AWS Graviton4 vs. AmpereOne 192-Core Benchmarks For Leading AArch64 Server Performance

With the Supermicro ARS-211M-NR R13SPD server that's in the lab for a few weeks for reviewing the AmpereOne A192-32X and delivering the first independent benchmarks of the AmpereOne 192-core AArch64 server processor, the AmpereOne benchmarks to date have been comparing to other Intel Xeon and AMD EPYC server platforms. But if looking up to the cloud is the closest AArch64 server competition to AmpereOne there is: Amazon's Graviton4. In today's article ia showdown looking at how AmpereOne and AWS Graviton4 compete at 192 cores for ARM 64-bit server performance.

AmpereOne can be found in the cloud with Oracle Cloud A2 being the only generally available public cloud instances powered by AmpereOne thus far. Or it's beginning to surface in the channel for those wanting on-premise AArch64 server platforms with up to 192 cores like we see with the Supermicro ARS-211M-NR R13SPD platform. With AWS Graviton4 it's only available via the Amazon Web Services EC2 cloud but has proven to be a formidable competitor. Graviton4 reached GA this summer with the R8g instances and delivered very competitive performance to AMD EPYC and Intel Xeon EC2 instances. It's also been striking how well the Graviton1 to Graviton4 performance has evolved in a short period of time. Mind you when Ampere Computing originally announced AmpereOne in May 2022 they said at the time they were sampling multiple customers with a formal launch expected later in 2022 -- with a very slow ramp, we are only beginning to see AmpereOne availability increasing now in Q3'2024. Graviton3 went GA in mid-2022 and now Graviton4 reached GA with R8g instances in the EC2 cloud this summer all weeks before Oracle Cloud's A2 instance availability.

Graviton4 as a reminder has 96 Arm Neoverse-V2 cores using the Armv9.0-A ISA and offers 2MB of L2 cache per core and 12 channel DDR5-5600 memory. Graviton4 can be available in a two socket solution for having 192 cores -- matching the top-end AmpereOne single socket processor configuration. AmpereOne scales up to 192 cores in a single socket, matches Graviton4 with 2MB of L2 cache per core, uses a custom core design by Ampere Computing, and currently offers eight channel DDR5 memory at DDR5-5200 speeds. Per this year's AmpereOne roadmap update in Q4 they are hoping to have AmpereOne M with 12 channel DDR5 memory support in Q4 that would match the memory channel count currently available with AWS Graviton4. Next year is when AmpereOne MX is on the way with up to 256 cores.

For those after on-premise AArch64 servers for CPU-focused workloads, Ampere Computing basically wins by default. Ampere Altra (Max) is widely available still with up to 128 cores and AmpereOne availability will hopefully continue to improve. With NVIDIA there is the likes of the NVIDIA GH200 for high-end AArch64 servers but there really only making sense if engaging GPU-accelerated workloads. And then with the Graviton processors they are only available within Amazon's cloud. But in any event I was curious to see how well the AmpereOne A192-32X bare metal server can compete with 192 cores of Graviton4 using the r8g.metal-48xl instance for 192 vCPUs between two Graviton4 processors.

Due to running the Graviton4 benchmarks via the cloud and even for the "metal" r8g.metal-48xl type there still is no access to any CPU or system level power monitoring exposed by the system to user-space. Thus for this article is simply a battle of raw performance with being unable to compare to the CPU or server level power consumption of Graviton4. Those wondering about AmpereOne power consumption can see my earlier AmpereOne A192-32X review with all the power numbers there.

With also being only in the cloud it's tough to compare TCO... Plus AmpereOne availability is still tight in the retail channel and I haven't been provided with any Supermicro ARS-211M-NR R13SPD cost estimates. All we know right now is the AmpereOne A192-32X has a $5,555 USD list price. Thus the best that can be done on the cost comparison side is going off cloud costs for each. Oracle Cloud prices the AmpereOne A2 at $0.014 USD OCPU per hour and $0.002 gigabyte of system memory per hour. So for 192 A2 AmpereOne CPU cores and 1536GB of RAM, that's around a base pricing of $5.76 USD per hour in Oracle Cloud. For AWS with the r8g.48xlarge / r8g.metal-48xl instances for 192 Graviton4 vCPUs and 1536GB of RAM is around $11.31072 per hour for an on-demand rate. In general though Oracle Cloud tends to be more competitively priced and not a 1:1 comparison for cloud instances from the different public cloud providers.

For this raw performance comparison the AmpereOne A192-32X (192 cores) was running on the Supermicro ARS-211M-NR with its eight channel DDR5-5200 memory compared to the Graviton4 with the r8g.metal-48xl providing 192 vCPUs of Graviton4 (Neoverse-V2) with its 12 channel DDR5 memory. Both the AmpereOne server and AWS Graviton4 instance were tested using Ubuntu 24.04 LTS with the Linux 6.8 kernel and GCC 13.2 as the default code compiler and other Ubuntu 24.04 LTS package defaults aside from running both Graviton4 and AmpereOne with the 64K kernel page size rather than the default 4K page size option.