Does anyone have experience using these machines that they could share? One of the draws for them is the large number of CPU logical cores available on a single chip, upwards of 100-200+ using SMT4 or SMT8, which means a single server could handle as many parallel threads as are typically spread across multiple nodes on regular Intel based HPC systems, for a fraction of the price.
But the obvious drawback is that software compiled for Intel architecture will likely not work. Some, but not all, of this solved by using the BioBuilds conda channel. Would obviously be easier if you could just build a Docker or Singularity container for your software, but you end up in the catch-22 of needing access to a Power server with admin rights to try to build your software for use on the Power servers you don't have admin rights to.
New supercomputers are boasting enhanced genomics performance on these machines, but what are other users' thoughts? Note that the retail price for a single AC922 unit as mentioned there seems to be ~$50k for a base unit (2x16 core CPU, SMT4 = 128 logical cores, SMT8 = 256 logical cores), with almost half the cost coming from the 2 GPUs, not sure if they're available without them.
Considering how expensive they are (and the specialized skills they likely need) it is not surprising that we don't find many mentions of Power CPU's in bioinformatics. They remind me of Itanium servers back in the day.
Developers mostly develop on hardware that is easily available to them and their users. While one can get around some problems expect to fight every software release that comes out, if you use specialized hardware such as this.
Possible but will need to be confirmed. A server is a sum of all parts. It is not enough to have the beefiest CPU around. It needs to be consistently fed with data. I am not sure what kind of high performance storage system Oak Ridge server uses. That likely contributes significantly to the overall power of the system. On a single AC922, I bet those cores would be starved for data with the fastest consumer SSD we can get today, if one tries to run it standalone.
as a follow up to this a few years later, its worth noting that AMD Epyc is now a thing, with core/thread counts rivaling that of IBM's Power CPUs, supposedly without the issue of needing custom software compilations.