SLI GTX580 -- Smallest / Lightest Possible Build

[lambchops511]

I want to build 2x GTX 580 (not 680) w. the smallest / lightest possible build. Weight is more important than size (i.e. constant transportation).

I would prefer dual 2x16PCIE if possible, but if the cost is too much, 2x8PCIE is acceptable too.

I will not be using any 3.5" drives, but 2x2.5" drives.

No DVD / 5.25" drives.

I place a high value on stability (i.e. I like Intel branded motherboards), but it is my understanding SLI does not work on Intel branded motherboards?

There will be no OC.

Any recommendation for Case + PSU + Motherboard.

 

[T_Yamamoto]

Smallest and lightest would be something like a matx in a vulcan

 

[ElFenix]

i dunno if i'd be comfortable sticking 2x580s in an matx case. those things need room to breathe.

if the length will fit a lian li pc-a05fn is very compact for an atx case and light weight.

 

[Zap]

Lian Li PC-A04B

one possible motherboard

Be careful choosing mATX boards. Some of them such as some Gigabyte boards have strange layouts such as the PCIe slots for SLI being 1 and 4 instead of 1 and 3.

 

[lambchops511]

Lian Li PC-A04B

one possible motherboard

Be careful choosing mATX boards. Some of them such as some Gigabyte boards have strange layouts such as the PCIe slots for SLI being 1 and 4 instead of 1 and 3.

wouldn't 1-4 be better? more spacing? or would it not fit w. case?

 

[Ken g6]

Oh! I know! If you insist on two GTX 580s in SLI, what if...you got them both on one card? :sneaky:

Now, techincially, that card (not its box) is 11 inches long, so it should fit in a BitFenix Prodigy! (Or similar mini-ITX case.)

 

[piasabird]

Sounds to me like you dont really want a small case. If you want to use two video cards how small can you go and still protect the video cards and provide cooling? Plus if it is going to be portable and moved around a lot it needs to be strong. I dont think you can get all these things and still be small and light.

Lian LI makes a case like this for $379 with one5.25 drive and a full ATX with lots of fans.

http://www.newegg.com/Product/Produc...82E16811112380

I dont know about being small and light.

This is a little smaller and less expensive:

http://www.newegg.com/Product/Product.aspx?Item=N82E16811112376

 

[lambchops511]

Oh! I know! If you insist on two GTX 580s in SLI, what if...you got them both on one card? :sneaky:

Now, techincially, that card (not its box) is 11 inches long, so it should fit in a BitFenix Prodigy! (Or similar mini-ITX case.)

I need the SLI for compute functionality testing (actual number crunching will be done w. some server cluster, but I want something local which I can dev on).

Is there any difference between 2xGTX580 and 1xGTX590? The clocks seems substantially lower? no?

 

[Ken g6]

Yes, the clocks are lower on the 590, but that's the only difference. It has the full shader count on each core.

The reason for the difference is power and cooling. You can OC a 590 to match the 580s, if you can keep it cool enough and give it enough power. Installing all possible fans, exhausting, on the BitFenix Prodigy case might do it, as the card's air intake comes directly from the outside.

But if you're doing compute, why wouldn't two GT520s (at the extreme minimum) be sufficient to verify multi-card functionality? You need GF100-style Fermi cores? GDDR5 RAM? If not, then I might suggest a refurbished laptop from eBay.

If you insist on two 580s (not 570s? Not 560ti 448s?) then two of these, or one of those and one of these, might help fit two cards in a micro-ATX case while staying cool.

 

[mfenn]

I need the SLI for compute functionality testing (actual number crunching will be done w. some server cluster, but I want something local which I can dev on).

Is there any difference between 2xGTX580 and 1xGTX590? The clocks seems substantially lower? no?

Don't forget that double precision is intentionally crippled on GeForce cards. So just forget about doing any kind of double-precision benchmarking or profiling on your local machine (unless you're willing to shell out for a Tesla C series).

Also, what kind of Teslas does your cluster use, M2070/M2075 or M2090 (I hope to God it's not the M2050)?

Really, I can see a few different scenarios:
- You need double-precision math. In that case, it doesn't really matter what GPU you buy, just get something that's a Fermi or Kepler and cheap because you're not going to be able to draw any conclusions about performance locally.
- You need single-precision math and your cluster uses M2070 GPUs. Get the GTX 560 Ti 448 because both have 448 CUDA cores.
- You need single-precision math and your cluster uses M2090 GPUs. Get the GTX 580 because both have 512 CUDA cores.

Obviously the case and PSU depend on the GPU that you end up picking.

 

[lambchops511]

Don't forget that double precision is intentionally crippled on GeForce cards. So just forget about doing any kind of double-precision benchmarking or profiling on your local machine (unless you're willing to shell out for a Tesla C series).

Also, what kind of Teslas does your cluster use, M2070/M2075 or M2090 (I hope to God it's not the M2050)?

Really, I can see a few different scenarios:
- You need double-precision math. In that case, it doesn't really matter what GPU you buy, just get something that's a Fermi or Kepler and cheap because you're not going to be able to draw any conclusions about performance locally.
- You need single-precision math and your cluster uses M2070 GPUs. Get the GTX 560 Ti 448 because both have 448 CUDA cores.
- You need single-precision math and your cluster uses M2090 GPUs. Get the GTX 580 because both have 512 CUDA cores.

Obviously the case and PSU depend on the GPU that you end up picking.

What's wrong w. M2050? We are a mixture of C2070 and GeForce cards (i.e. hybrid cluster). We only care about SP. So you are saying the GTX 580 don't warrant over the GTX 560? Development time is valuable as well.

 

[Ken g6]

Do you need this dev box to match your cluster performance as closely as possible per GPU? If so, get 2 560 Ti 448s (not just any 560 or 560Ti) and call that part settled.

If per-GPU performance doesn't have to equal your cluster, but per-CUDA core performance must match (proportional to clock speed), then the GTX 590 could be an option as well.

If performance doesn't have to match that closely, you have other (GF 104/114) options, which depend on your budget. (What is that, by the way?) That laptop I linked to above is an expensive option. Here is a relatively cheap dual GTX-460 card to go in a mini-ITX case. Since you only care about SPFP, GK104 should work as well; a GTX 690 would be mid-range in price, and fit in a mini-ITX case, but would be the fastest option so far.

 

[mfenn]

What's wrong w. M2050? We are a mixture of C2070 and GeForce cards (i.e. hybrid cluster). We only care about SP. So you are saying the GTX 580 don't warrant over the GTX 560? Development time is valuable as well.

The M2050 (C series are the same thing but with blower coolers) has only 3GB of RAM, which is extremely limiting in most computation environments that I've been exposed to.

You only care about single-precision, that is good news from a cost perspective. The next question is if you will be doing performance work on your local machine or if you will only be doing functionality testing.

If you're going to be doing performance work, you need to match your local card as best you can to the shader configuration of the cluster, which means getting a GTX 560 Ti 448. If you go out and buy GTX 580s and optimize the heck out of your code for 512 cores, it's not going to translate well to real production performance.

If you're only testing for functionality, get the cheapest Fermi or Kepler card with enough VRAM for your problem size. In this case, you would be testing with small enough data sets (or iteration count depending on your problem) that the actual time spend running vs. coding should be minimal.

 

[mfenn]

Do you need this dev box to match your cluster performance as closely as possible per GPU? If so, get 2 560 Ti 448s (not just any 560 or 560Ti) and call that part settled.

If per-GPU performance doesn't have to equal your cluster, but per-CUDA core performance must match (proportional to clock speed), then the GTX 590 could be an option as well.

If performance doesn't have to match that closely, you have other (GF 104/114) options, which depend on your budget. (What is that, by the way?) That laptop I linked to above is an expensive option. Here is a relatively cheap dual GTX-460 card to go in a mini-ITX case. Since you only care about SPFP, GK104 should work as well; a GTX 690 would be mid-range in price, and fit in a mini-ITX case, but would be the fastest option so far.

:thumbsup: Well said.