CPU and PCI-E speed on GPU performance
20 Oct 2023 14:04:24 UTC
Topic 230226
(moderation:
I'm trying to decide which of my systems would be best suited to serve as GPU-only Einstein-dedicated boxes.
To this end, I was wondering how important are the CPU and PCI-E interface speed. For instance, would an i5-4690 bottleneck a RX 6650 XT? I got the RX 6650 XT for einstein crunching only to replace older GPUs that were no longer usable. I'm wondering if I made the right choice. It seemed like a good option from a price to performance ratio. It's ok, but not doing as good as I expected compared to my 3080 ti. Thoughts on good current options (balancing price, performance and energy use)?
For my older GPUs (that remain usable), I'd like to reuse an old AMD A10-6790K. The single thread performance is about 70% of the i5. The mobo has two PCI-E (2.0) x16 slot but one is only 4x. It would run a RX 480 and a GTX 960. I will likely conduct some experiments but I'd be interested in knowing from the experiences for peoples here.
Thanks,
borghub wrote: I'm trying to
)
Einstein only requires a single cpu core to do it's thing with the gpu tasks, so if you set your settings to only uae 99% of the cpu cores you will be fine, or as you suggested just not run any cpu tasks. The thing for gpu's is to run a single task at a time and keep a log of how long each task takes, then try running 2 tasks at a time, then 3 tasks etc. at some point you will find it taking longer that running a single task at one time and that's the 'I went too far' stopping point for your gpu. Each gpu is different depending on the memory they have and the speed etc etc etc, sure there are generalities but your pc and your hardware running at your settings is what you need to test. AND each Application for the gpu will be different, again generalities apply but KNOWING BY TESTING is the only way to be sure. The other thing when running multiple tasks at one time is to pause the 2nd, 3rd etc task a little bit so the cpu intensive part isn't happpening at the same time for each task beyond a single task. This happens at the end of each task and others can tell you more prcisely when that heppens in the progress of the task, but what happens is the cpu does alot of work to finish the task and prepare it to be sent back to the Server and if you have multiple tasks doing this at the same time it will slow things down.
+41
)
+41
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
PCie is only going to affect
)
PCie is only going to affect how fast the data is loaded and stored to/from memory to/from the gpu. Once the data is in the gpu then PCie speed (Gen2 Gen3 etc) doesn't matter.
PCie is fast, really fast. You can look up the transfer rates with a google search or just go to wikipedia. About half way down the page is a chart. Each PCie gen about doubles the transfer rate for a given link width. A PCie 2.0 by 4 (correct terminology for x4, not times 4, x means "by") lane can transfer 2 GB/s (giga bytes, not giga bits) which is more than fast enough. Of course this is somewhat dependent on the DMA capabilities of he chipset, if it can keep up, and i am fairly confident it can, then you are good to go.
Oh wait: Whats your memory transfer speed, the low value, not using cache? That's going to be the limiting factor in transfers to and from the gpu. The slowest part rules the overall speed of a transfer to or from a gpu
The cpu speed is bound by its compute capability, the cache basically negates most of the speed issue to and from memory. **for the cpu** The i5-4690 is an old cpu (2014? not sure ) so it's not going to be all that fast but it is adequate. I think you can swap the cpu out for a 4790k which is overclockable - its something to think about as i am pretty sure the 4690 is locked at one speed and can't be overclocked.
Make sure your system is clean and dust free, and if you haven't done so already, replace the thermal paste.
borghub wrote: For instance,
)
In one of my older rigs, my i5-4460 was bottlenecking a RX 5500XT. It could be seen and heard (fans stopping intermittently) that the CPU could not deliver the data fast enough to the GPU. And it was a CPU similar to yours and a much slower GPU.
my blog about raspberry pis and diy life
stfn wrote:borghub
)
Does it happen around the 90% or so point of crunching? If so it's because of the 64 bit nature of the data and the fact that most gpu's now have very poor 64 but the data requires it so the task is 'off-loaded' to the cpu which dramatically slows things down. Once a new task starts it's back onto the gpu again and then back to the cpu for the final @10%.
This thread explains it alot better:
https://einsteinathome.org/content/fgrp5-cpu-and-fgrpb1g-gpu-why-does-crunching-seem-pause-90
In addition a study done a LONG time ago showed that Boinc does not benefit by having more than an 8x pci-e slot, now todays faster gpu's could mean that that study needs to be redone but I have no memory of who did the test the first time around.
mikey wrote: Does it happen
)
No no, it was happening for other task types, and intermittently throughout the whole task run time. I am very well aware of the 90%-move-to-CPU with the FGRPB1 tasks :)
my blog about raspberry pis and diy life