AWS Graviton4 vs. AmpereOne 192-Core Benchmarks For Leading AArch64 Server Performance
With the srsRAN 5G RAN software solution we see Graviton4 consistently delivering better performance both on a per-core/thread basis as well as for the overall multi-threaded performance.
AmpereOne pulled off a small lead with the Primesieve prime number benchmark.
When taking the geo mean of 60+ benchmarks conducted on both 192 core configurations, the AWS Graviton4 192 vCPU instance was 56% faster than the AmpereOne A192-32X on average. (Due to cloud costs in paying for the Graviton4 instance and the tight finances in the web publisher space these days, I had to keep costs down rather than my preference of typically running hundreds of benchmarks.) Particularly with HPC-type workloads and other scientific computing scenarios the Graviton4 with its Neoverse-V2 cores were delivering significant leads over the custom AmpereOne cores. Again, both AmpereOne and Graviton4 were running on clean Ubuntu 24.04 LTS environments with Linux 6.8 64K kernels and the GCC 13.2 default compiler. With some single-threaded simpler workloads like PyBench for Python performance we saw Graviton4 outperforming AmpereOne pointing to the Neoverse-V2 cores having stronger per-core performance than the custom Ampere-1 cores.
This though does just tell the story of the raw performance with unfortunately being unable to access any CPU/system power sensors on Graviton4 for comparing the total power consumption and performance-per-Watt. It's unfortunate not being able to share that insight for if for example the AmpereOne server is much more power efficient than Graviton4. Similarly, there isn't the ability to offer a 1:1 cost/TCO comparison. If going on the Oracle Cloud A2 pricing against the Amazon/AWS EC2 R8g pricing, the AmpereOne A2 instances look like on a performance-per-dollar basis they'd be roughly comparable or at least competitive to the AWS Graviton4 pricing on an on-demand/hourly basis. But with AWS Graviton4 servers being unavailable for purchase / on-premise deployments and the AmpereOne processor/server availability still ramping up, there isn't much more to definitively share on the cost/TCO at this point. At least with these raw performance numbers being out, those considering AArch64 cloud options can at least use these performance numbers for better comparing the value of Graviton4 against the likes of Oracle Cloud A2 instances.
Hopefully with the upcoming AmpereOne M processors offering 12 channel DDR5 memory and any other enhancements we'll see more competitive performance against Graviton4. AmpereOne M will hopefully be out next quarter but given the long ramp time we've seen with AmpereOne thus far, I'll wait to see it until considering it available. Next year with AmpereOne MX for up to 256 cores will put AmpereOne ahead on a maximum core count basis at least until seeing what Amazon Web Services has coming down the pipe for Graviton5.
It is also worth reiterating that AmpereOne was engineered as a 2022~2023 product but it's taken now until H2'2024 for its production to really ramp up to the point of seeing the Oracle Cloud A2 general availability and seeing this Supermicro server as a temporary review sample. If looking back at the Graviton1 / Graviton2 / Graviton3 / Graviton4 benchmarks I did earlier this summer, top-end Graviton3 to Graviton4 was a 55% jump in the geo mean performance... So had AmpereOne been readily available a year or two ago as originally envisioned, it would have either been a neck-and-neck competitor to Graviton3 or outperforming it more broadly. Instead AmpereOne availability is improving after the Graviton4 general availability this summer. Moving forward it will be interesting to see if the Ampere One M/MX ramp is quicker and how follow-on launches are timed especially compared to Amazon Web Services with its regular cadence for new Graviton processors.
It's certainly great seeing the rigorous AArch64 server processor competition and these parts becoming competitive to x86_64 AMD EPYC / Intel Xeon processors in more areas including power efficiency and value. Graviton4 performance continues to prove very impressive and I was surprised by how much faster it was than the AmpereOne A192-32X in a range of workloads. But for those after a local/on-premise AArch64 server, Ampere Computing wins by default at this point with Graviton servers not being available for purchase outside the confines of the AWS cloud.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.