In context: Amazon's AWS Graviton line of Arm-based server CPUs is designed by subsidiary Annapurna Labs. It introduced the processors in 2018 for the Elastic Compute Cloud. These custom silicon chips, featuring 64-bit Neoverse cores, power AWS's A1 instances tailored for Arm workloads like web services, caching, and microservices.
Amazon Web Services has landed a haymaker with its latest Graviton4 processor. They're exclusive to AWS's cloud servers, but the folks at Phoronix have somehow managed to get their hands on a unit to give us a peek at its performance potential.
Graviton4 packs 96 Arm Neoverse V2 cores, each with 2MB of L2 cache. The chip also rocks 12 channels of DDR5-5600 RAM, giving it stupid amounts of memory bandwidth to flex those cores. Positioning this offering for R8g instances, AWS promises up to triple the vCPUs and RAM compared to the previous R7g instances based on Graviton3. The company also claims 30 percent zippier web apps, 40 percent faster databases, and at least 40 percent better Java software performance.
However, the real story lies in those benchmarks, which the publication ran on Ubuntu 24.04. In heavily parallelized HPC workloads like miniFE (finite element modeling) and Xcompact3d (complex fluid dynamics), Graviton4 demolished not just its predecessors but even AMD's EPYC 'Genoa' chips.
One particularly impressive showing was in the ACES DGEMM HPC benchmark, where the 96-core Graviton4 metal instance scored a staggering 71,131 points, smoking the second-place 96-core AMD EPYC 9684X at 53,167 points.
In code compilation, the Graviton4 significantly outpaced the Ampere Altra Max 128-core flagship but lagged behind the varying core count Xeon and EPYC processors. However, it beat the EPYC 9754 in the Timed LLVM Compilation test.
The surprises kept coming with workloads not necessarily associated with Arm chips. Graviton4 demolished the competition in 7-Zip compression. Cryptography is another strong suit, with the Graviton4 nearly tripling its predecessor's performance in algorithms like ChaCha20.
After testing over 30 different workloads, Phoronix concluded that the Graviton4 is hands down the fastest Arm server processor to date. It's giving current Intel and AMD chips a considerable run for their money across various tasks.
Of course, this silicon arms race will only heat up further with new chips like Intel's Granite Rapids and AMD's Turin on the horizon. For now, AWS has a performance monster on its hands with Graviton4.