PCIE slot power (Very Successful Experiment Overclocking Maxwell Quadro)

[damric]

Been a while since this has been discussed. We know the slot is rated 75w. We know some slot powered cards can exceed that from time to time. I'm about to run some OC experiments on some cheap older hardware and I'd like to know what are the "tolerances" assuming a modern clean PSU and a high end motherboard. Example, do you think 100w will damage the traces on the card or the slot? Assume excellent cooling like cold water and lots of fans blowing cold air. I had a few motherboards in the past that had supplemental power connectors for slots but I no longer have them now. Anyone besides wiz experimented on stuff like that?

The next card in the test lineup is a quadro m2000 that looks like it needs a healthy dose of bios and perhaps driver foolery since it is soft locked (like afterburner) to prevent overclocking.

Last month I pushed a W4100 to about +100% overclock (bios tweak) and it's been folding for a month. I'm not sure that it even hits 75w with the 1200MHz OC and 1225mV. It's under water and on a cheap B450 board. Did the same with W2100 a month before that but that card is even weaker power draw. I'd love to push a few of these Maxwell cards.

 

[Deleted member 24505]

Is the slot max 75w or will it supply more than that?

 

[damric]

Is the slot max 75w or will it supply more than that?

It's supposed to be rated for 75w but we have seen that exceeded by a little bit on reviews here before.

 

[MachineLearning]

PCIe can provide 5.5A on the 12v rail (66W) and 3A on the 3.3v rail (9.9W), totaling a hair over 75W combined. This is the minimum spec by which every single motherboard on the market should abide by, and I expect that all or nearly all name-brand boards support at least 75W per PCIe slot.

Many early RX480 reviews will likely mention the PCIe slot wattage threshold at which issues occur (on their system, at least). This is not a review, but GN Steve talks about the PCIe power issues for the RX480 in this video on YouTube. It sounds like the RX480 was drawing around 100-120W through the PCIe slot which is obviously a lot higher than spec.

edit: I personally would feel comfortable running max 1 card only driven off PCIe at 100-110W, unless your mobo has a 6-pin for PCIe power delivery only (EVGA and ASUS often on high end boards).

 

[damric]

PCIe can provide 5.5A on the 12v rail (66W) and 3A on the 3.3v rail (9.9W), totaling a hair over 75W combined. This is the minimum spec by which every single motherboard on the market should abide by, and I expect that all or nearly all name-brand boards support at least 75W per PCIe slot.

Many early RX480 reviews will likely mention the PCIe slot wattage threshold at which issues occur (on their system, at least). This is not a review, but GN Steve talks about the PCIe power issues for the RX480 in this video on YouTube. It sounds like the RX480 was drawing around 100-120W through the PCIe slot which is obviously a lot higher than spec.

edit: I personally would feel comfortable running max 1 card only driven off PCIe at 100-110W, unless your mobo has a 6-pin for PCIe power delivery only (EVGA and ASUS often on high end boards).

Yes this is about what I was guessing myself, that it's a minimum spec that gets exceeded from time to time though it looks more often the radeons.

 

[seth1911]

Its for the ass, its just marketing.

The Board have a multiswitch from 24pin connector to every PCIE Socket, there are borads out with 4x PCIE sockets and every one can draw 75W,
the spec says only 75W per Socket not how many.

 

[LabRat 891]

Since these are PCIe 3.0 at best, might want to invest in some good risers w/ dedicated power (MaxCloudOn), or mod something like a SuperMicro or other good PCB-based x16 riser. You could probably solder onto the GPU's PCB somewhere to supplement +12V and Gnd from a PCI-E power lead instead.
By your own experience, you'll probably be okay. However, my experience differs. Managed to make a Chinese import board go slightly 'melty' in the 24-Pin with modified (Reynold's wrap shielded) cheapy x1 risers, when I started out mining years ago.

 

[ArdWar]

There's no tolerances, nor limit either per se. The card can draw as much power as they can and the only thing limiting that is the physical slot melting or the +12V rail droop too much that it triggers the UVLO or something in the MB or card itself. Both of those limits can be wildly different between cases depends on slot rating, contact quality, age, cleanliness, PCB design, airflow, etc.

The 75W limit is specification target. It's calculated (and guaranteed) that 75W power draw is safe and can be depend upon for reliable operation. It's the card's responsibility to limit the draw to 75 W, not the MB. You can even short it and it'll trip the fuse instead of limit or crowbar the slot.

 

[W1zzard]

There's no tolerances, nor limit either per se. The card can draw as much power as they can and the only thing limiting that is the physical slot melting or the +12V rail droop too much that it triggers the UVLO or something in the MB or card itself. Both of those limits can be wildly different between cases depends on slot rating, contact quality, age, cleanliness, PCB design, airflow, etc.

The 75W limit is specification target. It's calculated (and guaranteed) that 75W power draw is safe and can be depend upon for reliable operation. It's the card's responsibility to limit the draw to 75 W, not the MB. You can even short it and it'll trip the fuse instead of limit or crowbar the slot.

qft

 

[damric]

Good warnings, thank you. But consensus is go ahead and try this right?

The card in question for the experiment will be a quadro m2000, which is similar to a GTX 950 but normally runs at a much lower clock rate. I'll be removing the single slot cooler and hooking it up to cold water, putting heat sinks on the vrms and memory chips, and pointing fans on front and back of the card. I'm going to go past the 75w slowly, if it even lets me, and probably stop around 100w for long testing to see how everything is holding up.

 

[steen]

3090 pulls 90W+ from PCIEx16, so you're unlikely to have issues unless you destabilize power distribution through a shunt mod.

 

[damric]

3090 pulls 90W+ from PCIEx16, so you're unlikely to have issues unless you destabilize power distribution through a shunt mod.

I don't plan on any hardware mod at this point. I'm going to be screwing with firmware and if necessary the driver, and if absolutely necessary I might have to use a linux with a certain tweak. I have some more expert overclockers advising me on this one that have had good success with Maxwell firmware and driver mods. Whatever I do I'll be sure to post a small work log here whether this is success or failure. Success, in my mind would be at least getting the core over say...1400MHz without melting anything lol. I'm really hoping for over 1550MHz sustained but that might really be a stretch on the PCIE power.

 

[damric]

Whelp can't edit o.p. so here's some photos of what I'm going to be testing.

The stock cooler looks so bad and was barely keeping it cool so I took it apart today, removed the thermal pads, cleaned it all very well, and I put a nice MSI TFIII cooler on there that just barely fit with some very small modifications. I thermal glued some aluminum sinks onto the VRMs and the VRAM. I actually had to grind down about 1mm off of the sinks to get them to fit. I have a universal water block and can throw it on a water cooled bench if necessary. There's fans blowing on front and back of the card now and it's folding at 30C instead of 60C at stock clocks. So thermally it looks like putting the MSI cooler on there (from an old HD 7850) is probably going to be enough to get pretty far overclocking already. I forgot to snap a picture of the other cooler. I'll do that tomorrow probably.

Now the thing about most workstation cards from Nvidia and AMD is that they lock out software OCing ability with stuff like Afterburner, so it's pretty much going to be firmware foolery to get this baby clocking. I wonder how far this M2000 (similar to GTX 950) will go on stock voltage. We'll know in a couple days.

 

[looniam]

the card will stick w/firmware limits.

modded a lot of maxwell cards, usually just see a 75 watt entry though asus is famous for 88 watt limit there btw. but never saw a 66 watt (12v only) entry before.

 

[Tatty_Two]

Whelp can't edit o.p. so here's some photos of what I'm going to be testing.

You can now should you need to again.

 

[eidairaman1]

Whelp can't edit o.p. so here's some photos of what I'm going to be testing.

View attachment 233913
View attachment 233914
View attachment 233915
View attachment 233916
The stock cooler looks so bad and was barely keeping it cool so I took it apart today, removed the thermal pads, cleaned it all very well, and I put a nice MSI TFIII cooler on there that just barely fit with some very small modifications. I thermal glued some aluminum sinks onto the VRMs and the VRAM. I actually had to grind down about 1mm off of the sinks to get them to fit. I have a universal water block and can throw it on a water cooled bench if necessary. There's fans blowing on front and back of the card now and it's folding at 30C instead of 60C at stock clocks. So thermally it looks like putting the MSI cooler on there (from an old HD 7850) is probably going to be enough to get pretty far overclocking already. I forgot to snap a picture of the other cooler. I'll do that tomorrow probably.

Now the thing about most workstation cards from Nvidia and AMD is that they lock out software OCing ability with stuff like Afterburner, so it's pretty much going to be firmware foolery to get this baby clocking. I wonder how far this M2000 (similar to GTX 950) will go on stock voltage. We'll know in a couple days.

Get a gpu-z screenshot.

I remember a Fire GL/Fire Pro Card that was a cherry picked R300 being worthwhile for gaming

 

[damric]

Get a gpu-z screenshot.

I remember a Fire GL/Fire Pro Card that was a cherry picked R300 being worthwhile for gaming

This is going to be a tweaker card to experiment folding on, similar to what I did with the W2100 and W4100 I overclocked over the last couple months. Though I will run it through some benches for HWBOT points.

 

[eidairaman1]

Ugh thats a 950 under the hood

NVIDIA GM206 GPU Specs

1024 Cores, 64 TMUs, 32 ROPs

www.techpowerup.com

This is going to be a tweaker card to experiment folding on, similar to what I did with the W2100 and W4100 I overclocked over the last couple months. Though I will run it through some benches for HWBOT points.

View attachment 234019

Is that a M2200?

Upload your bios to the vga bios collection through gpu-z using the black arrow next to uefi checkbox, paste link here, i want to compare your bios to other GM206 based cards

 

[damric]

Ugh thats a 950 under the hood

NVIDIA GM206 GPU Specs

1024 Cores, 64 TMUs, 32 ROPs

www.techpowerup.com

Is that a M2200?

Upload your bios to the vga bios collection through gpu-z using the black arrow next to uefi checkbox, paste link here, i want to compare your bios to other GM206 based cards

It is very much similar to a GTX 950 but with reduced clocks to be able to work with no external power and the single slot cooler.

It's an M2000. Careful, as there like 30 versions of GM206 in various forms including unique OEM versions, weird cut-downs, many mobile versions, ect.

Funny if you look at the picture I took of the PCB it looks like the GPU can be slotted in and out modularly like an AMD CPU. I've never seen that before on any GPU.

Here's your GPU-z BIOS. Some funky looking pages when you open up MBE.

Edit: also attaching GPU-Z Sensor readout while folding at stock. There are some 80W+ PCIE draw spikes even at stock, but mostly like 50W. I'm done with my stock clock testing and going to be overclocking this little guy tomorrow hopefully.

 

[eidairaman1]

It is very much similar to a GTX 950 but with reduced clocks to be able to work with no external power and the single slot cooler.

It's an M2000. Careful, as there like 30 versions of GM206 in various forms including unique OEM versions, weird cut-downs, many mobile versions, ect.

Funny if you look at the picture I took of the PCB it looks like the GPU can be slotted in and out modularly like an AMD CPU. I've never seen that before on any GPU.

Here's your GPU-z BIOS. Some funky looking pages when you open up MBE.

Edit: also attaching GPU-Z Sensor readout while folding at stock. There are some 80W+ PCIE draw spikes even at stock, but mostly like 50W. I'm done with my stock clock testing and going to be overclocking this little guy tomorrow hopefully.

Yeah 768 shader tells me its a 950

 

[damric]

Ok so hitting a roadblock with core frequency. Even though the firmware programming was successful according to GPU-Z readout, it's still topping out at 1227MHz. I'm thinking the driver is ignoring the firmware. The raised power limits I programmed worked though, so it seems to run that 1227MHz constantly where as at stock it would fluctuate a bit. I'm studying a few successful maxwell quadro overclocks, and they had older nvidia 442.92 drivers that I can't seem to get from Nvidia's website.

If I can't get it to overclock in Windows 10, I might try Ubuntu since a buddy of mine has a sort of Nvidia driver hack for that. If that doesn't work, then I'm considering flashing to a GTX 950 firmware. Problem I foresee with that is the display outputs are different and could cause an issue. I can run the card with no display output though since this will be doing folding compute tasks.

Looking at these, thinking about which one to try first. EVGA has the only 4GB model. I wonder if that even matters.

EDIT:

Haha I cracked it, finally. I dug into the ASUS GTX 950 card firmware that uses slot power only, and copied some clocks and voltages over while leaving my enhanced power limits in place. So it's technically running a Quadro BIOS, but is performing now like a GTX 950. I see average slot power at 55W, max at 88W for 1354MHz steady boost clock. I'll push it further tomorrow if it survives folding overnight.

 

[eidairaman1]

Ok so hitting a roadblock with core frequency. Even though the firmware programming was successful according to GPU-Z readout, it's still topping out at 1227MHz. I'm thinking the driver is ignoring the firmware. The raised power limits I programmed worked though, so it seems to run that 1227MHz constantly where as at stock it would fluctuate a bit. I'm studying a few successful maxwell quadro overclocks, and they had older nvidia 442.92 drivers that I can't seem to get from Nvidia's website.

If I can't get it to overclock in Windows 10, I might try Ubuntu since a buddy of mine has a sort of Nvidia driver hack for that. If that doesn't work, then I'm considering flashing to a GTX 950 firmware. Problem I foresee with that is the display outputs are different and could cause an issue. I can run the card with no display output though since this will be doing folding compute tasks.

Looking at these, thinking about which one to try first. EVGA has the only 4GB model. I wonder if that even matters.

EDIT:

Haha I cracked it, finally. I dug into the ASUS GTX 950 card firmware that uses slot power only, and copied some clocks and voltages over while leaving my enhanced power limits in place. So it's technically running a Quadro BIOS, but is performing now like a GTX 950. I see average slot power at 55W, max at 88W for 1354MHz steady boost clock. I'll push it further tomorrow if it survives folding overnight.

Glad the research on my end helped lol.

Just dont stress the mobo too much

 

[oobymach]

Maxwell was more limited for overclocking than kepler, I had a 750ti which was maxwell and I got maybe 100mhz over stock, it was a slot powered card also, decent enough for 60fps gaming at the time. The ram oc is fairly substantial though for those cards, factory was 6ghz but you can push them to 7ghz without issue.

 

[damric]

Ok well it looks like it is stable for folding at over 1500MHz. I was able to run 3D benches at over 1550MHz. Anything much over that crashes at this voltage. I'm not going to bump voltage as that will probably severely increase the power draw. I'm happy with these statistics I pulled from GPU-Z logs. Even with the huge overclock, it's averaging only 60W and has occasional peaks over 100W. I'm ok with that. Next month it will fold non-stop for ETF, so we'll see if there is a longevity issue with the slot power, but I bet it will be fine. It's still on an air cooler. I have a better air cooler coming, and I do have a water block if I really want to get the temperature down, but as you can see it runs cool already.

While this M2000 card itself is too weak to be significant to most end users, there are greater implications now if someone wanted to overclock larger Maxwell Quadros now that we have a working theory that the overclocking locks can indeed be cut. I see the M4000s (similar to GTX 970 I think) are going for less than $200, which is still a very good 1080p card for a desperate gamer in desperate times.

I ran a 3DMark, got the M2000 world record, and of course HWBOT is down all weekend for upgrades.

 

[looniam]

Maxwell was more limited for overclocking than kepler, I had a 750ti which was maxwell and I got maybe 100mhz over stock, it was a slot powered card also, decent enough for 60fps gaming at the time. The ram oc is fairly substantial though for those cards, factory was 6ghz but you can push them to 7ghz without issue.

sorry man but maxwell left kepler in the dust ocing. the 780 and 780ti i had got 300-350mhz whereas both 980tis got 400-450mhz. its no contest. not to mention kepler was a pita to unlock the voltage, because if you didn't unlock voltage control - you would get nowhere ocing.