AWS Graviton is a family of 64-bit ARM-based CPUs designed by the Amazon Web Services (AWS) subsidiary Annapurna Labs. The processor family is distinguished by its lower energy use relative to x86-64, static clock rates, and omission of simultaneous multithreading. It was designed to be tightly integrated with AWS servers and datacenters, and is not sold outside Amazon.[1]
In 2018, AWS released the first version of Graviton suitable for open-source and non-performance-critical scripting workloads as part of its A1 instance family.[2] The second generation, AWS Graviton2, was announced in December 2019 as the first of its sixth generation instances, with AWS promising 40% improved price/performance over fifth generation Intel and AMD instances[3] and an average of 72% reduction in power consumption.[4] In May 2022, AWS made available Graviton3 processors as part of its seventh generation EC2 instances, offering a further 25% better compute performance over Graviton2.[5]
The first Annapurna Labs silicon product launched under the AWS umbrella was the AWS Nitro hardware and supporting hypervisor in November 2017.[6] Following on from Nitro, Annapurna began to develop general-purpose CPUs using its expertise.
The benefits AWS anticipated included:
The first Graviton processor reached these goals. Graviton2 now offers better performance compared to X86-64: 35% faster running Redis,[7] 30% faster running Apache Cassandra,[8] and up to 117% higher throughput for MongoDB.[9] In addition to higher performance, Graviton offers 70% lower power consumption [10] and 20% lower price.[11]
| Graviton | |
| Numcores: | 16x Cortex A72 |
| Size-From: | 16 nm |
| Clock: | 2.3 GHz |
| L1cache: | 80 KB per core (48 instructions + 32 data) |
| L2cache: | 8 MB total |
| Arch: | AArch64 |
| Instructions: | AArch64 |
| Extensions: | Neon, crc, crypto |
| Successor: | Graviton2 |
| Support Status: | Supported |
The first Graviton CPU has 16 Cortex A72 cores, with ARMv8-A ISA including Neon, crc, crypto. The vCPUs are physical cores in a single NUMA domain, running at 2.3 GHz. It also includes hardware acceleration for floating-point math, SIMD, plus AES, SHA-1, SHA-256, GCM, and CRC-32 algorithms.[12]
Only the A1 EC2 instance contains the first version of Graviton.[13]
The Graviton2 CPU has 64 Neoverse N1 cores, with ARMv8.2-A ISA including 2×128 bit Neon, LSE, fp16, rcpc, dotprod, crypto. The vCPUs are physical cores in a single NUMA domain, running at 2.5 GHz.[14]
EC2 instances with Graviton2 CPU: M6g, M6gd, C6g, C6gd, C6gn, R6g, R6gd, T4g, X2gd, G5g, Im4gn, Is4gen, I4g.[15] One or more of these instances are available in 28 AWS regions.
The Graviton3 CPU has 64 Neoverse V1 cores, with ARMv8.4-A ISA including 4×128 bit Neon, 2×256 bit SVE,, rng, bf16,, crypto. Organized in a single NUMA domain, all vCPUs are physical cores running at 2.6 GHz. Graviton3 has 8 DDR5-4800 memory channels.
Compared to Graviton2, Graviton3 provides up to 25% better compute performance, up to 2× higher floating-point performance, up to 2× faster cryptographic workload performance, up to 3× better performance for machine learning workloads including support for bfloat16, and 50% more memory bandwidth. Graviton3-based instances use up to 60% less energy for the same performance than comparable EC2 instances.[16]
Graviton3E is a higher power version of Graviton3.[17]
EC2 instances with Graviton3 CPU: C7g, M7g, R7g; with local disk: C7gd, M7gd, R7gd.
EC2 instances with Graviton3E CPU: C7gn, HPC7g.
The Graviton4 CPU has 96 Neoverse V2 cores, with ARMv9.0-A ISA.[18] It has 2 MB of L2 cache per core (192 MB total), and 12 DDR5-5600 memory channels. Graviton4 supports Arm's Branch Target Identification (BTI).
Amazon claims that Graviton4 is up to 40% faster for databases, 30% faster for web applications, and 45% faster for large Java applications than the Graviton3.
EC2 instances with Graviton4 CPU: R8g,[19] X8g,[20] C8g,[21] M8g.[22]