December 30, 2020
Earlier, software developers generally had to choose between Intel or Arm processors, which led to market dominance by these two products. In 2018, Amazon started offering its own chip: Graviton. Their key objective was to provide chips that cater to general needs, as well as analytical and AI workloads, at a fraction of Intel or NVIDIA chips. In other words, Amazon aims to shift the data-center market from one based on standard Intel and NVIDIA chips to one based on Artificial Intelligence and cloud-computing technologies.
The Graviton processors are custom built by AWS (Amazon’s cloud wing). The chips were co-created by created by Annapurna Labs, a chip company acquired by Amazon in 2015 for US$350 million. This Israel-based company has also built two generations of ‘Nitro’ ASICs that run networking and storage tasks in Amazon’s data centers. Nitro will also help in creation of AWS’s next custom chip, called AWS Trainium.
While Graviton’s first iteration was launched in 2018, its latest Graviton2, unveiled at AWS’ re:Invent 2019 is still making headlines. Graviton was based on 64bit Armv8 Cortex-72 microarchitecture. It contained 2MB of L2 cache for the four quad-core clusters. This chip is supported by operating systems like Amazon Linux 2, Red Hat Enterprise Linux, and Ubuntu.
Graviton2 is an Arm chip designed by AWS and built on 64-bit Arm Neoverse N1 cores. This 7nm chip has 30 million transistors. It provides a number of improvements over its predecessor. These features include two times faster floating point performance per core, optimized instructions for faster machine learning inference, customer hardware acceleration, always-on fully encrypted DDR4 memory and 50% faster per core encryption performance for enhanced security. It powers general purpose (M6g), compute-optimized (C6g) and memory-optimized (R6g) instances and their variants with local NVMe-based SSD storage (M6gd, C6gd and R6gd). These instances have up to 25 Gbps of network bandwidth, 18 Gbps of EBS-Optimized bandwidth and are also available in bare metal form. The M6g instances has the capacity to handle twice the load of a Graviton A1.
Graviton2 gives 40% higher throughput as compared to x86-based instances at 20% lower cost. Additionally, it can deliver up to 7x the performance of Graviton A1 (first-gen) instances. Also, it can support up to 64 virtual CPUs, 25 Gbps of networking and 18 Gbps of EBS bandwidth.
Moreover, Graviton2 supports the Large System Extensions (LSE) that enhance locking and synchronization performance across large systems. It also has support for fp16 and 8-bit dot productions for machine learning, and relaxed consistency-processor consistent (RCpc) memory ordering. Its additional memory channels and double-sized per-core caches help in expediting the memory access by up to 5x.
During the first week of re:invent, AWS introduced a new Graviton2-powered C6gn instances. This new instances can add 100 Gbps networking and Elastic Fabric Adapter (EFA) support to the C6g family. It can be used for workloads requiring high networking bandwidth such as high-performance computing (HPC), real-time video communications and data analytics, since they deliver up to 40% better price-performance over C5n instances for similar workloads. It also offers 4x higher network bandwidth and up to 2x higher EBS bandwidth, than C6g for improved networking performance.
Earlier this year, AWS announced that its fully managed Kubernetes service Amazon EKS is now generally available on its Graviton2 processor. The managed Kubernetes service now supports the 64 bit ARMv8.2 architecture among others, and also offers end-to-end multi-architecture support along with support for mixed managed node groups. Further, the EKS API and tooling such as eksctl take care of the architecture-specific configurations, for example, launching Arm-based control plane components such as CoreDNS or kube-proxy pods.
Today the chip manufacturers are caught up under tremendous pressure due race for dominant position in market. While it is leading to numerous advancements and innovations in processors, with recent demand in artificial intelligence (AI) capabilities during the pandemic, chips also need to address this demand gap. Graviton2 has an edge due to its relatively cheaper price and optimization for cloud-native applications. However, it is yet to dethrone chips maker giants Intel and NVIDIA.
The current market competitors of Graviton2 are AMD EPYC 7571 (Zen1) powered m5a instances, and Intel Xeon Platinum 8259CL (Cascade Lake) powered m5n instances.
Share This Article
Do the sharing thingy