AWS Graviton4 Benchmarks Prove To Deliver The Best ARM Cloud Server Performance
This week AWS announced that Graviton4 went into GA with the new R8G instances after Amazon originally announced their Graviton4 ARM64 server processors last year as built atop Arm Neoverse-V2 cores. I eagerly fired up some benchmarks myself and I was surprised by the generational uplift compared to Graviton3. At the same vCPU counts, the new Graviton4 cores are roughly matching Intel Sapphire Rapids performance while being able to tango with the AMD EPYC "Genoa" and consistently showing terrific generational uplift.
Graviton4 reached general availability this week with initially powering the new R8g instances. Graviton4-based R8g instances are promoted as offering up to 30% better performance than the Graviton3-based R7g prior generation instances. Graviton3 CPUs sported 64 x Neoverse-V1 cores while Graviton4 has 96 x Neoverse-V2 cores based on the Armv9.0 ISA. The Neoverse-V2 cores with the Graviton4 have 2MB of L2 cache per core, twelve channel DDR5-5600 memory, and other improvements over prior Graviton ARM64 processors.
AWS promotes Graviton4 as offering up to 30% faster performance within web applications, 40% faster performance for databases, and 40%+ greater performance for Java software.
Being curious about the Graviton4 performance, I fired up some fresh AWS instances to compare the R8g instance to other same-sized instances. The "16xlarge" size was used across all testing for looking at 64 vCPUs each time and 512GB of memory per instance. The instances tested for today's article included:
Graviton2 - r6g.16xlarge
Graviton3 - r7g.16xlarge
Graviton4 - r8g.16xlarge
AMD EPYC 9R14 - r7a.16xlarge
Intel Xeon 8488C - r7i.16xlarge
All instances were tested using Ubuntu 24.04 with the Linux 6.8 kernel and stock GCC 13.2 compiler.
It would have been interesting to compere to Ampere Computing's cloud ARM64 server processors but not really feasible, unfortunately. With Ampere Altra (Max) in the cloud like with Google's T2A Tau instances, only up to 48 vCPUs are available. And even then Ampere Altra is making use of DDR4 memory and making use of Neoverse-N1 cores... AmpereOne of course is the more direct competitor albeit still not to be found. We still don't have our hands on any AmpereOne hardware nor any indications from Ampere Computing when they may finally send out review samples. Oracle Cloud was supposed to be GA by now with their AmpereOne cloud instances but those remain unavailable as of writing and Ampere Computing hasn't been able to provide any other access to Ampere One for performance testing. Thus it's still MIA for what may be the closest ARM64 server processor competitor to Graviton4.
Let's see how Graviton4 looks -- and its performance per dollar in the AWS cloud -- compared to prior Graviton instances and the AMD EPYC and Intel Xeon competition. The performance per dollar values were based on the on-demand hourly rates.