Interesting workload to test EPYC?

Hello, dear L1Tech community,

I’m a couple years watcher, and lately Wendell has been calling for workload ideas to test CPUs. I happen to be both a Master’s student in computational/analytical chemistry and a hardware nerd. My research depends on using molecules simulated using DFT, a model that’s different and more complex (and accurate) than what biologists use for proteins (molecular dynamics), and it’s computed on CPUs. I do two types of calculations for my project; one scales very well with the number of cores, but the other… is interesting.

Now, I’m working on producing a specific benchmark that will hopefully be reproducible and a good indicator of hardware performance. The insight I’d like to gain is twofold:

  1. what affects its performance; core frequency? cache?? IPC?? RAM speed? My boss is working on computer funds (Haswell workstations falling behind), and I am to choose and/or build two workstations and a server (depending on the budget we get), so I want to choose the most optimal combination of hardware for what our lab does. (Also, I’m not sure if dual sockets means a significant efficiency loss for the software I use.)

  2. Is there a benchmark out there that’s available for most chips that’s representative of, and a good analogy to that specific load?

  3. HOW FAST CAN MILAN GO? WHO KNOWS? MWAHAHAHAHAHA :smiling_imp:

My profile has places to reach me to those interested in collaborating and help me gather datapoints. My software skill level is: I can SSH, install simple Linux distros (no Arch-level of complex, though), install software with a makefile if it doesn’t require too much configuration, and I built a small off-the-grid calculation cluster with OpenHPC (and a well-documented manual) and old core i7’s. I don’t know much about IPMI and servers… I’m a chemist after all, not a computer scientist. Hopefully I posted at the right place :slight_smile:

I’m also totally open to explaining in various levels of detail how those calculations work, and what I do with them. It’s interesting I think.

Toodles,
-Jason

2 Likes

Gimme something to run a script or preferably a walkthrough and I’ll be glad to

That would be awesome @wendell ! I’ll work on that and when it’s ready, I can reply in this thread, or you’d prefer a DM?

1 Like

I don’t know if there is another thread for test CPU workloads so I am posting here.

I’m interested in how nTopology runs on the Threadripper Pro “Sworkstation” beast for no other reason than to see what it can do lol.

nTopology is an implicit modelling and simulation software for designing very complex parts, generally for additive manufacturing. Specifically the topology optimisation and generative design simulation workloads which are very demanding.

I am just tinkering with a student licence of the software and my lowly R7 2700 CPU. I have had to simplify my simulations since it would run all day (or for some unknown length of time).

From my observation of the core loading, it seems to be mixed between single and multithreaded performance. Top-op and generative design works by doing repeated simulations, adjusting the shape of the part or removing material to make the optimum shape. With each iteration, there seems to be a multicore and single core component, the proportion of each depends on the exact simulation setup (larger meshes seem to be more multi thread loaded).

You can also set it to capture each iteration of the objects which then includes storage performance.

There is a free trial for the software, and it’s easy enough to run a simulation, just by opening the nTop file. I could send my file across which is an example case of optimised handlebars I did for one of my uni modules (with the setting tuned up to tax the machine a bit more). I could also look for some other big nTop files as a stress test.

I am only able to get to ~3.6Ghz on my BOINC workload on my 7443P at 200W TDP. That is 400Mhz off the max boost clock of the cpu.

Yet I am able to get to 3.345Ghz on my 7402P with the same TDP and workload. That is just 5Mhz shy of max boost clock.

Try ctdp 220w just for giggles

I did. Apparently just ignored it. Still saw ~ 200W being pulled max on my workloads