i run all my GDDR6X cards at +1000 mem clocks and that seems to be stable. I used to run my watercooled card at +1500, and that also seemed fine, but without memory temp readings, and without extensive testing, that might not be stable on an air cooled card. or maybe not necessarily even stable on a watercooled card and my sample was exceptionally good or something.
there's no reason to be running arbitrarily high clocks though. it gives you such little gains as it is. stick to 500-1000 mem OC on GDDR6X for stability. or even 0 will be the most stable.
running keepP2 can help with stability of high mem clocks too. but that usually prevents crashing outright rather than slowdowns. and if you keep mem OC at a reasonable level, you probably don't even need it at all anymore. it was most helpful with GDDR5X Pascal cards which could take a very high OC, but less useful for Turing and Ampere IMO. I don't even use it on some of my systems anymore since I don't have any GDDR5X cards anymore.
We went with the MSI 4090 Suprim Liquid (here) and I am really happy with these. I know there is some validation issues with the 4090 but I think this will be worked out with time. Our MSI 4090 GPUs stay incredibly cool under full load (never seen it above 56C, then the fans/pump speed up and it stays ~48C- fans/pump are usually at ~30%). They are basically silent and only take up two slots. Really amazing piece of hardware.
Still want to get my hands on a RTX 6000 Ada to see if there are validations issues with that GPU as well... maybe next school year (but probably not...).
The SUPRIM cards are also the only card that I know of where the vendor actually bothers to bin the chips before making the card.
All the other high end graphics cards are still more or less a silicon lottery. The Suprim cards are/were binned so that the chips on them could sustain much better than average overclocks. Paired with the AIO water cooling, they sell for a big premium since over clockers want a known quantity.
I'm sure many people have many various models of 3080Ti across the project.
pretty much any 3080Ti will output within a few percent of one another. the boost clocks help a little. not enough to worry about it or buy a specific model for IMO.
The invalid rate for the rtx 4090 models appears to be higher than the 3080 ti models.
I was eyeballing some results from an Rtx A5000 system. It looks like it also has a very high invalid rate.
Grp#1 tasks.
Anyone got good news?
Our RTX A4500 and A6000 GPUs are actually really good and have a relatively low invalid rate. Take a look at most of our systems running- most have A4500 GPUs. Our host with the highest RAC has two A6000s and one A4500.
i run all my GDDR6X cards at
)
i run all my GDDR6X cards at +1000 mem clocks and that seems to be stable. I used to run my watercooled card at +1500, and that also seemed fine, but without memory temp readings, and without extensive testing, that might not be stable on an air cooled card. or maybe not necessarily even stable on a watercooled card and my sample was exceptionally good or something.
there's no reason to be running arbitrarily high clocks though. it gives you such little gains as it is. stick to 500-1000 mem OC on GDDR6X for stability. or even 0 will be the most stable.
running keepP2 can help with stability of high mem clocks too. but that usually prevents crashing outright rather than slowdowns. and if you keep mem OC at a reasonable level, you probably don't even need it at all anymore. it was most helpful with GDDR5X Pascal cards which could take a very high OC, but less useful for Turing and Ampere IMO. I don't even use it on some of my systems anymore since I don't have any GDDR5X cards anymore.
_________________________________________________________________________
What is the current
)
What is the current make/model of pcie ribbon cables you recommend?
I am getting ready to configure a 4-5 GPU system. And will need the space that ribbon extension cables provide.
I am interested in pretty short ribbon cables as well as pretty long ones. Pcie Gen 3 is fine.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
https://www.anandtech.com/sho
)
https://www.anandtech.com/show/18860/asus-unveils-two-new-geforce-rtx-4090-gpus
And this is what liquid cooling will get you.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M
)
We went with the MSI 4090 Suprim Liquid (here) and I am really happy with these. I know there is some validation issues with the 4090 but I think this will be worked out with time. Our MSI 4090 GPUs stay incredibly cool under full load (never seen it above 56C, then the fans/pump speed up and it stays ~48C- fans/pump are usually at ~30%). They are basically silent and only take up two slots. Really amazing piece of hardware.
Still want to get my hands on a RTX 6000 Ada to see if there are validations issues with that GPU as well... maybe next school year (but probably not...).
The SUPRIM cards are also the
)
The SUPRIM cards are also the only card that I know of where the vendor actually bothers to bin the chips before making the card.
All the other high end graphics cards are still more or less a silicon lottery. The Suprim cards are/were binned so that the chips on them could sustain much better than average overclocks. Paired with the AIO water cooling, they sell for a big premium since over clockers want a known quantity.
Two questions. Are there
)
Two questions.
Are there any other brand/models of rtx 3080 ti GPUs besides EVGA and Founders Edition represented in the e@h project?
In a review of rtx 3080 ti's there were specs for boost speeds. Do the higher boosts usually translate into faster e@h GPU processing?
I am wondering about potential upgrades and / or replacement GPUs.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
I'm sure many people have
)
I'm sure many people have many various models of 3080Ti across the project.
pretty much any 3080Ti will output within a few percent of one another. the boost clocks help a little. not enough to worry about it or buy a specific model for IMO.
_________________________________________________________________________
The invalid rate for the rtx
)
The invalid rate for the rtx 4090 models appears to be higher than the 3080 ti models.
I was eyeballing some results from an Rtx A5000 system. It looks like it also has a very high invalid rate.
Grp#1 tasks.
Anyone got good news?
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: The invalid
)
known issue with 40-series cards. check the 4090 High invalids thread.
_________________________________________________________________________
Tom M wrote: The invalid
)
Our RTX A4500 and A6000 GPUs are actually really good and have a relatively low invalid rate. Take a look at most of our systems running- most have A4500 GPUs. Our host with the highest RAC has two A6000s and one A4500.
They have been rock solid.