I expect to run my Bit-coin GPU-based Mining box to the end of the month or so and then NNT till I am dry.

I believe that I can call running 16 gpu's a near success. I have had one of my gtx 1060 3GB's lose connectivity after running overnight. I think I know which one.

For Seti@Home it looks like I might be able to run a few cpu tasks without over-committing the cpu. But currently it looks more likely that it would run better as a pure gpu system.

I have set the E@H to run gpu only as one of my backup projects as well as WCG to run cpus if threads become available.

It looks like the only way to run a high-count gpu box and have enough processing power left to run cpu threads is with a cpu that has roughly 50% more threads available than you are using on the gpus.

So I am pretty confident you could reduce the available gpus to something like 10 or maybe 12 and run CPU tasks with the 8c/16t cpu I am using (Intel).

The main reason I wanted to run this experiment was because I have been bumping my head against upper limits of 6-7 (and twice unstable 9 arrangements of) gpus for a couple of years. I wanted to confirm that a purpose-built motherboard could in fact run as many gpus as it had slots...

Tom

---

A proud member of the OFA (Old Farts Association).

I have been having connectivity issues with the gpus. So I have been slowly but surely unplugging them. I only have 10 plugged in at the moment. Its that dreaded "paused..." message.

If I still have no "paused..." messages showing up by tomorrow I will start plugging the others back in. Since I could be having more than one cable failure it will take a while to confirm it.

Then I suppose I will have to buy some "regular" length U green cables because it looks like any failures I have are on the top of the rack which is the furthest away from the MB.

Tom

---

A proud member of the OFA (Old Farts Association).

I got to looking at the Intel 9th Gen cpu list for LGA 1151 sockets.

I was curious if there were any cpu with more than 8c/16t available for the 9th gen socket.

Nope.

So if you want a high core count/thread count high gpu count gpu system you are going to have to jump out of the Bit Coin specific multiple gpu motherboard series.
It is possible that an low-end 1st gen EYPC (16c/32t) with an riser expander board or two could approach 20+ gpus running at Gen1/Gen2 speeds.

If you are only interested in upto 18 video cards and "some" cpu processing you can probably get by on a MSI B360F-Pro motherboard with an 8c/16t 9th gen cpu.
If you are only interested in upto 19 video cards and no cpu task processing you can probably get by on the 19 gpu slot B250 motherboard with an 4c/8t cpu.

Good luck.

Tom M

---

A proud member of the OFA (Old Farts Association).

Well, the whole concept of the Mining boards is to use just enough CPU to run the GPUs in Mining configuration. Trying anything else is likely to be disappointing.
After some experience at Einstein@Home I've found that project to be Best for the Mining boards using AMD GPUs. Using nVidia GPUs is restricting as there is some strange conflict with the App needing a Full CPU even though the PCIe Bandwidth is basically Zero to 1% on a Gen3 x1 connection. How can you need a Full CPU core when the App uses 1% of a x1 connection? It does think it actually needs the Full core though as the run-times extend much more than the SETI times do when running just One extra GPU even though SETI uses much more bandwidth than 1%. Something strange going on there, uses 1% of the connection for a Full CPU core?

You don't have that problem with the AMD App, running Two tasks on an AMD 570 uses 10% of a Quad core i5, meaning you should be able to run up to 19 AMD 570s with an 8 core i7 at Einstein. The 570s although Cheap, are getting a little old. I'm thinking about buying One of the newer AMD 5700s just to see how it does. Currently I have the Two Biostar boards running 8 NV GPUs each and the ASUS board running the One AMD 570 with an i5-6600 CPU. I had to convert the Hackintosh over to Ubuntu as the nVidia OpenCL driver is trash on a Mac. Probably why Apple doesn't do nVidia any more, their OpenCL driver can't even count PCIe slots on a Mac or Hackintosh.

the Einstein app heavily favor the AMD cards. you have discovered what a lot of us already knew about the nvidia OpenCL app over there. it requires a lot of CPU support, where the AMD app does not. If I recall correctly, it's something about constantly polling the GPU to ask if it's finished the computation or something like that. so it's not sending a lot of data back and forth to utilize the PCIe bus much, but it's keeping the CPU constantly busy.

from what other's have posted about the 5700, it works well, it crunches fast. BUT you might have to move over to Windows for that. If you recall, as I'm sure you're very aware, the 5700 cards had a driver problem causing incorrect results. this problem affected Einstein also, and when AMD made the fix for SETI, it fixed Einstein as well. But as I recall, they only fixed the Windows driver, and not the Linux one. I don't think we ever had confirmation if the Linux drivers were ever fixed since there was almost no one running that configuration.

you could look into getting a Radeon VII also. not too expensive on the used market, and probably the fastest card available for Einstein at the moment.

It will be nice if petri can get another Special application over there, then maybe I'll crunch there as more than just a backup project with 0 resource share.

---

Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

Ian,
If I have the CPU cores and want to stay with Ubuntu, how do the Nvidia cards compare to AMD? Will the same $ get you about the same processing capability for E@H tasks?

---

The AMD cards are cheaper in general compared to the equivalent performance of Nvidia cards. They have better math capabilities. Most importantly, you can run concurrent tasks on AMD cards while you can't with Nvidia. So production is higher on AMD.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

This thread pretty much cured me of wanting a discontinued Radeon 7, I had some pretty bad problems running Einstein GRP work on my Radeon VII starting in late September... I'd much rather go with the Newer, much less power hungry 5700. If it doesn't work yet in Linux then I'll run the Hackintosh for a while longer until it does. The Mac AMD App does work at Einstein, it's just a little slower than the Linux App. From the 5700 run at SETI, we know the 5700 does work on a Mac, and the AMD Mac driver can count slots...

Polling isn't the problem at Einstein. Both the Older CUDA App and the Special App used Polling and there wasn't any problem running more tasks than cores. There is something else going on there. Even the SETI OpenCL NV App didn't use a Full CPU core in Linux or Mac, that was a Windows thing at SETI.

BTW, running Two tasks on a $100 AMD 570 makes it faster than a NV 2070 at Einstein.

Tbar,

I can afford a $100 card. Does it have to be under Windows to avoid the driver problem for Linux?

---

The AMD cards work fine in Ubuntu at Einstein, the only question is the newer 5700. This is my older RX 570, now running two tasks at a time, https://einsteinathome.org/host/12822598. If you check on the recent tasks they are under 20 minutes for two tasks, which is pretty normal for a 570. Also note there aren't any Invalids on the 570 yet. Invalids at Einstein are apparently pretty normal with at 1% Invalid rate normal. NewEgg has the NEW 570s for $119 right now, they are a little cheaper at eBay.

you can base your purchase off one user's bad experience if you want. but he probably had a bad card or some other conflict that he couldnt figure out. it happens. but you can't deny their power. all of the fastest systems are using them. the first system doing like 8,000,000 per day is using 6 of them. at least the RVII doesnt have the known driver issue of the rx 5700.

but yeah the nvidia app at Einstein is much less optimized than the AMD app. that's common knowledge there.

for comparison, a system with 2x Radeon VIIs is about the same speed as my 7x RTX 2070 system. undoubtedly using a lot less electricty, even on the much less efficient GPUs like the RVII

the second benefit of the RVII is it's MASSIVE double precision performance. this makes it a powerhouse for running Milkyway@home also. second only to the Titan V, at a fraction of the cost. certainly a Radeon VII system would make a great Einstein/Milkyway combo setup.

---

Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

Thanks, gentlemen! I'll look into the AMD options as part of deciding which rigs to put where.

---

Thanks, gentlemen! I'll look into the AMD options as part of deciding which rigs to put where.

There shouldn't be any driver issues under Linux for the older cards like the 570.
https://www.amd.com/en/support/kb/release-notes/rn-amdgpu-unified-linux
The only slip up is the possibility of not installing the OpenCL component of the drivers using the Auto install. You need to read the instructions and invoke the install script with the proper parameter to install the OpenCL component.
https://amdgpu-install.readthedocs.io/en/latest/install-script.html#invoking-the-amdgpu-install-script
Which would be:

$ ./amdgpu-install -y --opencl=legacy

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

So is there any science/technology project out there that comes close to being as efficient a user of Nvidia gpus as Seti@Home did?
I am not expecting "special sauce" performance just "good" performance.
And doesn't push the gpus as hard as Gpugrid does?

Tom

---

A proud member of the OFA (Old Farts Association).

So is there any science/technology project out there that comes close to being as efficient a user of Nvidia gpus as Seti@Home did?
I am not expecting "special sauce" performance just "good" performance.
And doesn't push the gpus as hard as Gpugrid does?

Tom

If you find one please tell us.
AFAIK The GPUGrid & E@H are the "less worst" for the Nvidia stuff at least on this areas.
Please correct me if i'm wrong, no other project actually uses heavily optimized CUDA 9 or 10 apps to crunch like Seti Linux Special Sauce.
And BTW a lot of them uses DP where the AMD stuff has a clearly superior capacity (most of them at least).

---

I’d say the GPUGRID apps are just as optimized as the Special App. They make full use of the GPU and even uses the PCIe bus more.

I got a line on some custom riser adapters. To add more power to the motherboard. So soon my 10-GPU system will have all 10 GPUs with x8 lanes, not just 8 that I’m currently running

---

Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

you can base your purchase off one user's bad experience if you want. but he probably had a bad card or some other conflict that he couldnt figure out. it happens. but you can't deny their power. all of the fastest systems are using them. the first system doing like 8,000,000 per day is using 6 of them. at least the RVII doesnt have the known driver issue of the rx 5700.
but yeah the nvidia app at Einstein is much less optimized than the AMD app. that's common knowledge there.
for comparison, a system with 2x Radeon VIIs is about the same speed as my 7x RTX 2070 system. undoubtedly using a lot less electricty, even on the much less efficient GPUs like the RVII
the second benefit of the RVII is it's MASSIVE double precision performance. this makes it a powerhouse for running Milkyway@home also. second only to the Titan V, at a fraction of the cost. certainly a Radeon VII system would make a great Einstein/Milkyway combo setup.

OK, I've looked it over some more, and agree the problem may associated with power limiting a Power hungry GPU. Someone made that claim later in the thread. So, if the govmint really is gonna send me $1200, I may think about a Radeon 7...or two.

BTW, the RX 570 I bought from eBay a couple of years ago, for $100, still hasn't produced an Invalid at Einstein....
But, I may be changing it to a different board later today.

So is there any science/technology project out there that comes close to being as efficient a user of Nvidia gpus as Seti@Home did?

Asteroids@home have released a new CUDA 10.2 application for Windows x64: 23 Mar 2020 news item

Some complaints in the news thread about all tasks running on device zero, but they seem to have sorted that out. I found a lot of screen lag when I ran one on the card driving the monitor, but it was OK on the headless secondary card.

Unfortunately, they only released a Windows application. Nothing for Linux. Still only the old CUDA55 available for Linux.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)