Cooler weather soon, what are your goals for your GPU(s)?

[TennesseeTony]

With the Fall and Winter getting closer, have you given thought to when you will kick off GPU computing again? What about your choice of projects?

As for me, I will probably dabble with a few projects in the fall, with limited GPUs in use, but my real kick off will be for the annual Folding@Home race in December. I currently have about 100M there, and my plan is to continue Folding after the race until I reach 250M sometime in January.

Then I will switch to POEM and GPUGrid for 4-5 weeks each, with a goal of doubling my current scores, bringing them up to 500M in each project. That may be enough to take me into late March, which, here, means warmer weather.

Time and cold temperatures permitting, Milkyway might get some more lovin' from me this winter, but not likely.

 

[TennesseeTony]

Ken, PPS-Sieve race coming up in December during the Folding race...is that the same sieve as last year? With GPU? If so, I guess I can add to my list above, a goal of 25M more for PrimeGrid.

 

[Markfw]

I will turn on a few more for the December race... Close to 3 million PPD now.

And up to number 63 worldwide for F@H. Hope to get a few more places before then.

Here http://fah-web2.stanford.edu/cgi-bin/main.py?qtype=userpage&username=Mark_F_Williams

 

[Ken g6]

Ken, PPS-Sieve race coming up in December during the Folding race...is that the same sieve as last year? With GPU? If so, I guess I can add to my list above, a goal of 25M more for PrimeGrid.

Yes, it's the same race.

I'm debating whether or not to get a new GPU for myself. Outside of GPU DC my i3 is doing very well so far all by itself.

 

[TennesseeTony]

...up to number 63 worldwide for F@H...

Out of 1,845,131 people!!! Very impressive!

...I'm debating whether or not to get a new GPU for myself...

Don't do it. Too addictive.

 

[TennesseeTony]

OrangeKid, you going for 50,000,000 in POEM?

 

[Orange Kid]

Don't think I'll get to 50M. The December F@H and PrimeGrid race will cut into that. I think 25M+ is doable though. I'll be down most of Jan and Feb due to a extended absence so don't look for much from me then.

 

[Pokey]

Luckily, I have a basement plus I can afford the power bill..... So I have kept my GPUs going.

I hit the 1 Bil milestone in Folding and decided to get off it for a while, and have switched over to seti for a bit.

My goal now is to get back to my seti world best before the Dec Folding event rolls around. But seti has a lot of moving parts these days due to the new Green Bank work units being sent out and the different apps being developed, so credit is hard to come by with a lot of ups and downs.

Anyway that's my story.................

 

[Markfw]

OOPS, I did it again......

Brown truck on the way....

And up to number 60 world-wide and 3.3 million ppd now (not for long)

 

[Orange Kid]

Hey Tony.
So I've been looking at the times to do WU's for POEM on your machines versus the times for my 970 and all of yours are taking longer than mine. So the question is are you running more than one WU at a time per GPU? It doesn't make sense to me why 980's 980ti's and 1070's would be running at a 20% to 30% deficit to a 970. No if you are running multiple WU's per GPU then that would probably make the difference. I'm trying to decide if a 1080 is worth the extra $$ or would a 1070 be better over all. By the looks of it I should get another 970

 

[Markfw]

BTW, aside from power usage, the 980TI is the best bang/buck for F@H, and I would bet for other DC also. My 1080 SOMETIMES will beat my 980TI, but most of the time they are equal.

I just got another 980TI factory OC'ed for $379, the cheapest 1070's are $430, and as I say the 1080 is the equal of the 980TI

Here is a pic of my farm. The descriptions tell you what they are, and this is an average ppd pic.

 

[TennesseeTony]

Orange Kid, I'm running either 2 or 3 POEM tasks per GPU, depending on the machine (I'm too lazy to edit the app_config or they'd all be running 3 per GPU). Dual tasks bump up the GPU utilization to almost 90%, the third task bumps utilization to mid 90's. POEM needs one CPU thread/core per task also.

I've been prepping my systems for GPU season, putting GPU's back into service that I should have sold a long time ago. I'm a bit overstocked, but I think I've found homes for all the Nvidia cards.

That being said, the 1070 is now down to $410. Preliminary tests (meaning only about 6 hours of run time) in Folding@Home is showing 1.1M PPD for a pair, which is on par for what a pair of my 980Ti's seem to put out in the past, and also in recent testing. (Mild factory overclocking only, on all my cards, hitting voltage limits on all, with only high 70s % GPU Utilization on the 980Ti's at about 1400MHz and near 100% usage at about 1750MHz for the 1070's)

Again, neither the 1070's nor the 980Ti's have been run more than a few hours (in recent testing), so that really doesn't count for much. Folding definitely needs one CPU thread/core per task as well. Mark, can you run multiple tasks on one GPU in Folding?

Edit:

Here's a sample app_config.xml for POEM (.333 for triple tasks)(.5 for double):

<app_config>
<app>
<name>poemcl</name>
<user_friendly_name>POEM++ OpenCL</user_friendly_name>
<gpu_versions>
<gpu_usage>.333</gpu_usage>
<cpu_usage>.333</cpu_usage>
</gpu_versions>
</app>
</app_config>

 

[Markfw]

Nope. One task per GPU, and one per TOTAL cpu's. So for 24 cores, and 2 video cards, I get 3 clients. See the picture os dual 5639's. Thats 24 cores and 2 video cars for 3 clients.

If you can do it otherwise, I have no idea. And you can see my idle CPU client that made my PPD skyrockets on my E5570 Xeon. (it says P3500, but could be T3500, not cure the model number of the MB.)

 

[Orange Kid]

Thanks Tony,the times make sense then.
Mark,ya the 980ti's are cheaper upfront but I looking long term and think that a 1070 or 1080 would do good and the power and heat would be less allowing for more run time without the wife wondering why it is always so warm in the family room

 

[Markfw]

Well, the power usage may be less, but I can't tell that the 1080 is putting out less heat, its a toaster !

 

[TennesseeTony]

Mark, are you running the sample 980Ti and the sample 1080 using equivalent CPUs? I'm wondering if, since the tasks are 100% of the time using a CPU core, if your scores are being affected by the CPU. I know my FAHBench scores certainly changed when placing the 980Ti in a faster system.

 

[Markfw]

The 980TI's are using E5570's, 4790K, and 3930K, but the 1080 is using the 6700k skylake, the best CPU, and still can't best the 980TI's most of the time.

 

[TennesseeTony]

Hmm. There goes that theory.

Regarding the heat output, I do see my 980Ti's running at about 80% TDP (stock), versus 98% TDP (stock) on the 1070's. So that's roughly 200 watts versus 150 watts, not nearly the gap of a full 100 watts I'd expected, looking at the specs. For the 1080, if it too is near 100% TDP, that's only 180w vs 200w, so yeah, not much difference at all.

At any rate, especially with some projects not yet supporting the 10 series, the 980Ti is an incredible performer for the price, no doubt!

PS: One last thought (for the moment), slide a Ti beside that lonely 1080, the driver (or the download at least) is the same. Could be something system specific? Not very likely though...

 

[Markfw]

The driver may be the same, but I am told to only use the exact same chipset/model in a dual setup. I have 2 980's one is an SC and one a FTW, but within a few mhz of each otther stock.

BTW, just to vent. So I saw the "4996" EVGA 980TI card for $379, and jumped on it. Then it went up to 489, then down to 439, then not available.(the best 980TI I saw)

So then the "4995" EVGA 980TI card went down to $379 and I snatched one. Now its up to $429 Is 100 mhz slower at "stock" than the 4996, but close.

What the heck is up with these pricing models ?? All refurbs BTW, but mine looked new. I will see on the second one.

 

[TennesseeTony]

The price fluctuations are strange indeed.

Of course no sooner than I mention my 1070's being nearly tapped out at 98% TDP, I notice that the next set of workunits has arrived, and the performance is different, and making me look like a fibber. On this set of workunits the TDP is nearer to 80%.

I do have one card though, that I think is begging me to free it from its artificial limits imposed by its maker. It's running faster, and it's hitting not only the voltage limit, but cycles to the power limit and also the utilization limit, AND it's using 50% of the bus (does that mean PCIe bandwidth?) while the other card is only using 32%.

But to be fair, if it's the secondary card (not connected to the monitor), it may be running at 8x, so the bus usage would appear higher, as a percentage, although not actually pushing more data.

EDIT: The Kill-A-Watt shows an even 400 watts at the moment, was 415w last night, with a max that I have seen of 420w, on a Seasonic 850w Gold (should be in the 90% efficient zone at this power level). Not bad for 2 980Ti equivalent cards and a 12 thread box.

 

[Markfw]

Just so you don;t think I am not watching, every once in a while this happens to my 1080 (but its rare) 901k ppd, and ONCE I saw over a million.

And here is my dual 5639' Xeon's with dual 980's

 

[TennesseeTony]

...Drools...

I can't wait, well, I can, but I don't want to, but I can hardly wait to try out the Precision tool thingy from EVGA. But for now I'm letting this system run a week, then I'll switch over to a different system for a week, and then another, seeing how each perform with their slightly different sets of GPU's and CPU's.

I find folding at home is hard to judge performance on, with all the variable types of work units, but surely a week will give a solid hint as to what to expect from each station?

 

[Markfw]

Yes, a week each. Please share your results. Remember, you have to watch every work unit on a box for at least a day, but I think I know the general performance. Below is my estimated summary:

CPU. Not worth it for less than 24 cores, and then only of 50k works for you

GPU: First, you can only have 2 cores less than max for every card, but for socket 1366 or older, don't even turn them on. You will loose way more than you gain by enabling it. My 6700k only gets 22k ! Now for the cards I have experience with, than range, and the average

780TI been a while so not sure on this, but I think they were 250k ppd

980 200k-500k, usually 400k average
980TI 300k - 650k (once 800k), but average 550-600k
1080 same as 980TI, but I saw one million ppd once.

And I think my picture above is a real example of averages, since my 7 day ppd avg is 3.3 mil

 

[Orange Kid]

Thanks for all the info here. (total confused as what to buy now)

 

[Markfw]

If you are for points/ppd/price, a refurb 980TI is hard to beat. If you are going for longer term, and want to save some electric bills 1070 might be your buddy Until I have one I can;t say.