Ampere Computing Announces AmpereOne With Up to 192 Cores Per Socket

Ampere Computing is reporting far more VMs per rack as possible with AmpereOne, to which they are basing their assessment just on the number of physical CPU cores possible and one VM per physical core. So this claim is rather dubious in effect just reporting the maximum number of cores per rack possible without any power/performance potential of the said VMs.

The initial benchmarks provided by Ampere for AmpereOne are rather limited: a benchmark of Stable Diffusion for generative AI and then using DLRM for AI recommendations. AmpereOne reportedly offers 2.3x more frames per second than an AMD EPYC 9654 "Genoa" and more than twice the number of queries per second for DLRM as the EPYC 9654. Unfortunately, these two AI performance benchmarks is all that's shown for today's announcements.

One item that is important to point out was when going through the end notes of their presentation, with the DLRM benchmark they are testing AmpereOne using FP16 while for AMD EPYC 9654 was tested using FP32. AmpereOne saw a system power average during DLRM at 534 Watts to AMD EPYC Genoa 9654 at 512 Watts.

Ampere Computing ended their briefing by showing their wide arrange of partners from hardware vendors like Giga Computing, Foxconn, HPE, and Supermicro to public clouds like Microsoft Azure, Tencent Cloud, Google Cloud, and others. This is indicative of their current partners and no word yet when we'll see any AmpereOne hardware or begin seeing these 192 core server processors in the public cloud.

I was told during my briefing that thanks to their early kernel upstreaming as well as already having the compiler support for GCC and LLVM Clang in place, the major AArch64 Linux distributions should be ready for running on AmpereOne. Those earlier compiler patches confirmed AmpereOne as being based on the ARMv8.6 ISA.

That's it for now with the information from today's Ampere Computing roadmap disclosure. Hopefully soon we'll find our hands on AmpereOne hardware for testing to provide an independent look at its performance and power efficiency across a wider variety of benchmarks.