Amazon Web Services (AWS) has announced the second generation of its own ARM processors and several EC2 instances: the Graviton2 These chips have 64 CPU cores and are based on ARMS current server design, with 7 nm technology for over 30 billion transistors. The C6g, M6g and R6g instances are expected to achieve up to 7x the speed of the previous A1 instances with the first Graviton processor.
The Graviton2 represent a variant of Neoverse N1 (Ares): this is modified Cortex-A76, with AWS on 64 cores of possible 128 sets. These use one MByte L2 cache per core, 32 MB L3 buffers and eight DDR4-3200 memory channels, as well as 64 PCIe Gen4 lanes for NVMe SSDs. The previous Graviton1 are 16 nm chips with only 16 cores based on Cortex-A72. Graviton2 is expected to double the floating-point performance per core.
By 2020 the C6g (compute with 128 GB DDR4), M6g (general with 256 GB DDR4) and R6g (memory with 512 GB DDR4) will be available, for the time being the M6g will start as a preview. There are also C6gd, M6gd and R6gd versions, where the d suffix identifies local NVMe SSDs. The main memory is encrypted with AES-256, the instances are connected with 25 Gbps Ethernet.
Graviton1 and Graviton2 in comparison (Image: AWS)
Compared to M5 instances with 24-core Xeon Platinum 8175: Here is a 64-core Graviton2 instance between 24 and 54 percent in front, whether integer or float performance, such as x264 video encoding. On average, the performance advantage should be 40 percent, with rental costs for users also being 20 percent lower, according to AWS.