Graviton2, Amazon's new CPU with 30 billion transistors

The division AWS (Amazon Web Services) of the ecommerce giant unveiled a new architecture-based processor during the re: Invent 2019 event ARM's Neoverse N1 (also known as Ares), the Graviton2.

Installed on the company's servers for the management of M6g instances (sixth generation of EC2 instances), it counts on four times the number of cores of the previous one Graviton1, for a total of 64 (connected by a 2 TB / s mesh architecture), and as many as 30 billion transistors.

In addition to the amount of cores, the chip is equipped with 1 MB L2 cache per core, with 8 DDR4-3200 channels with support up to 512 GB of RAM, is 64 PCIe 4.0 lines, in addition to support for encryption AES256.

Some benchmarks proposed by the company have shown how the new SoC manages to achieve performances up to 7 times higher compared to its predecessor in the A1 instances, and performance per core up to 2 times higher, combined with memory access 5 times faster compared to Graviton1.

What surely makes the project interesting, is the trust that Amazon has in the new ARM architecture. Neoverse N1 was in fact designed to compete directly with the largest supplier of current chip servers, Intel.

AWS has indeed provided some interesting numbers for comparison with Intel chip-based M5 instances. According to Amazon, the instances guarantee a 20% lower cost and up to 40% higher performance. More specifically, the company provided this numbers:

• SPECjvm 2008: + 43% (estimated)
• SPEC CPU 2017 integer: + 44% (estimated)
• SPEC CPU 2017 floating point: + 24% (estimated)
• Load balancing HTTPS with Nginx: + 24%
• Memcached: performance + 43% with less latency
• X.264 video encoding: + 26%
• EDA simulation with Cadence Xcellium: + 54%

At first glance, statements of some importance would seem: Graviton2 has more cores than higher performance per virtual core. However, it should be noted that Intel CPUs have Hyper-Threading, which leads to having two virtual CPUs per core. This means that the tests shown by AWS use only one thread of the two available per core on Intel CPUs.

Graviton2 will be the beating heart of the “general purpose” EC2 instance M6g, of the C6g optimized for calculation and the R6g dedicated to memory, with up to 512 GiB of memory and 25 Gbps of network bandwidth. Instances are available immediately for non-production loads, while general availability and other instances will arrive in 2020.