210k ppd on a geforce 970

[runpaint]

Ok, each card now has a core17. But instead of 250k and 350k ppd, I'm only getting 152k and 198k. I rolled back the drivers but that only took me to 344.11 and no change, so now I'm downloading 340.43 (I had the older driver version when I first installed the cards). But I have to go to work so it'll have to wait a few hours.

[runpaint]

Ok, I tried the older drivers. I have one computer with the 970, the 980, and a 750ti. So I paused folding, installed the driver, rebooted, and all 3 slots lost all progress and failed all the work units and then couldn't fold any new ones either. So I'll either keep it just like that, or go back to the latest driver.

[runpaint]

Well, I went to rollback the drivers, but my computer didn't have any video cards in the device list (the screen resolution had gone very low after the driver change). So I rebooted, and everything was back to normal with the current driver installed. Also, now one of the cards is emitting a faint whine/hiss.

[ChristianVirtual]

just installed a new Zotac GTX 970 into my Linux test box with Ubuntu 14.04 (i3, 3.3GHz) and 343.22 driver. With a 9201 WU it makes TPF 1:53 and 270kPPD; pulling total system 213 Watt from the wall (100 W idle; 113 W for folding). Looks really nice until here ...

[runpaint]

Well you won't be getting the core15 work units with lower PPD, but you might end up not folding anything at all sometimes. But that might give you better overall ppd, since your slot will be ready for the next available WU while mine will be busy on a core15. Of course it'll have to make the "next attempt" at the right time, which I still don't fully understand.

Sometimes I'll notice that I've lost wi-fi and need to reset it, so it's waiting to send and receive work. But the next attempt is in 45 minutes, so I pause and unpause the slot to make the next attempt 1 minute. Why would the next attempt be so long? I haven't watched it, but I think it waits longer after each attempt so it's not tying up the server with constant requests. But when there's a temporary loss of connection on my end, I'd rather it keep trying every minute since it doesn't use anything but my computer.

[bruce]

Sometimes I'll notice that I've lost wi-fi and need to reset it, so it's waiting to send and receive work. But the next attempt is in 45 minutes, so I pause and unpause the slot to make the next attempt 1 minute. Why would the next attempt be so long? I haven't watched it, but I think it waits longer after each attempt so it's not tying up the server with constant requests. But when there's a temporary loss of connection on my end, I'd rather it keep trying every minute since it doesn't use anything but my computer.

You've figured out pretty well except for the fact that you're looking at it from the perspective of one person, whereas Stanford looks at it from the perspective of one Server. First, suppose a server is down for an extended period of time. Many donors will complete their assignment and will try to upload the results and repeatedly fail. If all clients were retrying every 1 second, then in the first few minutes after the server comes back on-line, the server will be flooded with upload requests, overloading the server and successfully uploading almost nothing. Second, assume that every time an upload fails, the time before the next retry is a bit longer. Now when the server comes back on line, a manageable number of clients will retry, avoiding the problem of saturation and spreading the load out over a progressively longer period of time, commensurate with the duration of the outage (and therefore commensurate with the number of clients with work queued for upload).

[runpaint]

I guess what I'm saying is that the client doesn't differentiate between the server being down and my internet connection being down. Like when your web browser says "Can't find the server at http://www.google.com, check the address and try again" instead of saying "Unable to connect to the internet". If the client could detect the status of the internet connection, then it could automatically try resending as soon as a connection is detected. But it works well enough as it is, and any slight downtime isn't enough of a problem to warrant a "fix".

[Sn1ken]

Ok, I tried the older drivers. I have one computer with the 970, the 980, and a 750ti. So I paused folding, installed the driver, rebooted, and all 3 slots lost all progress and failed all the work units and then couldn't fold any new ones either. So I'll either keep it just like that, or go back to the latest driver.

If you are using X99 2011-3 motherboard it is possible to end up with nice score. But if you are using 1150 CPU, you better remove one card.

Of some reason the GPU is polling the CPU and this needs PCIe lanes and you need a solid CPU speed to cope with the GPU. AMD is not using the same amount of resource.

What kind of motherboard and cpu do you use?

A lot of people deny this fact, but I know myself.

My 970 pulls 301k PPD on core 17. And 78k PPD on core 15.

[runpaint]

Yeah, I looked at 2011, but what's the point of paying extra for 2 PCIe 3.0 x16 slots when the processor only has 28 lanes?

I have some 1155 and some 1150, and some AMD boards. I'd rather spend $600 for a new 980 and plug it in to one of my computers, rather than having to spend another $900 for a new computer every time I buy a new card or two. Don't get me wrong, if I had the money I'd get the fastest processor and the fastest motherboard to go with these cards. But I'm looking on the bright side and hoping that the limited performance of these mobos will act as a slight underclock on the cards and make them last longer. (it's not the same as underclocking, but they're not working as hard)

I've been looking at the i7 3820 with an X79 motherboard, which could let me run 2 cards at pcie3.0 x16 and one at x8. And I can get the entire new computer with that setup for under $800. But the next time I have $800 to spend, would I rather buy that and no cards, or some new cards and no new computer?

[Sn1ken]

Don't get me wrong.

As you are using 1150 or 1155 you are limited by 16 PCIe lanes to your CPU.

That equals two GPUs for one motherboard. If not the PCIe will drop along with the PPD. Trust me. I only run two GPUs on each motherboard. And the reason is simple. PPD pr $
The 2011 is not just expensive, but the heat from 4 GPUs on the same motherboard is also a pain in the neck.

Hope u see my point. Just get rid of one card. Skip the 750 ti and your PPD will rise.

[Kjetil]

On 1150 or 1155 2x 970, 980 is 8x,8x.

[runpaint]

And it's x8 & x8 on a 2011 if the cpu doesn't have 32 pci lanes. But pcie3.0 x8 is as fast as 2.0 x16.

But hopefully you're right about removing one card, I had the same thought & took out the 980 yesterday & moved the 750ti from the x4 slot on this fx990 motherboard (FX9590 8 cores @ 5.0GHz, my most expensive processor and getting the least ppd out of the same cards. Most of my 750ti are MSI, and the only one getting 70k ppd is on an H81 with an i3).

So I have a 970 and a 750ti running together, but the points are still low for both. And that still doesn't explain why the 970 and 980 were getting 250k and 350k ppd when I first installed them, before I changed the driver version.