rpi_logo
Which GPU?
Which GPU?
log in

Advanced search

Message boards : Number crunching : Which GPU?

1 · 2 · Next
Author Message
Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67456 - Posted: 16 May 2018, 1:10:11 UTC

I prefer AMD GPUs for price/performance ratio. But I can't find any information on which models are best for double precision. All the benchmarks just quote certain games, or single precision. And AFAIK Milkyway uses double precision. What do other projects use? Does anyone know where I can look at a list of currently available AMD GPUs to compare double precision?

All I have at the moment is an old R9 290, and a brand new RX 560. The 560 should be half the speed of the 290 (going by the single FP speed on reviews), but it's running at about 1/4 of the speed for Milkyway, so I assume they scrimped on the double precision. Anyone running an RX 580 (which I plan to get for gaming)?

Jim1348
Send message
Joined: 9 Jul 17
Posts: 45
Credit: 2,315,069
RAC: 12,963

Message 67457 - Posted: 16 May 2018, 1:19:22 UTC - in response to Message 67456.

These give the performance figures:
https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units

I don't see any particularly good buys though.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67458 - Posted: 16 May 2018, 1:56:45 UTC - in response to Message 67456.

Nothing seems to have much double precision anymore. Is this because games don't use it? What about bitcoin mining? Cards seem to be getting made for that nowadays. My 5 year old R9 290 outperforms everything at double!

And do you know if other BOINC projects use single or double? I think SETI is single, as it's working on my very old Nvidea card that will only run that and Prime.

Jim1348
Send message
Joined: 9 Jul 17
Posts: 45
Credit: 2,315,069
RAC: 12,963

Message 67459 - Posted: 16 May 2018, 4:37:28 UTC - in response to Message 67458.

The R9 290 was one of the better ones. Insofar as I know, nothing else uses it. I am not a gamer, but if they needed it, cards would be great at it; same for bitcoin. Only the professionals use it, so they pay for it.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67460 - Posted: 16 May 2018, 12:50:28 UTC - in response to Message 67459.
Last modified: 16 May 2018, 12:51:32 UTC

I got the 290 as a gamer recommended it to me years ago, and I build gaming PCs as a small business. I then used it and several other 290s back when altcoin mining was vaguely profitable.

So most games only use single precision? Also Einstein and SETI etc?

Compared with CPUs, graphics cards don't seem to be improving so well over the last few years.

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67461 - Posted: 16 May 2018, 12:51:32 UTC

Hi, I have an entire spreadsheet with info. I can't link it at work, google docs is blocked :( I'll reply tonight from home.

It depends on how hands-on you like to be, if you have no problem rigging up your own cooling solution, the FirePro (S91xx) server cards are fantastic for FP64 and a *very* good value for cost/performance ratio. The server cards are passive, heat-sink only, you need to figure out a way to actively cool them. I use 40mm fans, but these are VERY LOUD.

There are Workstation versions (W9100 etc) of the card that have built in fans like a normal GPU, but they tend to be more expensive. Still offer the best raw performance however.

Just as a point of reference, depending on the exact model, the S9100 and S9150 offer somewhere in the neighborhood of ~2500 GFLOPS of FP64 performance.

Stepping down to the consumer Radeon cards, you're best bet is still the 280X. Depending on the exact model, clockspeeds etc, they offer, roughly, around 1000 GFLOPS.

Similar cards, (AMD does a lot of re-branding, with the same/slightly tweaked hardware) such as the preceding 7970 offer roughly the same performance. 7950 slightly below that etc.

Keep in mind the xx90 (6690, 7790) cards are dual-GPU on single card, and offer slightly better performance-per-watts than running two separate cards. I can't recall the exact GFLOPS for these off the top of my head (though easy enough to look up on wiki). I personally run a 6990 and am pretty happy with the performance, but that was before I discovered the S9150.

The 290 is 'ok' but not great with FP64, less performance than 280x. All the newer 3xx/4xx/5xx series simply do not offer strong FP64. Better single precision yes (that's all games need) but even Vega is not particularly strong in double precision.

By far the best bang for the buck is the FirePro server cards if you can deal with the cooling and resulting noise. or the regular 280X for an easy drop-in solution.

Have fun! :D

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67462 - Posted: 16 May 2018, 13:00:38 UTC - in response to Message 67461.

I had a 7990 (I think) - a dual AMD card for altcoin mining. The only way to get it to run at full speed was to actually replace the 12 inch side fan on the computer case with a car radiator fan (an 80 watt fan that made a racket). It's simply impossible to cool two GPUs on one card sensibly!

Firepro 9000 series are a few grand aren't they? (in GBP £) - I think I'd rather get three of 2nd hand 290.

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67463 - Posted: 16 May 2018, 13:04:58 UTC

Rough numbers, FP64 performance. will vary with clockspeeds etc

290: 606 gflops
560: 163 gflops
580: 385 gflops

280x: 1024 gflops
6990: 1276 gflops
7990: 1894 gflops

S9150: 2530 gflops

Keep in mind this is only half the story. Other important question is the energy (watts) required for this performance. My spreadsheet goes into all of this and makes calculations based on TDP of gflops-per-watt and other considerations (including purchase price, for $cost-per-gflops)

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67464 - Posted: 16 May 2018, 13:08:08 UTC - in response to Message 67463.

I bought my S9150 for $340 USD on ebay. Currently see other S9100 cards for $300-400 USD.

W9100 card (basically the same card, with built-in cooling fan) is $800-1000 or more. Crazy. lol.

So if you can rig up the cooling, the price/performance just blows everything else out of the water. Otherwise, 280X all the way.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67465 - Posted: 16 May 2018, 13:11:16 UTC - in response to Message 67464.
Last modified: 16 May 2018, 13:11:45 UTC

If I was only doing double precision I might consider it, but AFAIK most projects use single. I also use the cards for playing games.

Oh and I hate noise!

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67466 - Posted: 16 May 2018, 13:22:12 UTC - in response to Message 67465.
Last modified: 16 May 2018, 13:24:30 UTC

Ahh, well Milkyway is the only project I crunch so I bought the cards specifically for it. 580 should be good for games then. :)

Profile mikey
Avatar
Send message
Joined: 8 May 09
Posts: 2212
Credit: 250,022,407
RAC: 131

Message 67468 - Posted: 17 May 2018, 10:49:32 UTC - in response to Message 67456.
Last modified: 17 May 2018, 10:50:42 UTC

I prefer AMD GPUs for price/performance ratio. But I can't find any information on which models are best for double precision. All the benchmarks just quote certain games, or single precision. And AFAIK Milkyway uses double precision. What do other projects use? Does anyone know where I can look at a list of currently available AMD GPUs to compare double precision?

All I have at the moment is an old R9 290, and a brand new RX 560. The 560 should be half the speed of the 290 (going by the single FP speed on reviews), but it's running at about 1/4 of the speed for Milkyway, so I assume they scrimped on the double precision. Anyone running an RX 580 (which I plan to get for gaming)?


Try here too:
http://www.geeks3d.com/20140305/amd-radeon-and-nvidia-geforce-fp32-fp64-gflops-table-computing/

GeForce GTX 580 1581 197 FP64 = 1/8 FP32

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67469 - Posted: 17 May 2018, 13:13:59 UTC

https://docs.google.com/spreadsheets/d/1ImSDoLeuZFvmO6xoMpy2VMs9Du_6sHT5GqcORMtx2tQ

That link should work for my Spreadsheet. Still a work in progress and there's some missing data and some extra fluff like the "fake rank" column where I decided to try and assign a very-rough estimate of overall general performance based on Cores*Clockspeed.

The main focus is showing 3 main variables (GFLOPS, Price, Watts) and how they interact.

Unfortunately due to locking/merging certain cells I cannot dynamically sort the columns how I would like. Still determining if there is a way I can do that.

Profile BeemerBiker
Avatar
Send message
Joined: 18 Nov 08
Posts: 95
Credit: 753,495,459
RAC: 1,300,589

Message 67470 - Posted: 17 May 2018, 14:29:39 UTC - in response to Message 67469.
Last modified: 17 May 2018, 14:32:54 UTC

https://docs.google.com/spreadsheets/d/1ImSDoLeuZFvmO6xoMpy2VMs9Du_6sHT5GqcORMtx2tQ

That link should work for my Spreadsheet. Still a work in progress and there's some missing data and some extra fluff like the "fake rank" column where I decided to try and assign a very-rough estimate of overall general performance based on Cores*Clockspeed.

The main focus is showing 3 main variables (GFLOPS, Price, Watts) and how they interact.

Unfortunately due to locking/merging certain cells I cannot dynamically sort the columns how I would like. Still determining if there is a way I can do that.


Nice spreadsheet, thanks!

Want to mention that the S9000 is listed at "<225" at WiKi, and I suspect it is no more than 200 and may be closer to 150 than 225. It has a single 8 pin PCIe which has 3 12v power leads unlike the "200w" 7950 which has 4 (two 6pin ) or 5 (6pin+8pin). I lost a pair of 7950 due to overheat and the remaining I converted to liquid cooling. When I got my first S9000 (new, unused for $150 but passive cooling), I discovered the left over HD7950 fan cooling fit perfectly on the S9000. Depending on the OEM, the cooling assembly also cooled the memory chips on the GPU side of the board. MSI and PowerColor heat sinks serviced only the GPU chip. I now have 3 of the S9000 and they all run cool. I had to use a dremmel to cut a small amount of plastic from the shroud exhaust end on a gigabyte cooling assembly so one of the S9000 would fit in case. Anyone with a dead HD7950 but working cooling assembly can essentially get a brand new, unused replacement for under $160. Crossfire is not supported and a 3pin molex must be forced into the 4 pin cooling cable to make the fans run. Since I run 24/7 this as fine for me.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67471 - Posted: 17 May 2018, 14:39:33 UTC - in response to Message 67470.

Don't GPUs have thermal protection like CPUs? Even if a CPU fan fails completely, the CPU just slows down. I actually had a guy bring me a computer he said "quickly slowed down once I started using it", I discovered the heatsink had actually come away from the CPU! But no damage done, it just got slower to maintain a sensible die temperature.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67472 - Posted: 17 May 2018, 14:53:20 UTC

Can't Milkyway use Single Precision instead? From what I've read, a double precision calculation can be emulated using two single precision calculations. This would speed things up on almost every GPU out there.

Profile BeemerBiker
Avatar
Send message
Joined: 18 Nov 08
Posts: 95
Credit: 753,495,459
RAC: 1,300,589

Message 67473 - Posted: 17 May 2018, 15:37:19 UTC - in response to Message 67472.

Can't Milkyway use Single Precision instead? From what I've read, a double precision calculation can be emulated using two single precision calculations. This would speed things up on almost every GPU out there.


Single precision is accurate to 7 digits and double to 15 as illustrated here

If your accuracy requires more than 7 digits of precison you use double precision: either in hardware or using a library that emulates double precision.

Possibly, some optimization could be done to speed things up as SETI lunatics did for setiathome.

I had to program in CMS-2M on navy systems that did not support floating point hardware. It was all done using scaled arithmetic, binary angles, and trig lookup tables. Floating point hardware would have made a huge difference in cost, especially labor.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67474 - Posted: 17 May 2018, 15:55:51 UTC - in response to Message 67473.

It's the emulation I was considering. I read somewhere something like "for some reason Milkyway doesn't use well known algorithms to run on single precision hardware, but Einstein and SETI do". As far as I know the emulator would use two single precision calculations to get a double precision one done, but single precision on most graphics cards is well more than twice as fast.

melk
Send message
Joined: 10 Dec 17
Posts: 47
Credit: 629,888,759
RAC: 295,367

Message 67475 - Posted: 17 May 2018, 18:59:52 UTC - in response to Message 67470.

Want to mention that the S9000 is listed at "<225" at WiKi, and I suspect it is no more than 200 and may be closer to 150 than 225.


I suspect that could very well be. Wiki isn't necessarily 100% accurate itself anyhow. Really just a rough estimate at best.

Peter Hucker
Send message
Joined: 5 Jul 11
Posts: 138
Credit: 27,075,696
RAC: 0

Message 67482 - Posted: 18 May 2018, 1:40:29 UTC - in response to Message 67475.

I notice graphics card manufactures are designing cards to work well with coin mining, so why don't they design them to work well with boinc? We want double precision! And we want it now!

1 · 2 · Next
Post to thread

Message boards : Number crunching : Which GPU?


Main page · Your account · Message boards


Copyright © 2019 AstroInformatics Group