Maximizing PPD discussion / testing results

[the garynator]

Figured I'd start a thread on this. This was prompted by some posts I made in this thread (https://www.overclockers.com/forums/threads/settling-into-the-6-spot.805575/).

The purpose of this thread is to quantify differences in performance of cards on different platforms, and to post the results of my testing, as well as observations. It would be awesome to get some more data to compare / other suggestions / etc. My goal is to gather enough info in here to show how much PPD can be garnered via upgrades other than a different video card. I believe our team has a lot of PPD available with just optimizing things like matching video cards with the right systems in the case of someone with a farm, free things for dedicated folding rigs as well as things like dual booting systems like a gaming PC that you want windows on but doesn't get used when folding, and cheap upgrades like finding cheap CPU/Mobo/Ram upgrades that can garner similar amounts of points vs upgrading to a faster GPU.

Some info I posted in this post (https://www.overclockers.com/forums/threads/settling-into-the-6-spot.805575/post-8213466):
"I'm snagging some data from the 4070ti while it's in the 2500k system. Preliminary results show a significant decrease in PPD vs pcie 3/4 systems with faster cpus. Not sure how close these differences translate compared to older cards but I've seen some people running 30xx and 40xx cards in older systems. There could be a significant amount of efficiency and PPD on the table with fairly cheap upgrades.

I'll have to quantify it based on the data I've collected, but for example, my overclocked 4070Ti in my 2700x @4GHZ (pcie 3.0 x16) is pretty much on par ppd-wise with a stock 4070 super in the 5900x (pcie 4.0 x16) in Ubuntu. There's normally about a 20% difference between a stock Ti vs Super. The stock Ti in the 5900x, even with the super in it (so pcie 4.0 x8) in ubuntu, out performs the overclocked Ti in the 2700x system by a couple million PPD.

A Pcie 4.0 board, CPU and ram can usually be had for like $350 or less on Facebook marketplace. Heck, the Dell 5820s Kyle has can be snagged for as little as $100 (add a ~$60 CPU upgrade) and you have room for 2 cards at full pcie 3.0 x16. Removing or minimizing bottlenecks on newer cards can give big gains!"

Some more info regarding comparisons of CPU power and platform:

Settling into the #6 spot.

We still have a slight downhill trend in our production but as of late, we have picked up production considerably. With summer fighting to stay active here, Fall temps will soon be here. I know that my production will increase as will that of several other teammates in addition to members of...

www.overclockers.com

Mainly wanted to drop some more info regarding questioning performance of 4070 Super vs 4070 Ti in this post (https://www.overclockers.com/forums/threads/settling-into-the-6-spot.805575/post-8213603) by @EarthDog : .

On a high PPD WU, here's a comparison between the 4070Ti and 4070 Super. Both are running PCIe 4.0 x8 in the same system. This is with the Super in the top slot and Ti in bottom, from what we've observed, Top slot almost always performs better than bottom slot even with identical cards (even when running in a 4u so temp differences aren't as drastic as in a tower).



This is about a 21% performance increase over the Super.

Also regarding 4070Ti in 5900x vs 2700x system, The 5900x Ti is on PCIe 4.0 x8 vs the Ti being single card in the 2700x system on PCIe 3.0 x16 so differences are mainly Ubuntu and CPU (Ram slightly, but both are DDR4, 2700x is 3000, and 5900x is 3600). I will have to dig for some more direct comparison results to quantify Ubuntu gains vs CPU.

I'll try to post more data in here as I gather it / post it.

 

[the garynator]

Threw together a comprehensive list of platform average comparison of all of the combinations we have running in WI for Project 18223 over the last few weeks. There wasn't much deviation in ppd across the WUs for the given time, so it should give a pretty accurate comparison. Samples column is how many WUs are included in the average.



18223 is a pretty average example. There are some WUs that show larger differences and some that show smaller differences.

As you can see, if you're running a card in a bottlenecked system, you could be losing a significant amount of points. It also illustrates how big of an increase Linux offers. CPU, PCIe, RAM speed aren't huge gains individually, but they all add up. If you're running a newer card in an old, slow system on windows, you could potentially see ~50% gains moving to a faster system and ubuntu.

I'd imagine you'd see similar results with 3080s and faster cards as well but I don't have any data to compare or a card to get it from. Just might not be as big of gains. There's a certain point at which it doesn't make a huge difference as well. It will, however be even more pronounced on faster cards and I would imagine it will be on the new 50x0 cards coming out soon.

I know it's only one project, but it was a bit time consuming to do this comparison manually so I'll probably just do the rest when I have time to set it all up to do the averages / grouping automatically.

EDIT: Also, top slot typically has the highest PPD even in biforcated or full speed PCIe's, this is evidenced by the 3900x system having higher average than the 5900x system. The 3900x has a Zotac Ti in the top slot, and an Asus 4070 Super in the bottom slot (riser cable and mounted vertically). The 5900x system is a 4u Rackmount case with a PNY 4070 Super in top slot, and Zotac 4070 Ti in bottom slot (temps are much more even in this machine due to the cards being vertical). The Dell 7920 is about 400-500k higher PPD on the slots that would be comparable to the top slots per CPU. All of the cards in the 7920 are on risers and mounted vertically.

 

[dfonda]

Thanks for the hard work. Monkeys rule Humans drool.

 

[the garynator]

Just a little tidbit. 1.2mppd difference between my 2700x and 5900x systems. Difference is 2700x is overclocked to 4ghz, Ti is running on PCIe 3.0 x16, Ti is overclocked, Windows 10 and single card folding. 5900x is stock, Ti is in bottom slot on PCIe 4.0 x8, Ti is running at stock, Ubuntu, and dual cards. 1.2mppd difference on same WU. 3sec diff in frame time average. It's quite crazy how big of a difference even a second or three makes in bonus points. If your CPU is bottlenecking checkpoints, it can easily add a second or 2 to your TPF.


On another note, something I've seen in one of our systems was that uploading completed WUs was taking a long time for some reason (I believe it was due to my VPN). sometimes up to a minute or 2 to upload. Keep in mind that HFM calculates PPD and does not grab the credit that F@H returns once the WU has been uploaded (this is going to be an issue with V8 since they removed that functionality). The only way to see this is either watching advanced control as it's uploaded or checking the logs. That could mean a million or 2 ppd discrepancy between HFM / Advanced Control / Lars vs what you're actually getting.

 

[the garynator]

Instabilities in my 5900x system have returned since putting the Ti back in it. So to rule out the Ti as the issue, I moved it to the top slot and put the super in my 2500k system (now running Ubuntu) and have an interesting data point, all the same Project. Ti in 2700x is on Windows 10 and overclocked. 5900x Ti is stock and in top slot, only card, 2500k @ 4ghz with Super running stock.


Keep in mind, this is a WU that only does checkpoints every 5% instead of every 2%, so CPU won't factor in nearly as much as on most of the other high PPD WUs.

 

[the garynator]

Here's some direct comparisons between the 4070 Super at stock speeds in the 5900x system vs the 2500k system on relatively high PPD WUs from Project 18228. Almost a 2m ppd difference in Linux. I've also included some other data sets for comparison as well.

You can mostly extrapolate PCIe bandwidth differences, CPU differences, and Ubuntu vs Windows differences from this data. It's a small dataset but it's pretty much on par with everything I've seen for similar WUs.



Here's with the 4070Ti data on these WUs I had handy mixed in:


I scraped some of the data to remove outliers from when I was using my pc / reboots / instabilities / etc.

I also started setting up a pivot table so I should be able to assemble results for other projects and such in quite a bit less time going forward.

Anyway, I'm kinda impressed with how well the 4070 super is doing on 18228s in the 2500k system. I think I'm going to rip through the results from 18224s next (we call them wonder units lol), they are the best performing WUs and we've seen some of the biggest discrepancies between platforms on them. The checkpoints tend to hit CPU pretty hard.