AWS Announces Graviton3

Amazon Web Services announced Graviton3 at re:Invent 2021 (December 1, 2021, Las Vegas) as its third-generation custom Arm server processor. The chip used 64 Arm Neoverse V1 cores (a high-performance variant of the Armv8.4 ISA, compared to the Neoverse N1 cores in Graviton2) manufactured on TSMC’s N5 (5nm) process versus Graviton2’s TSMC 7nm. AWS claimed 25% better compute performance, 2× better floating-point performance, 2× better cryptography performance, and 3× better machine learning performance compared to Graviton2 (the last comparison due to Graviton3’s support for BF16 and INT8 matrix multiply instructions via Arm’s Scalable Vector Extension). Graviton3 was also the first AWS-designed chip with DDR5-4800 memory support — a full year before DDR5 became widely available in consumer desktop platforms — yielding 50% higher memory bandwidth than Graviton2’s DDR4. AWS claimed 60% better energy efficiency versus comparable x86-based EC2 instances in the same workload category.

Graviton3 EC2 instances launched as the C7g family in May 2022, with instance sizes from c7g.medium (1 vCPU, 2 GB RAM) to c7g.16xlarge (64 vCPUs, 128 GB RAM). AWS’s custom chip strategy began in 2018 with the first Graviton (Annapurna Labs, 16 Arm Cortex-A72 cores, announced AWS re:Invent 2018), followed by Graviton2 (64 Neoverse N1, announced December 2019, generally available May 2020) which had already attracted major workloads: Amazon.com’s internal services, including the retail website itself, migrated significant compute to Graviton2, and AWS cited partners including Netflix, Honeycomb, and Formula 1 as early Graviton2 adopters. The Graviton2’s combination of lower instance pricing (~20% cheaper per vCPU than comparable x86 instance types) and competitive single-thread performance made it a strong option for horizontally scalable web workloads written in Go, Java, Python, and Node.js.

The Graviton3 announcement illustrated the broader cloud provider custom silicon trend. Google had deployed its first custom Arm-based server chip (Titanium) for networking and security functions since 2016, and announced its general compute Arm chip (Axion, based on Neoverse V2) in 2023. Microsoft Azure announced Cobalt (also Neoverse V2) in November 2023. The motivation was consistent across providers: server CPUs designed around actual cloud datacenter workloads could optimize the instruction mix, memory hierarchy, and power envelope differently than Intel’s Xeon and AMD’s EPYC lines, which served a broader range of deployment environments. For AWS, Graviton also reduced dependence on Intel and AMD supply chains and chip pricing — Intel’s Xeon SP dominated AWS EC2 in 2018, with AWS spending hundreds of millions per year on server CPUs. Graviton4 (Neoverse V2, TSMC N3E, 96 cores) announced at re:Invent 2023 continued the cadence.