That's a lot, and yes, my Ubuntus are are AMD. But its detailed above for someone to try it out for AMD.

BTW the numbers I gave above exclude servers. Don't have the stats but I bet they are reversed in favour of Linux. If I had to build servers it would be one of the flavours of Linux. When I think of it that is exactly what I did except I used a desktop install.

I've seen this quote in a number of places

Can't help with the 7900X, but the 9354 takes about 52mins per task - although that's in a VM on top of VMware ESXi...

It's a dual-socket Dell R7625 server so it's got two 9354's in it. I need to have hyperthreading turned off for some of the work stuff I do now and then though, and I wanted to leave 2 cores free for ESXi and another 2 for the vCenter VM, so the BOINC VM is running on 60 physical cores.

Also have an R6625 box too, which has another two 9354's in it, although I need to use that for work more often than the other box, so the VM on that one doesn't process as much.

Pretty damn cool seeing 60 tasks running concurrently though, and they proper burn through the tasks :)

The R7625 is usually running for about 14hrs too, although it depends on whether I have to pause BOINC to do some work stuff. Most days I need the R6625 too much to allow BOINC to run, so that one usually only crunches a couple days a week. Also, with the R6625 being a 1U server rather than 2U, it's quite a bit louder under load and for some reason that I've not had chance to look into yet, the fans cycle up & down rather than staying constant, which is distracting and annoying with my lab being in the next room to my office.

A full 14hrs or so of BOINC on the R7625, mainly running MCM with a few SCC tasks thrown in, gives about 31 days runtime, returning about 810 results for about 492k points.

I only came back about a month ago after having taken a few years out. Initially I just set BOINC up on my workstation which has an i9-12900K, then I realised I may as well set a couple of servers in my test lab to work. Once I saw how much they were doing it makes my workstation look insignificant, so I decided to stop the workstation and have a cooler office instead!! :)

I've not tried reducing the core allocation, no, but with hyperthreading disabled and two CPUs in the server, all 60 cores allocated to BOINC are 'real' physical cores, so by reducing to 30 cores I'd basically only be using a single CPU, so I doubt it would increase throughput. There are no other bottlenecks in the server; it's got 1TB RAM and 8x high-end enterprise NVMe drives in RAID10 - the best part of working for Dell is getting lots of cool tech to play with :)