Radeon VII @ newegg for 500 dollars US on 11-27

[ewmayer]

In checking the status of my Mlucas-on-CPU, GpuOwl-on-Radeon-7 runs today, I noticed the former is running ~20% slower today than it should be. 'top' shows a process "kworker/0:0-events" eating a significant % of the CPU cycles. Does anyone have a clue what that is? A hint to reboot, perhaps?

I also see tasks named "kworker/0:0H-kblockd" and "kworker/1:0H-kblockd", though they're not consuming significant cycles.

EDIT: Found this, which points to USB power management in the kernel. I did unplug/replug the USB keyboard a little while ago, maybe that triggered the problem.

[kriesel]

Originally Posted by preda

on Linux you could redirect output to a file, or to /dev/null

./gpuowl options > /dev/null

Yes. On Windows I think what was getting directed was the batch file invocation. On colab there was a compound command line mucking things up.

[preda]

I see the same. I'm not sure what it causing it, but I suspect some aspect of ROCm interacting with the kernel. In my case, what seems to fix it somewhat is to stop *all* gpuowl instances (i.e. reach a state with none running), after which I can restart them and it will take a while until the kworkers pick it up again.

Originally Posted by ewmayer

In checking the status of my Mlucas-on-CPU, GpuOwl-on-Radeon-7 runs today, I noticed the former is running ~20% slower today than it should be. 'top' shows a process "kworker/0:0-events" eating a significant % of the CPU cycles. Does anyone have a clue what that is? A hint to reboot, perhaps?

I also see tasks named "kworker/0:0H-kblockd" and "kworker/1:0H-kblockd", though they're not consuming significant cycles.

EDIT: Found this, which points to USB power management in the kernel. I did unplug/replug the USB keyboard a little while ago, maybe that triggered the problem.

[ewmayer]

Originally Posted by preda

I see the same. I'm not sure what it causing it, but I suspect some aspect of ROCm interacting with the kernel. In my case, what seems to fix it somewhat is to stop *all* gpuowl instances (i.e. reach a state with none running), after which I can restart them and it will take a while until the kworkers pick it up again.

Ah, so you're running Prime95/mprime on your CPU, similar workload as my own. Do you see the CPU-job slowdown at fairly regular intervals? If so, could you try the simple edit of the /etc/default/grub file described at the link I posted above, and see if - following the next reboot following the edit - the problem goes away?

[paulunderwood]

Originally Posted by ewmayer

Ah, so you're running Prime95/mprime on your CPU, similar workload as my own. Do you see the CPU-job slowdown at fairly regular intervals? If so, could you try the simple edit of the /etc/default/grub file described at the link I posted above, and see if - following the next reboot following the edit - the problem goes away?

If you do the edit

GRUB_CMDLINE_LINUX_DEFAULT="<existing stuff> usbcore.autosuspend=-1"

don't forget to run sudo update-grub.

You can always try out the argument by non-permamently altering the kernel options at boot -- select the kernel and press "e" (for edit) at the boot selector screen for Ubuntu when you boot the computer. It is a bit of a clunky operation to add the argument. Having added the argument to the kernel entry press "b" for boot.

[ewmayer]

So I'm looking at building a multi-GPU ATX-case system - got a 850W PS, idea is to have that power 3 Radeon VIIs, if that can be done while leaving a little headroom for LL/PRP-testing on the CPU, fine, but the latter is dispensable-with.

I'm going to track a few eBay Radeon-VII-lightly-used auctions over the next few weeks (or months) to get a feel for pricing. In the market for 3 of these, hoping to be able to score them for ~$400 each, perhaps by targeting odd-hours-ending ones and whatnot.

But first, gonna be looking for a (likely new) ATX-compatible mobo+cpu, needs at least 3 PCI3.0 slots, maybe an octocore AMD Zen* CPU ... any recommendations welcome. Maybe something like the following combo?

CPU: https://www.amazon.com/AMD-Ryzen-370.../dp/B07SXMZLPK

MoBo: https://www.amazon.com/MSI-X570-Gami.../dp/B07T2FWCYG

But after pasting the above links, another thought occurred to me re. conventional case-system build: Since the CPU is gonna be a minor player GIMPS-work-wise, and may not even being doing any crunching due to the poor flops/watt vs GPUs, no point blowing extra $ on it - just get the cheapest reasonably-recent CPU model, more important is that MoBo has at least 3 PCI3.0 slots. Hell, even a used CPU+MoBo would work, as long as they're in good working condition.

Mike/Xyzzy suggested looking into possibly purchasing a used mining rig, but e.g. on eBay I didn't spot any based on Radeon VIIs, those GPUs seem to be in high demand, for good reason. So any ATX-case-system remains the go-to option. (Ah, I will need a new or used case, since I plan to keep my old Haswell system up and running with its single Radeon VII - the pronounced glitchiness of the Haswell CPU actually makes it good for Mlucas QA testing, perversely enough.)

Target is to keep the entire build cost under $2000, so let's strategize!

[Prime95]

If you get an eBay Radeon VII, will it be covered by manufacturer's warranty? When RMA'ing mine I had to upload a copy of the invoice.

[preda]

Originally Posted by Prime95

If you get an eBay Radeon VII, will it be covered by manufacturer's warranty? When RMA'ing mine I had to upload a copy of the invoice.

amazon has new ones available (XFX).

BTW, I recommend against physical modifications that void the warranty, as these GPUs do die from time to time for obscure reasons and the RMA is a lifesaver then.

[ewmayer]

Originally Posted by Prime95

If you get an eBay Radeon VII, will it be covered by manufacturer's warranty? When RMA'ing mine I had to upload a copy of the invoice.

Had an e-mail exchange with PhilF about this, using one current eBay used-4-sale one as example:

Originally Posted by ewmayer

Hi, Phil - so let's say I want to buy several of these to put into a future multi-GPU dream build. New on ebay are going for more than Amazon charges for them, so let's look at used. I see something like this:
That says "No returns, but backed by eBay Money back guarantee". What does that mean, specifically? "If it arrives nonworking, I don't want it back but you get your money back?"

Originally Posted by PhilF

It means you can't return it just because you changed your mind, but eBay guarantees it will arrive in the condition stated by the seller. In this case, the seller says in the description:

Tested and 100% perfect working condition.

So if it arrives in any other condition, eBay will force them to refund you after you return it because it wasn't "as described". Also, eBay almost always sides with the buyer if there's a problem.

Originally Posted by preda

amazon has new ones available (XFX).

BTW, I recommend against physical modifications that void the warranty, as these GPUs do die from time to time for obscure reasons and the RMA is a lifesaver then.

Yes, my current one was bought new there - $550, so ~$600 including CA sales tax. That's still a possibility obviously, but I'd like to see, as noted, if I can get the total build cost under $2000, by way of a budget-crunching challenge.

Only physical modification I would consider - and I hope to avoid this by buying a different ATX case than current one - is tweaking the mounting bracket, like I had to do with my Amazon-bought card to be able to screw in the case-back PCI-device fixing screws. But this time, should any such fiddling prove necessary, I would first lay the case on its side so I can slide the GPU into the PCI slot without needing the case-back screws to hold it there, so as to be able to test it and make sure it works.

[Prime95]

Three cards in an ATX will run pretty (very?) hot. I do recommend a cheap open-air type mining rig ($20 - $25) unless you have cats or small kids that might injure themselves in the fans or exposed circuitry.

[PhilF]

Originally Posted by Prime95

Three cards in an ATX will run pretty (very?) hot. I do recommend a cheap open-air type mining rig ($20 - $25) unless you have cats or small kids that might injure themselves in the fans or exposed circuitry.

There's another reason for a full case. I am a ham radio guy, so I can't tolerate the radio frequency interference caused by unshielded computer equipment.

I can highly recommend the case linked to below. I got one on sale because the sale price was so cheap (even cheaper than the current sale price), and I quickly became highly impressed.

First, it doesn't have a needless glass side. It is fully shielded. Second, in addition to the 3 case fans that comes with it, there are provisions for 2 more large fans right at the place where the heat from the GPU(s) needs aggressive removal.

The coolest thing of all (no pun intended) is that the power supply mounts at the bottom of the case, and can be mounted upside down. This means it can draw its own cool air supply from outside of the case at the bottom (it even comes with a removable dust filter for it), meaning you don't have to worry about powerful case fans starving the power supply for air.

The only drawback is there could be a problem with accommodating full size cards like the Radeon VII without having to modify it. It has all kinds of room for drives that might have to be removed in order to make room for cards.

https://www.newegg.com/p/N82E16811147153

Three cards may be too much heat for a full case, but they are still worth taking a look at in certain cases. Sorry, again no pun intended.