Any recommendations for inexpensive secondhand GPUs?

[MeeLee]

Not sure how the PCIE 2.0 x1 will work out to be.
Most motherboards and FAH software is kind of requiring PCIE 3.0 x4 slots...
Will be interesting to see what GPU gets you what score on that x1 slot.

The GPU depends on the PCIE slot to provide power.
At x1, it'll only get 20W. At x4 it gets 25W. At x8 it gets 35W, and at x16 it gets up to 75W.
Which is why you'll preferably need to run a powered riser on anything but x16 slots.
When you connect a GPU to an open slotted PCIE x4 slot, or a PCIE x16 slot, with only x4 lanes, your GPU won't be able to reach the maximum wattage target, and might run at 80W (or lower) instead of 100W, even if the top 6 pin can provide a full 75Watts of the 100Watts the GPU needs.

[ZePompom]

Not sure how the PCIE 2.0 x1 will work out to be.
Most motherboards and FAH software is kind of requiring PCIE 3.0 x4 slots...
Will be interesting to see what GPU gets you what score on that x1 slot.

The GPU depends on the PCIE slot to provide power.
At x1, it'll only get 20W. At x4 it gets 25W. At x8 it gets 35W, and at x16 it gets up to 75W.
Which is why you'll preferably need to run a powered riser on anything but x16 slots.
When you connect a GPU to an open slotted PCIE x4 slot, or a PCIE x16 slot, with only x4 lanes, your GPU won't be able to reach the maximum wattage target, and might run at 80W (or lower) instead of 100W, even if the top 6 pin can provide a full 75Watts of the 100Watts the GPU needs.

Ok, thanks.

In my case, it'll be a 4x riser plugged in a 16x slot. So 25W, plus the 75W of the additional 6 pins cables = 100W, so should be enough since the TDP of the P106-90 seem to be around 75W (Plus on my GTX 1660 I noticed FAH doesn't make it reach their max consuption, but more like around 75% of it).

If not enough power, well, I can try with a 16x riser, less flexible, but more power supply.


For the 1x through USB, indeed power supply is not an issue, because it's a powered riser. I am (but not as much than I was) worried about the data transfert, but I got some GPU-Z screenshots on LTT forums and it's encouraging.

These ones :





His P106-90 doesn't max-out a 1x 1.1 PCIe (uses "only" 82% of it) and make about 160k PPD on average (can go up to 200k, he said he's probably bottlenecked by his old CPU struggling to manage 3 GPUs folding simulataneously, when he stops one of them, the PPD f the 2 other rises. So his P106-90 is around 160k when all 3 GPUs are folding, around 200k when only 2 are, still on its PCIe 1x 1.1. The screens above are in the 160k PPD scenario, and we can indeed see the GPU not being loaded at 100% but 88%).

So with 1x in 2.0 I should be fine #fingers_crossed.

Someone else made the statment, about Nvidia P106 : "I do not have a -090 to play with, but I can say that my -100 is doing around 450-500kPPD, even when plugged into a PCIe 1.1 x16 slot."

[MeeLee]

82% is definitely bottlenecked. The measurments are very rough, usually jump every ~20% or so. It means on average you're using 82% of the bandwidth, but probably have many spikes at 100% that are slowed down.You'll be able to notice this when you plug the GPU in an x16 slot, and see how high the PPD increases.
Should be around 330-350kPPD in Linux (300-330k in Windows on Core 21).
PCIE bandwidth is best kept under 75%.
Your GPU also shows 64W of usage, an indicator you either capped power too far, or your riser isn't getting the power it needs (like if you use SATA to PCIE 6 pin adapters to power the riser).
You could also increase fan speed, which brings the temps down. I run mine at 100%, and personally prefer to run them at 45-65C.

there are 2 versions of the P106. The ones made after the GTX 1060 3GB (300-360k PPD), and the 6GB (500-650k PPD).

[ZePompom]

The P106-90 is based on the GPU of the GTX 1060 but have less CUDA cores. On this aspect it matches the GTX 1050 Ti

768 cores for 1050Ti, 1152 for 1060 3GB, 1280 for 1060 6GB.
P106-90 gets 768 and the P106-100 gets 1280.




So to me it's unrealistic to expect the P106-90 to get as much PPDs than a 1060 3GB, since the specs look more like a slightly overclocked 1050 Ti with faster but smaller RAM.
The P106-90 TDP is also much closer to the 1050Ti one than the 1060 3GB one.

In this context. Yes 160k PPD is still bottlenecked, as we can see on the screenshots and as the owner admits. But not that much. Plus I don't think it's by the pci bus since the owner says it jumps to 200k when one of the other gpu of his system stops folding (wich is approx what we can expect for a full unleashed 1050 Ti).

Regarding spikes, I found the pci bus usage very stable when monitoring my GTX 1660 during folding but indeed, spikes are way less visible on a 16x 3.0 slot than on a 1x 1.1
But since I should get a 1x 2.0, so twice the PCIe bandwidth than what we can see in the screenshots, it should be fine ... I guess

I should get my first riser next week, the "normal profile" 1x to 16x through USB, the only one shipped from my country, the others are shipped from Asia so I'll have to wait much more
This one is meant to be used in combination with the low profile one, for my hidden 1x connector, but once I get it I'll still be able to do some testing by plugging it in my free 16x slot (the one who should host the 4x to 4x riser in the end)


The final configuration should look like this is things goe smoothly :



So if 1x to 16x riser is not fast enough, and 4x to 4x doesn't bring enough power, I'm screwed

[foldy]

One option to get better PPD on x1 risers is to use Linux instead of Windows.

[ZePompom]

One option to get better PPD on x1 risers is to use Linux instead of Windows.

That's what I planned to do initially (64 bits Lubuntu), with a dedicated folding machine, but since the PC I had for this project is too old, now it's plan B : putting the P106-90 I got in my daily life PC.

But since I do gaming on this PC, I keep it under win 10, for convenience


But I'll use linux if plan B fails and I go for plan C : Back to dedicated folding machine but with parts I would buy specifically for this purpose (since I can't use what I already have).

It would cost around 170€, Ryzen 1200AF or Athlon 3000G, Motherboard with two PCIe 16x slots, 4GB RAM (PSU, HDD and PC case would be reused from the old PC I can't use the P106-90 into because MB is too old)


I slowly increase the budget but the plan C have a huge gap

Plan A : Just put P106-90 in my old PC
90€ for two P106-90
=> Failed because the old MB doesn't support the much more modern Video Cards

Plan B : Putting the two P106-90 in my main/daily life PC
+20€ for PCI risers and power supply cables adapters
=> in progress

Plan C :Buying brand new parts for a folding machine able to host the two P106-90
+170€ for brand new Motherboard, CPU and RAM


That's why I'm ok if the P106-90 are slightly bottlenecked by plan B, because since plan C is much more expensive, I don'twant to do it only for few thousands of PPDs
It's more in case they are hugely bottlenecked by plan B, or it simply doesn't work. (And even if that case, not even sure I'll do it, there is fair chances I just give up rather than spend 170€ by still trying to make it work)

[MeeLee]

The bottom port is actually a PCIE x16 slot, not an x4 slot.
But it appears it's only populated with x4 pins. (which means it won't get enough power to the GPU, even though the board has a 24 pin PSU connector, which means it was lazy of Gigabyte to exclude the remaining pins there)...
Still, if you could find a powered riser (x4 or x16), your GPU will benefit from it.

Instead of using the x1 slot, check to see if the board has an M.2 slot on the back or front (underneath some heat sink) somewhere.

There are some m.2 to PCIE x4 adapters available on the internet.
Some m.2 slots operate at x2 speed only, but even they are an improvement over the x1 slots.

[ZePompom]

The bottom port is actually a PCIE x16 slot, not an x4 slot.
But it appears it's only populated with x4 pins. (which means it won't get enough power to the GPU, even though the board has a 24 pin PSU connector, which means it was lazy of Gigabyte to exclude the remaining pins there)...
Still, if you could find a powered riser (x4 or x16), your GPU will benefit from it.

Instead of using the x1 slot, check to see if the board has an M.2 slot on the back or front (underneath some heat sink) somewhere.

There are some m.2 to PCIE x4 adapters available on the internet.
Some m.2 slots operate at x2 speed only, but even they are an improvement over the x1 slots.

Yup, it's populated in x4 only, that's why I ordered only a 4x riser because 8x or 16x wouldn't be useful.

There is indeed a M.2 connector, just above the top PCIe 16x slot, I'll try this if the 1x bottleneck too hard.
If the GPU can't have enough power through the fake 16x slot (even if the total should be enough with PCI + 6 pins cable), yes, powered riser it'll be.
But the powered risers I can find are all in 1x to 16x, so if 1x is not fast enough it won't solve my problem
( https://www.amazon.fr/KALEA-INFORMATIQU ... 61&sr=8-14 )

I'll already see if it works that way, even bottlenecked. If it does but bottleneck too hard I'll explore these options.
If it does not work, with both kind of risers, I think I'll just stop it there

But that's exactly what I'm affraid about. Not enough speed from 1x riser and not enough power from the 4x


EDIT : found a powered 16x to 16x, yay!
https://www.ebay.com/itm/Powered-PCI-Ex ... 3040624000

[ZePompom]

Something came through my mind.

Even if wired only in 4x, the second 16x slot is still made for a video card, in a crossfire or SLI config.

So could it have the same power output than a regular 16x slot to be able to power the said video card plugged in it?
Since you're supposed to be able to plug in it the same videocard than the one in the fullspeed slot.


That could explain why Gigabyte didn't remove the extra pins for power supply.

So I believe this "fake" 16x slot could actually deliver the same power output (75W) than a real one.

What do you think?

[MeeLee]

No. It's clearly only populated at x4.
Which means it'll only get 25 Watts from the motherboard, instead of 75W.
In SLI it could get some power from the top connector, if your GPU supports it.
But FAH doesn't use SLI.

The GPU will still work in the x4 slot, but there's a possibility your GPU can get anywhere from a maximum of 50W to a max of 100W in that slot. Chances are closer to 60-75W.
The GPU won't be able to run at it's 100% power ratio. But it'll still be overclockable. It'll also run a lot cooler.
With a powered riser, you can actually raise the GPU power and overclock, thus also increase PPD over stock in an x16 slot.

Chances are without the powered riser, you can run your GPU overclocked, at a lower wattage; at about the same performance as the GPU stock in an x16 slot.

[cayenne187]

Radeon RX 570-$125 new-plenty on auction at ebay.

[ZePompom]

No. It's clearly only populated at x4.
Which means it'll only get 25 Watts from the motherboard, instead of 75W.
In SLI it could get some power from the top connector, if your GPU supports it.
But FAH doesn't use SLI.

The GPU will still work in the x4 slot, but there's a possibility your GPU can get anywhere from a maximum of 50W to a max of 100W in that slot. Chances are closer to 60-75W.
The GPU won't be able to run at it's 100% power ratio. But it'll still be overclockable. It'll also run a lot cooler.
With a powered riser, you can actually raise the GPU power and overclock, thus also increase PPD over stock in an x16 slot.

Chances are without the powered riser, you can run your GPU overclocked, at a lower wattage; at about the same performance as the GPU stock in an x16 slot.

Weird, since the power is only supplied by the "header" of the PCIe slot, regardless of its speed, what could prevent it to give 75W even if only populated in 4x?
What I read about PCI norm is 4x cards must drain maximum 25W from the motherboard and 16x cards can go up to 75W. But this slot, even if populated in 4x, is clearly meant for 16x cards, since it's a 16x slot. So it would make sense than even if there is less lanes/speed, it could feed cards made for full 16x slots, no?
I mean. Does it even respect the PCIe norm to have a 16x slot (regardless of how it's populated) who can not provide the power the PCI norm says a 16x card should be able to get ?
It's still a 16x slot, meant to host 16x cards (if not it would be just a 4x slot), but it would be unable to power most of them?


I know FAH doesn't use SLI, I just find kinda weird than a Motherboard who is supposed to be crossefire-friendly would not provide the same amount of power to both PCIe slots, while they are clearly intended to both host the same Video Card model
It would be more logical, to me, than these two slots can provide the same amount of power, and that would explain the additional pins on MB (even if it can still be lazyness from Gigabyte as you suggested, though)


But indeed, these Cards have an additionnal 6 pins conenctor, meaning +75W, and their TDP is 75W, so theorically, they could be powered by the sole 6 pins connector (if not using sata adapter), so 25W from the PCI slot should be enough to go full power, if the bandwidth is adequate.


I don't plan to overclock or undervolt the P106-90 as I had stability issues with FAH (and only with FAH) when I slightly overclocked the GTX 1660.

Other softwares were stable, but on regular basis FAH core was crashing. Explicitely asking me if my system was overclocked
So I'll maybe do some tests, but I didn't plan to modify them. i'll be glad already if it works okay-ish with vanilla settings, even if some PPDs are lost

Can't wait to test it out. I'm tired to worry about it, not knowing if it'll work or not

[MeeLee]

That assumption would be incorrect.
each increase in lanes, adds both data and power lanes to the GPU.

The problem with the 6-pin plug, is that in theory they're rated for up to 75W, but the GPU is engineered to only use a specific amount of that power (usually less than that); as it's designed for being entered in a PCIE x16 or x8 slot, and get a certain percentage of it's power from that slot.
If you're power capping your GPU, this would not be an issue, so long the GPU still gets enough watts.

[ZePompom]

That assumption would be incorrect.
each increase in lanes, adds both data and power lanes to the GPU.

Are you sure? When I check the PCIe pinout I don't see additional power pins in a 16x compared to a 4x. In both case all power pins seem to be in the header, who is common to all speeds.

https://pinouts.ru/Slots/pci_express_pinout.shtml

The extra pins are data and ground, but I don't see any additional power.

So I don't see why a motherboard manufacturer couldn't desing a 4x populated slot with the same power than a full 16x since the power pins are the same in both case ?

Plus, I insist, but it would make no sense to have a slot designed for 16x GPU rated at 25W max, meaning you would be unable to use in it almost every modern gpu.

If it's the case, I would call itba scam. But I doubt that what Gigabyte did.

It all make sense to me. The motherboard rated for crossfire (aka the same gpu in both slots), the 4x pci havi'g as much power pi's tha' a 16x, the aditional power pi's to motherboard (24 instead of 20)

Sadly thee is not enough room to plus my gtx 1660 there. If it was possible I would plug it in, to make sure it can or can't provide the power.

To me, it's just a slower pcie 16x slot, but still meant to host 16x cards, and to feed them properly ^^

You were right on the "this old MB will never work with that kind of gpu", so I hope you're wrong this time, rofl

[MeeLee]

Well, test it out, and report.
I'm telling you, your PCIE x4 seated GPUs will more than likely run at a lower TDP than in an x16 slot, because mine did.
Most of the X8-x16 slots, is data and ground. some of the data connections carry a 12V signal, since it's a digital line.