This would not be possible without team at F@H: Hugo, Sukrit, Ben, Justin and the rest of the GPU FAHCore team, not forgetting Joseph, F@H lead dev. I know, I probably (definitely) was getting on your nerves a lot in the past couple of years, I wouldn’t even blame you if I was on your ignore list by now
OpenMM team led by Peter Eastman was essential to this becoming a reality, thanks for getting on board with this, and responding to requests for any fixes swiftly.
All the volunteers who provided great insight and ideas on various issues regarding this, we appreciate you all.
Last, but definitely not least, this would not be possible without AMD. Really epic group of people (you know who you are), thanks for opening the doors for this, and enduring all the questions and requests coming your way. When we first had an idea to commit to HIP in 2023, ROCm was in its infancy, and it just exploded since then, and it seems that they are just getting started. Really really appreciate the support.
What is HIP?
In simple terms, this is AMD’s CUDA. It brings AMD GPU compute performance on par with equivalent nVidia cards. Here is the link to the testing results comparing AMD’s OpenCL with HIP (and few nVidia cards running CUDA). The performance increase is massive. (Seeing 60-100% perf jump and 2x-5x PPD jump across various AMD GPUs):
https://drive.proton.me/urls/6J14R2SM38#WVHe5HMIe4Rm
For anyone afraid of the link above:
Note: PPD in the spreadsheet is just for reference, and does not show what you might get in the future. Look at it this way: whatever you were getting before in PPD, double it (as a minimum)
HIP wen?
Hip support will come with the new core_28, you will start seeing new projects coming out for this core (there still going to be old cores in the mix until they are phased out) in next couple of weeks.
What do you need to do to start?
FAH Client 8.5.5 or later and a following guide to get you started:
viewtopic.php?t=43533
The GOOD:
Massive perf increase.
All future GPU FAHCores will support HIP as default
Linux shows no CPU folding impacting GPU folding (take that nVidia
), so you can fold on CPU alongside the GPU, just leave couple of threads free for each AMD GPU in the system.Other, non AMD, cards are seeing up to 4% perf increase with the new core!
The BAD:
As a first try FAHCore_28 supports ROCm/HIP v6.4.4 only.
Fahclient v7 at the moment has issues with the core, so will be excluded from assignments of that core WUs
No Polaris or older support (on Windows or Linux)
Windows performance might be impacted by driver overhead.
Windows drivers might not ship rocm 6 libraries (unlikely, but possible).
On Windows you must leave 2 CPU threads free for HIP compute. Half of the CPU threads free would be optimal, ideally no CPU folding should be happening if you want full PPD benefits from HIP on Windows.
What’s next?
Future FAHCores (core_29) will support other ROCm versions, and more GPUs.
Possible driver optimisations for FAH HIP workloads from AMD
Any issues?
Please, do not hesitate to tag me in here or any other new thread you create with questions. Just be kind and provide as much info about your system as possible.
Good luck and:
Let’s fold them all
