (and bandwidth), and lots of fast I/O to deliver support for the most demanding workloads.

However, there’s one other capability that truly flips the script with regard to the x86 players: multi-socket support. Until now, mainstream Arm-based servers have steered clear of packing two CPUs in a server to deliver support across the workload performance stack. With Graviton4, R8g instances will support both single-socket and dual-socket configurations.

Graviton4 dual-socket support
Dual-socket support enables Graviton4 to meet the full range of customers’ needs — Source: AWS

What does a dual-socket configuration mean in terms of workload support? Well, we live in a data-driven world. Supporting the vast amounts of data being generated is critical to the success of a business. As a cloud provider, the ability to deliver a data factory (if you will) tailored from the ground up for optimal performance within the cloud environment can be a killer differentiator. And this is what Graviton4 is about. Workloads such as high-performance databases, in-memory caches, and big data analytics were once off-limits for Graviton-based instances. With R8g, these workloads are now right in the Graviton wheelhouse. Now, a business can run both its front-end applications and back-end data stores from Graviton-based AWS instances.

While performance is one vital element for enterprise-grade computing, so too is security. In addition to the inherent security built into V2, AWS has also included encryption in the system interfaces that move data. Performance without protection is useless today, and Graviton accounts for this with encryption built around DRAM, Nitro cards (AWS accelerators), and Graviton coherent links. Thanks to this, data at work, data at rest, and data in flight are all protected.

How well does Graviton4 perform?

Analysts were not given raw performance numbers at the time of the briefing, so it’s nearly impossible to do a direct comparison among Graviton, EPYC, and Xeon. Instead, AWS showed performance of R8g instances (in preview) relative to previous Graviton-based instances. These numbers were impressive when looking at database performance and cloud-based workloads. The HammerDB benchmark showed a 40% advantage for R8g versus R7g, and similar gains when testing a Groovy/Grails application on eight vCPUs.

Graviton4 versus Graviton3 performance — Source: AWS

While these numbers are strong, I’d like to see an instance-versus-instance comparison of Graviton4 instances (R8g) relative to AMD and Intel-based instances. These numbers matter most if a company is considering consuming a big data analytics service, for example.

What does this mean for the market?

While the release of Graviton4 is significant in many ways, the biggest questions I have are: What does this mean for the competitive cloud landscape? And what does Graviton4 mean for Intel and AMD?

On the cloud landscape front, the competition is fierce. The most recent market data from Synergy Research shows AWS with roughly 33% of the market for hosted cloud, infrastructure-as-a-service, and platform-as-a-service. Meanwhile, Azure and Google Cloud Platform are each gaining market share. Both competitors have gained roughly 5% or more over the past five years, with Azure now commanding about 23% of the market and Google Cloud 11%. Further, Azure recently introduced two new silicon innovations of its own. (You can read my coverage of that, coauthored with Moor Insights & Strategy CEO and Chief Analyst Patrick Moorhead, here.)

In this competitive landscape, innovation is key to growth. While I don’t believe Graviton4 will shift market share or cause customers to move workloads from Azure or Google Cloud, I do believe it will deliver value to existing AWS customers. That value can lead to AWS being the landing spot for existing on-premises workloads that have yet to be migrated to the cloud.

Looking at the impact on the merchant silicon providers is a little more interesting. It is easy to publicly say that nothing changes and relationships are stronger than ever. However, if Graviton4 performs as advertised, why wouldn’t AWS deploy these servers everywhere possible? And how does this not impact the volume (and resulting pricing) of x86 chips? Further, I would venture to say that Microsoft is looking at this and thinking about how its own homegrown CPU (Cobalt) will evolve over time.

I see the launch of Graviton4 as a potentially significant inflection point in both the cloud and silicon markets—a point where the dominance of vendors including Intel and AMD are challenged, and a point that requires a shift in how these vendors design and sell silicon to the cloud market.

Time will tell, and it won’t take long.

+ posts

Matt Kimball is a Moor Insights & Strategy senior datacenter analyst covering servers and storage. Matt’s 25 plus years of real-world experience in high tech spans from hardware to software as a product manager, product marketer, engineer and enterprise IT practitioner.  This experience has led to a firm conviction that the success of an offering lies, of course, in a profitable, unique and targeted offering, but most importantly in the ability to position and communicate it effectively to the target audience.

+ posts

Patrick founded the firm based on his real-world world technology experiences with the understanding of what he wasn’t getting from analysts and consultants. Ten years later, Patrick is ranked #1 among technology industry analysts in terms of “power” (ARInsights)  in “press citations” (Apollo Research). Moorhead is a contributor at Forbes and frequently appears on CNBC. He is a broad-based analyst covering a wide variety of topics including the cloud, enterprise SaaS, collaboration, client computing, and semiconductors. He has 30 years of experience including 15 years of executive experience at high tech companies (NCR, AT&T, Compaq, now HP, and AMD) leading strategy, product management, product marketing, and corporate marketing, including three industry board appointments.