Amazon Tests Graviton4 Chip Performance for Cloud EDA

ARM recently conducted benchmark tests on Amazon's latest custom data center chip, the Graviton4, to evaluate its performance in running EDA tools in the cloud. The Graviton4 chip, featuring around 95 billion transistors for the ARM Neoverse V2 CPU core and CMN-700 Coherent Mesh Network, is now available as the EC2 R8g instance family. This chip is designed to cater to memory-intensive workloads, such as large in-memory databases and big data analytics.

ARM is currently assessing the R8g instance for its in-house EDA workloads and has discovered that the enhanced performance and memory capabilities further strengthen the advantages of migrating high-performance engineering workloads to the ARM architecture. Amazon has positioned the Graviton4 as a significant advancement over its predecessor, the Graviton3E, particularly targeting HPC applications. However, detailed information about the chip has been limited, prompting ARM to conduct practical performance tests.

The architecture of the Graviton4 boasts twice the level 2 cache and 50% more DDR5 memory channels at higher data rates compared to its predecessor. This results in 16.7% more main-memory bandwidth per core, allowing for 50% more virtual CPUs per socket utilizing the Neoverse V2 core. Additionally, the chip supports more chips per instance, enhancing scalability. It also incorporates encrypted high-speed hardware interfaces and Branch Target Identification for improved security.

David Lecomber, leading the high-performance computing tools group at ARM in Cambridge, commended AWS for systematically enhancing the capabilities of the Graviton CPUs over the years. With the general availability of AWS Graviton4 CPUs, ARM is enthusiastic about the progress made by Amazon and Annapurna Labs in this achievement.

ARM conducted performance comparisons between the AWS Graviton3E and Graviton4 single socket instances, noting the higher core count in the Graviton4 (96-core) compared to the Graviton3E (64-core). While higher core counts can introduce scaling inefficiencies, the benchmarks revealed a +24% per vCPU performance advantage for the Graviton4. The tests encompassed a range of workloads, showcasing the chip's improvements in both memory-bound and compute-bound scenarios.