We benchmark the recently released HPC platform: Amazon-Graviton3E.
Amazon recently made available the HPC version of Graviton3 named Graviton3E. According to them, the new Hpc7g instances provide up to 35 percent higher vector instruction processing performance in relation to the simple Graviton3. Additionally, Graviton3E provides two times better floating-point performance in comparison to Graviton2. All this comes with 60 percent more energy efficiency, in comparison to comparable x86 AWS instances.
We have measured the performance of the newly available Graviton3E. In regard to memory bandwidth, we used SVE enabled STREAM TRIAD. As we can see, in the following figure, we reach ~230-240 GB/s from the 8 available DDR5 controllers of the machine, when we use 32-64 threads. The theoretical peak for the memory syst,em is 307.2 GB/s according to this nextplatform post.
The next benchmark that we looked into is Likwid peakflops, which can measure the peak performance of the machine in Gflops/s. With that we can see that Graviton3E achieves astonishing scaling performance, as it maintains >99% per CPU core efficiency even when using all 64 threads. The simple Graviton3 platform, drops below 90% per core efficiency when using only 8 threads, and 63.54% when using all 64 threads. This shows that Graviton3E can deliver maximum execution efficiency even when all threads are running and performs much better in this regard in relation to it’s sibling Graviton3. And for the record, the Graviton3E reaches 2,644 Gflops/s, with 64 threads.
Leave a Reply