GPU PPD tanks when I start folding on CPU

[BenDreadful]

Ok, so this is what I did.

 

I was folding like normal, only on my gpu like I have been for months and I decide to include my cpu again, to see what the max PPD would be and because I learned about the max-packet-size small tag or option or w/e you want to call it.

 

So I start it up, I add a cpu slot, add the option and it seems to be working. I get a core with base credit of 600 so I know it's short. It starts folding, ETA is 4.5 hrs. Perfect! I leave my pc to do it's work (doing all this through TeamVeiwer since this pc is in another country).

 

After about half an hour I go and check how it's going, are the temps still the same and all that. And I get greeted with my 7970 folding at 100% and getting about 30k PPD compared to the previous 115k+. 

 

I was shocked and try to figure out the issue, go to the logs, but they're normal. Absolutely nothing changed. So I delete the cpu slot and add it again, but specify only 6 cores thinking maybe the gpu needs one more to work with. 

 

Same thing happens.

 

So if anyone knows what should I try next, pls help. 

 

I'm running a 3770k at 4.5ghz, the cpu folding seems stable, I don't think the issue is there. the gpu is an overclocked 7970, when the ppd drop happens, the card stays on 98-99% usage so that's not it.

 

Thanks in advance.

 

EDIT: oh, and it seems that the "max-packet-size" "small" thing doesn't work anymore, I got a 1800 base point WU which takes 14hrs to complete. Am I missing something?

[ProGearUk]

Your CPU has hyper threading so go into configure then slots and tell the client to only use 7 cpu threads. This will allow that AMD gpu to use the cpu as it requires. 

[BenDreadful]

So I delete the cpu slot and add it again, but specify only 6 cores thinking maybe the gpu needs one more to work with. 

 

tried it, same thing happened

[ProGearUk]

tried it, same thing happened

Ok well I would maybe delete both slots and start them again and then let the client choose the amount of threads to use.

[BenDreadful]

ok, so the only thing that worked was dropping the threads to 5, but that means my cpu gets 18k ppd.

 

and the small packet size doesn't work, I get 8703 wu's which take me 15hrs to complete. is this just small, but my cpu makes them seem large?

[Whaler_99]

My stock 3770 CPU usually get's WU's that take a day plus to execute. And, when you are making changes, let things run, like for 10 or 15 hours. The client will do all kinds of random stuff on PPD calculation. Most of us have seen those magical WU's that yield MASSIVE PPD, only to see it normal a few hours later. For the 3770k run at CPU:7 with no options set. Let the GPU run. Watch it for a day or so and see if it sorts itself out. 

 

If it doesn't, it may be an underlying hardware issue with your system and then we need to start nit picking hardware specs and such....

[BenDreadful]

ok, as a couple of days passed, I went with this.

 

I put the cpu to work with only 4 threads since I don't see a real difference when folding with more, but maybe that's me. It's now getting 25k PPD which is ok, I thought it would do more, but I'm not complaining. It still drops my gpu PPD, but only a couple thousands so it's not that significant.

 

The real problem I have is no small WU's. I ran it without any options set, like you said, but since that gave me about 29k PPD I don't see any reason why put it at CPU:7 compared to CPU:4. Then I tried to enable the max-packet-size option, but still nothing. Maybe it needs a day or so to do it, who knows, I've been running it for about 5hrs, 8 more and the WU will be finished.

[Whaler_99]

I know a bunch of the older flags no longer work, especiually as the WU's are becoming more standardized. The only benefit to going CPU:7 versus 4 should be a bit more dedicated core affinity, so your TPF should improve, boosting your PPD - but, and here is where some teasting comes through, if you are running HyperThreading, 4 of your "cores" are virtual, so how much more can you get from a few extra virtual cores? One reason why AMD rules in the bigger systems - all the cores are real cores, whereas many Intel Xeon systems still use HT for a high core count.

[BenDreadful]

How are amd cpu's in smaller, "normal", 1p configs? Would my 3770k beat the amd equivalent, whatever that may be, if we compare them 1vs1?

All I could find on that subject was 3+ years old or regarding a 2p-4p system.

In short, if I were to build a pc for everyday use and I plan to fold with it, should I go amd or intel?

And yeah, I gave up on small wu's for my cpu, it would be great to have them, but the option isn't working :(

[Whaler_99]

For "raw" computational power AMD has always had an edge over Intel - so if you are buidling single CPU boxes dedicated to Folding, go high end AMD. If more a general use system AND folding, go Intel as they have a slight edge there in day-to-day use....

[BenDreadful]

Yeah, thought so. And the whole linux for cpu and windows for gpu makes it a bit hard to justify getting a good cpu for a rig which will do most of the work with the gpu.

I guess we'll have to wait until gpu's can get at least some points in linux.

[Theo]

Oh just a random thing, is it better to fold on windows, or on linux VM running on windows? Ive seen some posts like that and always wondered

[BenDreadful]

Oh just a random thing, is it better to fold on windows, or on linux VM running on windows? Ive seen some posts like that and always wondered

Linux will give you about a 20% ppd increase in cpu folding compared to windows or sometimes even more (that's just what I've heard, could be wrong), and running a vm on windows should in theory give you a boost, but the thing is your cpu is working extra to run the vm so it could just cancel itself out.

Worth a try imo, I'm thinking of running a vm, just don't have the skills or knowledge in linux so I'm hesitant to even start :P

[Whaler_99]

OK... Linux in VM on Windows would ONLY be worth it if for some reason you HAVE to have windows on your box and you have more the 16 real cores, not hyper threaded ones. Then, you could use something like an old version of VM Server that still allows you to assign 16 virtual CPU's. Hyper-V and new VMWare don't. I believe also the Sun/Oracle one does as well.

 

Basically with 16 cores, you get access to bigadv WU's that grant QRB (quick return bonus). But, bigadv are only supported under Linux, not Windows. Also, they are BIG. Having an 8 core box with hyper threading, so 16 cores, will mostly fail, because you cannot finish it in time. 

 

Now, a install of Linux on a server will get you more PPD then a Windows server running FAH. But running Linux in a VM on Windows, not much point.... Just run FAH natively, unless you have a LOT of cores. :)