Amazon’s Graviton4 CPU, the latest in its line of Arm-based server processors, promises the usual performance gains; the cloud-computing giant claims it provides up to 30 percent better compute performance, and 75 percent more memory bandwidth, than the chip’s predecessor.
Beyond expected performance, however, Graviton4 places the spotlight on security. It’s the first of the line to include features such as Branch Target Identification (BTI), a component of the Arm architecture designed to protect against sophisticated cyber threats while delivering significant performance gains.
“AWS is the only cloud provider that we know of to announce that all high-speed interfaces on a processor such as Graviton4 are encrypted, or to have made a CPU that supports Pointer Authentication and Branch Target Identification generally available,” says Rahul Kulkarni, director of compute and AI/ML at Amazon Web Services.
Graviton4’s Security Focus
Introduced as part of the Armv8.5-A architecture’s security extensions in 2018, BTI mitigates against branch prediction attacks, a type of side-channel attack that exploits the branch prediction mechanism found in nearly all modern processors. Branch prediction improves a CPU’s performance by guessing the outcome of conditional operations. This speeds computing, because it allows the processor to continue working while it awaits the answer to conditions such as if/then. Any work is “speculative”, and is undone if the guess was wrong. But attackers have learned to weaponize branch prediction and other speculative execution functions to read or manipulate data; this led to the Spectre and Meltdown attacks in 2018. Dozens of new branch prediction attacks have appeared since.
“[We allow] customers using Amazon Linux 2023 to get the defense in depth protection from these features by default,” says Kulkarni.
BTI combats these exploits by marking certain target memory addresses as “valid”. If the branch predictor attempts to address an unmarked memory address, a security exception occurs, preventing speculative execution of the potentially malicious code. BTI is backed up by Pointer Authentication. Pointers are variables that store the memory address of another variable. The authentication feature adds a cryptographic signature to authenticate memory pointers, thus helping thwart attacks that attempt to alter data in memory.
Amazon makes these features of the Arm architecture accessible by building them into Amazon Linux 2023, the customized version of Linux available through the company’s “elastic cloud computing” service, EC2. “This allows customers using Amazon Linux 2023 to get the defense in depth protection from these features by default,” says Kulkarni.
Graviton4 also hedges against attacks by encrypting data across high-speed hardware interfaces. This includes Graviton’s memory and AWS Nitro cards, which house Amazon’s proprietary input/output hardware and move data within AWS data centers. Encrypting data across these links should decrease the possibility of man-in-the-middle attacks designed to intercept data as it passes between different elements of Amazon’s server infrastructure.
Graviton4’s Computing Power
Graviton4 is the fourth generation in Amazon’s Graviton CPU architecture. Launched in November of 2018, Graviton is an Arm-based 64-bit CPU designed in-house specifically for Amazon’s Web Services (AWS). The original Graviton CPU, designed by Amazon subsidiary Annapurna Labs, was part of a wave of Arm-based server CPUs that began to arrive through the later end of the prior decade.
“Many people told me it was impossible to build a chip that could compete with the x86 CPUs and didn’t use the x86 architecture,” Ali Saidi, senior principal engineer at AWS, said in an interview published by Amazon’s A to Z blog. “But 25 years ago, x86 wasn’t the dominant architecture. The innovation and economies of scale of the PC drove success in other areas like servers. Since it happened before, I knew it could happen again.”
Amazon Web Services’ chief evangelist Jeff Barr poses with the Graviton4 processor.Amazon
Graviton4 has improved significantly since the first edition; the core count has increased from 16 to 96 cores, the architecture has moved from Arm’s general-purpose Cortex cores to the server-specific Neoverse N2, and the L2 cache size has expanded from 8MB to 192MB.
Kulkarni says ECS2 instances powered by Graviton4 should deliver “up to a 30 percent performance increase,” over Graviton 3, and that “third-parties, such as SmugMug [which owns Flickr] and SAP have corroborated 20 to 40 percent improvements for the same instance size.”
First announced in November of 2023, EC2 instances with Graviton4 are now generally available; Amazon says that besides SmugMug and SAP, Epic Games, Honeycomb.io, and ClickHouse are among the customers that have leapt from prior Graviton instances to the new hardware.
