• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

PCIE5 on PCIE3

T

Toconator

[H]ard DCOTM January 2026
Joined
Jul 8, 2005
Messages
878
Just an observation. My new RTX5060Ti PCIE5x8 on a PCIE3x16 slot is almost always @ 100% PCIE bus load according to GPU-Z. PPD is also down vs the PCIE4x16 mobo I had it in at first but with variability of FAH projects its hard to say how much but at least 300K ppd, probably a bit more. It's current home is on 24/7 tho so I'll take the small sacrifice of ppd. Seems we're at the time where PCIE version is starting to matter.
 
It's not a new thing with regards to F@H. I noticed a difference in output years ago when I swapped a bunch of mining systems to F@H for a competition. Same exact GPU's on the same system produced more output if they were connected to an x16 or x8 versus the x1 connectors I used for mining.

I went so far as to buy a bunch of pciex16 extenders and swapped some cards off x1 and x4 systems so instead of going with 1 system and 4-6 GPU's I'd have 1 system and 2. Some only 1 gpu with no extender.

I concluded F@H is sensitive to PCIe specs. So, a 3.0 versus a 4.0 makes sense to me. Limiting bandwidth is a measurable differential. This is not the case for BOINC projects. I only noticed it in F@H. I think others have written about it.

Since I'm power limiting my GPU's I decided the bump in performance isn't worth the hassle and extra expense.

I've not tested my x8 5060's for this particular scenario. But, I also had a MB completely fail to utilize any of the x8 variants I've picked up. To the extent I RMA'd the MB and bought a new test bench system.
 
It is on a gen3x16 slot and totally saturates it most of the time. I'm not willing to move it just yet but did grab a used 5600X CPU and got a new B550 mobo on sale for a good deal. Gotta swap some parts around but should have a gen4x16 slot available soon.
 
but they are electrically limited to x8 so it would be at 3x8, thats whats gimping it and why it "saturated".
 
Last edited:
I was fine being blissfully ignorant.

Swapped a 4060 out for a 5060 which is on an older board with a gen3x16 slot. My test bench is apart right now so I didn't try it on a newer board first.

Conveniently it's in the basement rack where I don't see it very often. So I'll try and forget about it before I find myself at Microcenter again.

Considered keeping both GPUs. The bottom card would get pretty choked out though. Could use an SFX PSU with an ATX adapter to free up some space for the bottom card. Or put waterblocks on them like all my other rigs... Every project always starts snowballing. I have other things I need to spend money on.
1762789454494.png
 
Fluxmaven said:
I was fine being blissfully ignorant.

Swapped a 4060 out for a 5060 which is on an older board with a gen3x16 slot. My test bench is apart right now so I didn't try it on a newer board first.

Conveniently it's in the basement rack where I don't see it very often. So I'll try and forget about it before I find myself at Microcenter again.

Considered keeping both GPUs. The bottom card would get pretty choked out though. Could use an SFX PSU with an ATX adapter to free up some space for the bottom card. Or put waterblocks on them like all my other rigs... Every project always starts snowballing. I have other things I need to spend money on.
View attachment 765871
Click to expand...
Snowballing is most of our middle names lol 😂 I know what you mean 😁
 
xDiVolatilX said:
Snowballing is most of our middle names lol 😂 I know what you mean 😁
Click to expand...
Yeah my last snowball a few days ago netted me a new B550 mobo and used CPU mostly for the PCIE4 slot. Got an unactivated Win11 Pro on it which wouldn't run FAH, stalling while updating core so I injected the cores folder from another boxen and it's off to the races so far. Might 'snowball' another NV 5xxx series for folding next ...
 
Last edited:
So, the snowballing happened.

I wanted to upgrade this rig as cheaply and simply as possible. I picked up an open box Ryzen 4500 for $35... Which had 8 slightly bent pins that I thankfully was able to straighten out before realizing that CPU only supports PCIe Gen3 as well. So that's on the pile of stuff to bench for HWBOT now.

With RAM prices going to the moon, I ruled out using any of my spare AM5 CPUs since I don't have any extra DDR5 kits. So, it is receiving a new ASUS B760M-AYW WIFI D4 II because it was the cheapest board I could find that gets me a Gen 5x16 slot and can use it with DDR4 that I already have. That's being paired with a 12490F.

Stealing the 12490F meant I was short a CPU for a different system... So I bought a Core Ultra 245T and a B860 Tomahawk to upgrade my Plex server. The 14600K that was going to go in the Plex server is going in the machine that was going to use the 12490F. This also leaves me with a spare Z790 Tomahawk. Couldn't use it in the folding machine because the 2U rack case only supports up to mATX. Also it's a DDR5 board.

That's not the end of the story though... While keeping an eye on sales acquiring all of this hardware, I came across an open box Supermicro workstation at Microcenter. Ryzen Pro 8700G/32GB DDR5/2TB NVMe/RTX 4070 for $660... Couldn't pass it up. So now I have another 4070 to fold on. Still haven't finished building the system with the last open box 4070 I bought to fold on lol.

So thank you for bringing the whole Gen3 VS. Gen5 issue to light. :D
 
Fluxmaven said:
So, the snowballing happened.

I wanted to upgrade this rig as cheaply and simply as possible. I picked up an open box Ryzen 4500 for $35... Which had 8 slightly bent pins that I thankfully was able to straighten out before realizing that CPU only supports PCIe Gen3 as well. So that's on the pile of stuff to bench for HWBOT now.

With RAM prices going to the moon, I ruled out using any of my spare AM5 CPUs since I don't have any extra DDR5 kits. So, it is receiving a new ASUS B760M-AYW WIFI D4 II because it was the cheapest board I could find that gets me a Gen 5x16 slot and can use it with DDR4 that I already have. That's being paired with a 12490F.

Stealing the 12490F meant I was short a CPU for a different system... So I bought a Core Ultra 245T and a B860 Tomahawk to upgrade my Plex server. The 14600K that was going to go in the Plex server is going in the machine that was going to use the 12490F. This also leaves me with a spare Z790 Tomahawk. Couldn't use it in the folding machine because the 2U rack case only supports up to mATX. Also it's a DDR5 board.

That's not the end of the story though... While keeping an eye on sales acquiring all of this hardware, I came across an open box Supermicro workstation at Microcenter. Ryzen Pro 8700G/32GB DDR5/2TB NVMe/RTX 4070 for $660... Couldn't pass it up. So now I have another 4070 to fold on. Still haven't finished building the system with the last open box 4070 I bought to fold on lol.

So thank you for bringing the whole Gen3 VS. Gen5 issue to light. :D
Click to expand...

It has been my pleasure sir. MOAR PPD for the [H]orde!
 
Update: swapped the 5060Ti into a PCIE 4.0 mobo. PCIE bus load still reaches 100% but spends less time there. PPD might be slightly better but it's hard to say until it runs a few more WU's. Another observation: I'm not sure if this is inherent for 5xxx series cards or it's a factor of not being on PCIE 5 but the GPU load seems to fluctuate far more than my other cards. All cards dip periodically very briefly (maybe as a data point is done?) but this one has the load all over the map. For comparison my 2060S hums along at a constant 93-99% GPU load with the occasional dip. I'm using MSI Afterburner and GPU-Z sensors to monitor performance. PerfCap Reason is voltage reliability so I've unlocked voltage control (which I don't usually do) and forced constant voltage and even trial boosted +10% to no avail. I'm wondering if I got a flaky card.

edit: 2060S just started the same WU as the 5060Ti. I'm seeing more variability on the 2060S now but not as drastic, maybe because it's not as powerful and isn't clocked as high. It's one of these 182xx WU's some of which really push the GPU's. I'll try to performance track an easier WU if/when the 5060TI gets one. I'll set the voltage back to default in the meantime.
 
Last edited:
I can't speak to other 50x0 cards, but the 5090 I had running on PCIE 4.0 had 25% lower PPD than the one running on PCIE 5.0. After swapping to B650-E + 9800X3D, PPD increased 25% immediately. I was too impatient to test the 5070, but I'd assume there might be similar bus bandwidth limitations when working on larger WUs.

Same SSD/OS installation/driver. Straight motherboard / CPU swap.

Edit: To clarify, the frame time on PCIE 4.0 and 5.0 were within 0-2 seconds, the major difference was the checkpoints. Writing the checkpoint would take 5-10 seconds longer. That can delay can be very significant when PPD is calculated.
 
Last edited:
russnuck said:
I can't speak to other 50x0 cards, but the 5090 I had running on PCIE 4.0 had 25% lower PPD than the one running on PCIE 5.0. After swapping to B650-E + 9800X3D, PPD increased 25% immediately. I was too impatient to test the 5070, but I'd assume there might be similar bus bandwidth limitations when working on larger WUs.

Same SSD/OS installation/driver. Straight motherboard / CPU swap.

Edit: To clarify, the frame time on PCIE 4.0 and 5.0 were within 0-2 seconds, the major difference was the checkpoints. Writing the checkpoint would take 5-10 seconds longer. That can delay can be very significant when PPD is calculated.
Click to expand...
Yes and the fact that the 5060Ti is only an x8 card magnifies the problem. Watching the GPU load graph on Afterburner I noticed a few of the drops flatlining for a few seconds, not the usual drop-point. I have the checkpoints maxed to 30 min so that is definitely it. I was wondering if a x16 card would suffer as much but I'm not willing to shell out for a 5070 to find out. On the flip side I wonder if a weaker card, say a 5050 would be hampered. I might be willing to try that. Worst case it does suffer some but still get >PPD than my older Turing cards for less power draw.
 
Got the low profile 5060 back up and running on the newer motherboard. So I'll give it a couple days to get some data to compare. It's running linux so no GPU-Z or Afterburner unfortunately.

This is the last couple weeks with it on a Gen3 slot. Hopefully see a bit of a jump with it on Gen5 now.

5060Gen3PPD.png


Even if it doesn't pick up a ton of PPD, I just want it to be stable. After dropping in the newer GPU, that system had several instances where it dropped the network connection and didn't recover on it's own. Which involves a trip to the basement that I can't make when I'm out of town for work. It had been stable for years prior so I guess the constant 100% load on the PCI bus was making it unhappy.
 
Looking over what hardware I still have and it appears that I'm running only on PCIe 3.0 or older. So, for me to upgrade GPU's I would basically need a upgrade to an entirely newer system.
 
Gilthanis said:
Looking over what hardware I still have and it appears that I'm running only on PCIe 3.0 or older. So, for me to upgrade GPU's I would basically need a upgrade to an entirely newer system.
Click to expand...
as long as you dont jam a 4090/5090 in there or one of the neutered x8 cards you'll only loose a bit of fps, it will still be an upgrade. make sure you have rebar support though...
 
pendragon1 said:
as long as you dont jam a 4090/5090 in there or one of the neutered x8 cards you'll only loose a bit of fps, it will still be an upgrade. make sure you have rebar support though...
Click to expand...
My rigs are old anyways. But I was debating whether to get a 5060 ti in the near future or not. It would be a theoretical performance boost of 30-50% over my 2080S in various tests. I'm still using X99 and X399 boards. My Plex server is still an Ivy Bridge era chip IIRC.
 
Gilthanis said:
My rigs are old anyways. But I was debating whether to get a 5060 ti in the near future or not. It would be a theoretical performance boost of 30-50% over my 2080S in various tests. I'm still using X99 and X399 boards. My Plex server is still an Ivy Bridge era chip IIRC.
Click to expand...
i didnt see that this was in the DC section, whoops... the 5x8 5060ti will be limited to 3x8, which will gimp it.
 
Average PPD on my low profile 5060 jumped up to 6mil after moving to a PCIe gen5 slot.

Lifetime average on my 5060 Ti is 7.4mil but it's running Windows and significantly undervolted.

A used 4070 might be a better option. Significantly more PPD while still drawing a tad less power than a 2080S. They are also gen4x16 so they won't be gimped on your older platforms.
 
Gilthanis said:
My rigs are old anyways. But I was debating whether to get a 5060 ti in the near future or not. It would be a theoretical performance boost of 30-50% over my 2080S in various tests. I'm still using X99 and X399 boards. My Plex server is still an Ivy Bridge era chip IIRC.
Click to expand...
pendragon1 said:
as long as you dont jam a 4090/5090 in there or one of the neutered x8 cards you'll only loose a bit of fps, it will still be an upgrade. make sure you have rebar support though...
Click to expand...
I have a 4090 currently running on an old X470 motherboard with PCIe 3.0 16x.
Using the original 2700X, of which the 4090 was getting around 24-26 million PPD, was what was holding it back severely.

Moving to a 5800XT allowed the 4090 to now get around 32-34 million PPD.
I really think the biggest limiter with FAH 8.4.X is the CPU during the decompression/compression and checkpoints, much more so than the PCIe bus; Nicolas_orleans has brought this in other threads and I tend to agree with him on this from my experiences.

A proper test of the PCIe bus being the limiter would be to use a PCIe 5.0 platform with all 16 lanes and a fast CPU.
Then manually drop the PCIe bus down from 5.0 to 4.0, 3.0, etc. in the BIOS/UEFI and see if there really is much of a difference, compared to dropping to a much weaker CPU on that same platform.

I also have a 3070 running on an old AMD Jaguar thin client as an eGPU with PCIe 2.0 4x.
With FAH client 7.X.X it would give around 5 million PPD, but with FAH client 8.4.X it now gets around 3.6 million PPD.

The PCI 2.0 4x bus should be extremely limiting (uses around 75% of the bandwidth, easy to monitor in Linux), but that same GPU on a much more powerful platform and PCIe 3.0 16x, still only nets around 6 million PPD.
This again backs up the CPU being far more important than the PCIe bus for FAH GPU tasks.
 
Last edited:
Well said, on pcie >= 3.0 I would not blame the behavior of the v8 client (I have seen no report v7 behaves differently with the same core), but the one of the newer cores + the fixed checkpoint frequency that is the same for a 1080 and a 5090...

Interesting to see the pcie 2.0 behavior you experienced !

Also, I think in newer cores there is an optional parameter for researchers to do some deep sanity checks on input and/or output. It's activated on some projects but not all and is reportedly CPU intensive.
 
Last edited:
I noticed a performance gain with a 3080 Ti going from gen 3 to gen 4 a few years ago when I switched from mining to folding, plus the 3060 Ti cards were gimped on x1 cards used for mining.
 
Red Falcon said:
I have a 4090 currently running on an old X470 motherboard with PCIe 3.0 16x.
Using the original 2700X, of which the 4090 was getting around 24-26 million PPD, was what was holding it back severely.

Moving to a 5800XT allowed the 4090 to now get around 32-34 million PPD.
I really think the biggest limiter with FAH 8.4.X is the CPU during the decompression/compression and checkpoints, much more so than the PCIe bus; Nicolas_orleans has brought this in other threads and I tend to agree with him on this from my experiences.

A proper test of the PCIe bus being the limiter would be to use a PCIe 5.0 platform with all 16 lanes and a fast CPU.
Then manually drop the PCIe bus down from 5.0 to 4.0, 3.0, etc. in the BIOS/UEFI and see if there really is much of a difference, compared to dropping to a much weaker CPU on that same platform.

I also have a 3070 running on an old AMD Jaguar thin client as an eGPU with PCIe 2.0 4x.
With FAH client 7.X.X it would give around 5 million PPD, but with FAH client 8.4.X it now gets around 3.6 million PPD.

The PCI 2.0 4x bus should be extremely limiting (uses around 75% of the bandwidth, easy to monitor in Linux), but that same GPU on a much more powerful platform and PCIe 3.0 16x, still only nets around 6 million PPD.
This again backs up the CPU being far more important than the PCIe bus for FAH GPU tasks.
Click to expand...

Interesting. I'm still running FAH 7xx and therefore slightly older GeForce drivers but if I'm forced to upgrade my main PC's drivers for a game fix I'll have to switch to FAH 8xx. It's on my best CPU atm tho 5700X with tweaked RAM so it should suffice. Nice to know going forward. Thx RF
edit: I did see a PPD increase when I switched from PCIE 3 to 4 on the 5060Ti even tho its a bit gimped by the x8. Up to ~1m PPD in best case. This also going from 8600K to 5600X CPU (FAH 7xx)
 
Last edited:
I just ran into this myself.

Using a hand-me-down Intel 10700K on a B460M. Or rather.... A 10700K on a Z490, no post no nothing, but it had a resistor on the board that was almost hot enough to boil water and the chipset got warm. Previous owner stated that it died in a power surge. He bought replacement 10700K and no luck (well yeah, your MB is fried). Gave the parts to his nephew, and after no luck, he bought a new MB (the B460M) which should have fixed the issues, but...... didn't. I got the parts in a mixed batch at the beginning of the month and did some testing. Both CPUs good, RAM Good, new MB BENT PINS in the CPU socket. Final diagnosis: no dual-channel RAM and PCIe 1 is now 3x8 instead of 3x16.

3x16 to 3x8 brought my 4070TiSup down a few hundred thousand PPD but some of that might be no dual channel RAM.
I also noted that non-XMP vs XMP ram was MILLIONS of PPD.

I went back to my rig with a AMD 9950X3D + 5080.... weird, XMP was on before my processor upgrade... turned it back on and that 5080 gained +3,000,000 PPD.

So TL/DR, yes PCIe channels make a significant difference, but RAM bandwidth can choke a GPU.

Now I have 2 free 10700Ks, a free tower, a crap MB for free, and a free RTX3070. All it cost me was the time to take 2 towers, 3 Mainboards, 2 GPUs and build one tower that actually worked. I took a gamble and grabbed a B560M of newegg so I can see what I can do with my spare parts.
 
Last edited:
jfb9301 said:
I just ran into this myself.

Using a hand-me-down Intel 10700K on a B460M. Or rather.... A 10700K on a Z490, no post no nothing, but it had a resistor on the board that was almost hot enough to boil water and the chipset got warm. Previous owner stated that it died in a power surge. He bought replacement 10700K and no luck (well yeah, your MB is fried). Gave the parts to his nephew, and after no luck, he bought a new MB (the B460M) which should have fixed the issues, but...... didn't. I got the parts in a mixed batch at the beginning of the month and did some testing. Both CPUs good, RAM Good, new MB BENT PINS in the CPU socket. Final diagnosis: no dual-channel RAM and PCIe 1 is now 3x8 instead of 3x16.

3x16 to 3x8 brought my 4070TiSup down a few hundred thousand PPD but some of that might be no dual channel RAM.
I also noted that non-XMP vs XMP ram was MILLIONS of PPD.

I went back to my rig with a AMD 9950X3D + 5080.... weird, XMP was on before my processor upgrade... turned it back on and that 5080 gained +3,000,000 PPD.

So TL/DR, yes PCIe channels make a significant difference, but RAM bandwidth can choke a GPU.

Now I have 2 free 10700Ks, a free tower, a crap MB for free, and a free RTX3070. All it cost me was the time to take 2 towers, 3 Mainboards, 2 GPUs and build one tower that actually worked. I took a gamble and grabbed a B550M of newegg so I can see what I can do with my spare parts.
Click to expand...
Interesting. I would've thought the VRAM bandwidth would've been more of a factor as the folding is run on the videocard, not so much the system. I'm still on DDR4 rigs and I recently inherited a 16GB kit of 3200DDR4. Not sure if replacing the DDR4 2400 in the Server is worth the effort as its only running a stock 8600K and 2060 Super for now .
edit: 'snowballed' to a 3060Ti, seeing over 1M+ ppd with current setup. might try the RAM speed upgrade unless I snowball another rig.
 
Oh, I am sure VRAM bandwidth matters. Far more that VRAM size. I am just saying that your RAM bandwidth matters too, much less so for slower cards like that 3060, RAM bandwidth doesn't affect my RTX3070 anywhere near as much as my RTX5080. It also depends on the WU. Some projects are more affected than others.

It makes sense though. If your PCIe data rate has an effect during the WUs run, then RAM bandwidth will also have an effect. The data that is bottlenecking on your PCIe port, has to be going somewhere. Not just to the CPU and back, your RAM has to be involved.

Your sever is actually in a really good spot for a science experiment. Put the kit in and see what you gain. My DDR 4 rigs are capped because 1) my 9700K memory controller refuses to run >2933 (still better than 2133 though), and 2) my 10700K might run 3200 but it's only single channel. Like my crap B460M, the B560M that I am waiting for states that with Gen 10 chips the memory is limited to 2933, but it may still allow manual OC of the the memory. If that works, I might just RAM swap my DDR3600 from my 9700K that won't run over 2933.
 
Although the thread title is PCIE 5 on PCIE 3 I just wanted to add an observation after upgrading the Server to a PCIE 4 GPU on the PCIE 3 mobo. PCIE bus is barely being utilized, both mobo and GPU being x16 makes a world of difference over the x8 constrained 5060Ti. I would guess a x16 5xxx series GPU won't be as constrained by the slower PCIE bus as well but am unwilling to shell out the $$ to find out. Seeing some great deals on used Ampere cards here locally. Might have to revive an older Kaby Lake mobo and 'snowball' another rig for the challenge. I'll see if I can source an old i7 or i5 .
 
Toconator said:
I'll see if I can source an old i7 or i5 .
Click to expand...
I've got a bunch of Skylake CPUs that will work in a Kaby lake motherboard. Assuming you're in the US, PM me and I'll send you one since it's for the cause.
 
Fluxmaven said:
I've got a bunch of Skylake CPUs that will work in a Kaby lake motherboard. Assuming you're in the US, PM me and I'll send you one since it's for the cause.
Click to expand...
Well thx fluxmaven but i'm in the great white north (and its been pretty cold & white lately). Just bid on a B250 miner mobo with 7600K and 4GB of DDR4 for a measly $50. I don't want the mobo and will just relist it but the 7600K will probably do the trick. If it falls thru and you have a 6700K/6700 I'll pay shipping if it's not too much. Make a decent backup Server.
 
Toconator said:
Well thx fluxmaven but i'm in the great white north (and its been pretty cold & white lately). Just bid on a B250 miner mobo with 7600K and 4GB of DDR4 for a measly $50. I don't want the mobo and will just relist it but the 7600K will probably do the trick. If it falls thru and you have a 6700K/6700 I'll pay shipping if it's not too much. Make a decent backup Server.
Click to expand...
Shipping to a friendly neighbor shouldn't be too bad if that 7600K doesn't work out. I definitely have 6700 and 6700T (35W TDP version). Just need to see if I have any unopened ones left. I have a board setup for non-K overclocking and delidded most of the Skylake chips. They are all re-lidded with liquid metal, but I know some people shy away from that. If all fails I for sure have at least one unopened 6700 running in a NAS that I could swap to something else.
 
You must log in or register to reply here.
Back
Top