This Week in MLC@Home
Notes for Feb 1 2021
A weekly summary of news and notes for MLC@Home

Summary
Paper coming together. The first round of datasets are cut and I will release them by the end of this week. Only issue is their size. I will likely make the full, larger datasets available as a torrent. Stay tuned for some new results and the big dataset release this week. The paper will follow shortly afterwards.

Detailed News

• A few people have asked if they can help with the client itself. The MLDS app is open source, and available for perusal at https://gitlab.com/mlcathome/mlds . I've created a number of issues under https://gitlab.com/mlcathome/mlds/-/issues to show thigns that need to happen. Some of which could potentially be tackled by other developers.
  
• Only issue is their size. I will likely make the full, larger datasets available as a torrent. When released, please consider seeding if you can.
  
• We've made our goal of having over 1000 entries for DS1/DS2, which allows the dataset release. This means we'll start mixing in other types of WUs than just DS2 WUs into the GPU queue again, balancing between the remaining DS1/DS2 and DS3 WUs. When DS4 is ready, we'll rebalance again.
  

Project status snapshot:
(note these numbers are approximations)






Last week's TWIM Notes: Jan 26 2021

Thanks again to all our volunteers!

-- The MLC@Home Admins(s)
Homepage: https://www.mlcathome.org/
Twitter: @MLCHome2projuct-summary-

his means we'll start mixing in other types of WUs than just DS2 WUs into the GPU queue again, balancing between the remaining DS1/DS2 and DS3 WUs. When DS4 is ready, we'll rebalance again.

Is there a "minimum" gpu for this project?
I have an entry level RX 550x and i don't know if it is usable...

GPUs need 2GB ram min.

RX550 is the minimum AMD GPU that will be supported, as I think the RX540 is not POLARIS-based (gfx803 in AMD nomeclature). POLARIS needs an updated client compiled against a later version of rocm (rocm 3.8 has a bug that keeps it from running on POLARIS). It's on the todo list but not ready yet. For now, the only AMD GPUs supported are VEGA-based (gfx900/906/908, not gfx902 APUs) and must be running Linux. Last night I just got pytorch working with rocm4., so that should be easier in the future. There's *some* support for NAVI in rocm, but its untested.

Note AMD APUs are not supported by rocm, and thus aren't supported by pytorch. Yes that's sad.

On the CUDA side, you need 2GB ram and compute capability 3.5 or higher. That's,... most GTX 700 series and up I think?

I am not clear on how to get tasks from the "test" queue. According to the apps page, it requires "Linux running on an AMD x86_64 or Intel EM64T CPU". But if I understand correctly from the above posts, the app is for a GPU, not CPU. FWIW, I do have an AMD VII and the settings are configured to allow both GPU and CPU for the "test" application. Also the server status page says there are 40 tasks available with 8 in progress. Yet I am unable to get any, either for CPU or GPU. What am I doing wrong?

---

Reno, NV
Team: SETI.USA

Hey everyone,

I'm glad to see the RX 550 GPU is staying on the list of supported GPUs. This is the only BOINC Project I am running atm

I am happy to answer any questions the leaders would like to ask about my rig. Or if someone could go more into detail about running the program more efficiently or strenuously. I plan on upgrading very soon my GPU to put it in a smaller rig. I will stick around the boards for the next few posts




So long and thanks for all the fish.

---

I've seen you are running only cpu-wu's. If you get your RX550 up and running please leave a note here. I have a backup system with an RX570 and could easily install another disk with linux - if there is a chance to get it working.

The AMD client *really* needs an update. Maybe I'll try and give it some love this weekend. PyTorch 1.8 with official ROCm support was just released today, so maybe its time to refresh all the clients. Plus pytorch updated their static compilation options, so maybe we can go that route this time too. That would be a huge win.

The current AMD client won't support polaris (rx5xx) due to a bug in the version of rocm its linked against (3.8). Also, I know of only one other person who has gotten the rocm client to work, and their WUs revealed there's still a library trying to link against their system's version of miopen (a rocm library) instead of the one we ship (not a huge deal, it'll work, but it should not be doing that).

Let me see what I can do this weekend.