tictoc

The Fury-X looks like it will be pretty competitive in the AMD category. My daily average over the last 15 days would put me at the top of the AMD cat if I hadn't missed the first part of the month.

A home for the cluster.

10M ppd is just nuts. How much power is it pulling on the more demanding WUs?

That is adequate flow assuming your meter is accurate. .5 gallons/minute is about the minimum acceptable flow rate, and pretty much anything above 1 gallon/minute you'll see increasingly diminishing returns. You are right in the middle, so it should be fine as long as your water and component temps are good.

In the future you shouldn't have to use a secondary GPU to recover a bad BIOS flash. All you should have to do is switch to a good BIOS, boot, and then once in the OS flip the switch to the BIOS that you want to flash. To make things easier I would just run on the OC BIOS, flip both the voltage switches to on, and see how far you get with Precision X. I think this is what I ultimately ended up doing on Windows, since I was able to pretty much max the card out that way. I think you might have to use an old driver to use the xoc BIOS I posted. I honestly don't really remember for sure. I wasn't nearly as thorough in taking notes and keeping track of changes back then. You might try the 347.88 drivers.

I guess you're running Windows on the headless machine? Do you have ssh set up? You can flash the GPU from a terminal (even on Windows), so if you plan on keeping it headless, I would set up ssh. I'm not positive if the Windows NVIDIA driver has the bits necessary to OC from the terminal, but you can definitely use nvidia-smi to monitor the card. As far as where you're at right now, the easiest thing to do if you have another machine, would be to throw the card in that machine and flash back to the original BIOS. It looks like that GPU only has a single BIOS, so rather than blind flashing the card, throwing it in another machine would be the easiest. Once you're back up and running, I would personally just tweak a stock BIOS with Maxwell BIOS editor, rather than trying to cross-flash. Unless you have really good cooling in your 4u chassis, you will probably be temp limited before you are volt limited. The other benefit to flashing your own BIOS to the card is that it will just be a set it and forget setup. You can just set the card to run at the max stable frequency, power, and fan speed, rather than messing around with clocks and fans from the OS.

The A2000 is listed on the latest GPUS.txt. https://apps.foldingathome.org/GPUs.txt Looks like it was added fairly recently, since it is not listed on the GPUS.txt on any of my machines. @NBrock if it is not listed in your GPUS.txt, then you should replace your GPUS.txt with the latest.

Link to original source. https://www.pugetsystems.com/labs/articles/Most-Reliable-PC-Hardware-of-2021-2279/

If you raise the power limit to max, it should just go there without doing anything special. On Linux I just used nvidia-smi (command line tool) to set the power limit to 500W. You should be able to do the same thing on Windows, if you find that you can't bump the power limit up to max. I think nvidia-smi is now in the default $PATH on Windows 10, so you should just be able to open an elevated command prompt and then set the power limit with the below command. nvidia-smi -pl 500 Here are a few BIOS for the 980 KPE, the Classified Voltage Controller (software voltage control, either version should work, but 2.1.0 might have been for 780ti), and an old version of Precision X. Precision X will read the actual voltage on KPE cards and you can also use kboost to lock the card to the P0 power state. You might not get too far without going cold. Maxwell voltage doesn't really start scaling until you get the temps way down, but you should be able to push it further than where you are at right now. Happy OC'ing. kpe980_1.3v.zip kpe980_xoc.zip classified_v2.1.0.zip classified_v2.1.2.zip evga_precision-x_v5.3.6.zip

Is the power limit set at 450W, and just not pulling that much power, or are you unable to set the max power limit? My 980 KPE will push quite a bit of power (more than 250W) without issue depending on the work load. It's been a bit since I messed around with it in Windows, but I am pretty sure I have an xoc bios and the classified voltage tool. I'll see what I have in my notes for the 980 KPE when I get home this evening.

For whatever reason the CPU-Z bench is unaffected, it also doesn't care what priority is set. For Cinebench R11.5/15/20/23 the above is always worth a few more points.

copyright tictoc Running Linux kernel 5.15.13 with ROCm 4.0. *Edit* Memory OC for Fiji was removed from the Windows drivers many years ago, but this is Linux, so a little kernel hacking will fix that right up.

Here is a proper run not with WINE or on a VM. Since I was already rebooting to run Windows, I gave the 3960X a little more juice. 4550MHz all-core OC. Memory is pretty average running 8x 16GB of ECC UDIMMs at 3200, but I'm not sure if the CPU-Z bench even cares about memory.

I lost a bit of time with some failed WUs yesterday trying to see if the Fury X could magically run at higher clocks after sitting on the sidelines for about 4 years. The answer to that question was a big NO! The only tuning left would be to bump the memory up a bit, but the ppd gain compared to the down time is probably not worth it.

Maybe I'll boot into Windows later and do a proper run, but here's an unoptimized Windows 10 VM, running on a mostly idle host.

Looks like 3x 1m. https://www.corsair.com/us/en/Categories/Products/Custom-Cooling/Tubing/hydro-x-hardline-tubing-14-config/p/CX-9059010-WW#tab-package-contents

Back up and running. I just remebered that I had a spicy BIOS flashed to the second slot on my Fury X, so now I am running that and slowly increasing the clocks a bit. I also forgot just how good of a clocker this card is relative to other Fiji GPUs.

All of my QD3s are 4 years old or older. I wouldn't worry too much about my recent experience. I used to run a bunch of the QD3s without issue, and I sold all of my other ones when I was getting rid of a bunch of unused fittings a few years back. While I'm here, here's a run of the CPU-Z bench in Linux via wine. Multi-thread bench is a little bugged with CPU usage topped out at 68%.

I have a few that stick open and one with a slow drip, which is why I currently don't have any on the main bench. Using CPC QDC's on my main machine and on the testing bench, and they have performed flawlessly.

k4m1k4z3 was a member of the legendary team Infinity during our reign of dominance.

You should be able to do this all from the command line without too many issues. Can you post your current working xorg.conf ?

I'll be back up and running later today. The 2080S and a Fury-X are currently on the bench, and I'm just giving everything a bit of a cleaning with some Blitz Part 2. @BWG I'll let you know if I'll still be temping in a slot. I have a few messages out, but so far no replies. ;)

No points from me to finish out December, since the year ended with an extended power outage. There must still be some blown transformers or downed lines, because the power is kind of back now at 90-95v.

I'm going to call it #1 in the world. Anonymous - default setting for new installs ICCluster - Spare time on the GPU compute cluster at EPFL NVIDIA SaturnV - NVIDIA AI/ML clusters EarlRuby - VMWare engineer who works on GPU accelerated containers. I'm assuming this is all running somewhere in one of VMWare's datacenters. BWG - One dude crushing it from his home office.

Missed the bot on this one. Nice to see you here @LarsL.