Intro and Question... DC system build...

[LightningZ71]

Hello all!

Many moons ago, I was a SETI@Home cruncher, following that, I was a Folder and also involved in a few other projects... Then I got married and had kids...

Now, I'm beginning to find myself with a little extra time and low energy costs. I want to get back into distributed computing again. One thing that I do NOT have a lot of spare of is money.

I've currently got:
1 X AMD Phenom II X6 1100t
1 X Nvidia geforce GTX 550ti
8GB Ram
Sparky Linux Game Over edition

1 X Dell Optiplex 780 Small Desktop
Intel Core2Quad Q8400
8GB Ram
Win 10 64bit

Potentially 2 X HP SFF 4GB Windows 7 32 bit boxes (can be converted to linux) Core2Quads (don't know the vintage yet)

A whole bunch of Nvidia NVS 510 (Kepler GK107, 2GB 128 bit DDR3) Low Profile Video Cards (with full height brackets)

Is ANY of this stuff even worth it to get into any of the current DC projects with? I'm considering possibly getting an old DL370 G6 (available used for under $300, that's the limit of any budget that I have) to put a bunch of those NVS 510s in (it has 8 PCIe slots, though electrically, they are x8 and x4) to all crunch away at once. Does that even make any sense, or would I be better served by just investing in a pair of ~$150 video cards for the AMD system instead?

Any advice would be appreciated.

 

[Kiska]

What you currently got is good for SETI
Like I have a used HP Proliant DL360 G5 that is doing seti currently
Just be mindful of power consumption

 

[Smoke]

Hello and welcome, LightningZ71

You have a lot folks here that can relate to your "cycle of life" story. Glad you joined us.

 

[StefanR5R]

A whole bunch of Nvidia NVS 510 (Kepler GK107, 2GB 128 bit DDR3) Low Profile Video Cards (with full height brackets)

https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units says that Quadro NVS 510 consists of 1 Kepler SMX. I have got a laptop with Quadro K2100M, which has 3 Kepler SMXs. I tried this one in the double-precision project Milkyway@Home, and it didn't fare well: https://forums.anandtech.com/thread...on-different-wu-gpu-cpu-times-wanted.2495905/

I don't think I tried any of the more common single-precision projects on it yet...

...OK, PrimeGrid PPS-Sieve for example takes 38 minutes per task on the K2100M. To put this into perspective, a GTX 1070 takes 6 minutes, and a Broadwell-E 11 hours: https://forums.anandtech.com/threads/primegrid-llr-races-thread-2016.2466149/page-10#post-38636580

 

[Kiska]

And a Geforce 940MX takes about 3000 seconds

The 940MX also takes about 1 hr to do a seti task

 

[StefanR5R]

Single precision throughput, from wikipedia:
GTX 550 Ti....................691 GFLOPS
Quadro K2100M..........753 GFLOPS
Quadro NVS 510..........throughput not listed, but probably at the ballpark of 1/3 of K2100M
GeForce 940M..............823 GFLOPS
GeForce 940MX............card not listed

(940MX/DDR3 is ~5 % above 940 M, 940MX/GDDR5 is ~20...30 % above 940M according to notebookcheck.net)

 

[LightningZ71]

First, thank you for the welcome guys (and any gals that may be present)!

I've been looking around at this site and others for information on the performance of the nvs 510, and it's retail twin, the GT 630 2GB DDR3, and it looks fairly grim. There isn't much out there to begin with on either card, and precious little on the 510 at all. As I'm practically swimming in the things (did a contract to retire and wipe a bunch of old PCS that all had quad monitors), I was mainly hoping to find some sort of project that they might be OK for. Their advantages, other than being free, are that they have very low power draw (needing only bus power) and each have 2GB of video ram. From the few I've tinkered with, they appear to be late production cards and have a surprising amount of overclocking headroom.

It just seems a shame to not find something constructive for them to do. I do think that I'll go the buy two better cards route instead of investing in the old server, but I can still stick them in the dell and the 2 HPs to do... something.

 

[LightningZ71]

Single precision throughput, from wikipedia:
GTX 550 Ti....................691 GFLOPS
Quadro K2100M..........753 GFLOPS
Quadro NVS 510..........throughput not listed, but probably at the ballpark of 1/3 of K2100M
GeForce 940M..............823 GFLOPS
GeForce 940MX............card not listed

(940MX/DDR3 is ~5 % above 940 M, 940MX/GDDR5 is ~20...30 % above 940M according to notebookcheck.net)

The nvs 510 is largely based on the GK107 gt 630 128 bit ddr3. That card has a listed throughput of 336 single precision and 14 double. Now, the base clocks on the nvs 510 are lower, but as I've found, overclocking to 1ghz are easily done on the batch that I have. I had heard that Nvidia made some tweaks to the quad and nvs drivers that affect their performance in some situations, does anyone know how that affects any of the DC projects?

 

[StefanR5R]

The nvs 510 is largely based on the GK107 gt 630 128 bit ddr3. That card has a listed throughput of 336 single precision and 14 double. Now, the base clocks on the nvs 510 are lower, but as I've found, overclocking to 1ghz are easily done on the batch that I have.

So, 336 GFLOPS * 1000 MHz / 875 MHz = 384 GFLOPS = 1/2 K2100M.

I checked SETI@home v8 task times now:
K2100M..........00:59:15
GTX 1080........00:06:20

I had heard that Nvidia made some tweaks to the quad and nvs drivers that affect their performance in some situations, does anyone know how that affects any of the DC projects?

Not sure about that, but I am fairly new to GPU computing in general and DC in particular. So far my Quadro K2100M and a FirePro W7000 performed in DC applications like similarly spec'ed consumer cards.

 

[TennesseeTony]

Welcome back!

From the standpoint of energy efficiency combined with output, I'd say to ebay everything not currently in use and put it toward a 10-series GTX (1070's get a price cut yet?), depending on the projects you're interested in running. You may only net $10-25 from each card, but based on the quantity you have, combined with your budget, you should easily be able to snag a $350 1070, which will outperform roughly 15-18 of the NVS 510's, all for less than 150watts.

If you are interested in running Milkyway, Einstein, and (coming "soon") Universe, you'll want a card with high double precision, which really only leaves you one choice, ancient AMD 280X/7970 cards off ebay.

That said, the majority of the projects still don't have/can't use GPU apps, but for those that do have GPU apps, it is often a major leap forward in productivity compared to the CPU app. In Folding, for example, the GTX1070 can earn you 550,000-700,000ppd, depending on the CPU it's paired with, compared to maybe 80,000-120,000ppd for a newer 14 core/28 thread Intel Xeon.

My "get a brand new, current generation, mid-high performance GPU" recommendation will also save you a lot of headaches from trying to manage more complex setups.

 

[StefanR5R]

Folding@Home is a bit special, as you get points not just for the amount of work done in one WU, but also for how fast the WU was completed. This means one big fast node gets a lot more points than two half as fast nodes in F@H, contrary to most DC projects.

 

[LightningZ71]

Well, I think that that was pretty definitive. I'll start unloading these cards and try and scrape together enough dollars for a top tier card.

One other question, is it worth it to swap out the Phenom II x6 for a Vishera FX-8300 series processor? I know it will drop the energy usage a bit, but will it make a significant improvement in project throughput. I know that the FPU changes (for the worse) when you get to the Bulldozer/Piledriver cores, but there are other improvements that can help things too.