Advise on risers?

[pegot]

I just got a GTX 780 6GB card (EVGA).

I'll pop it into my stem with a pre existing GTX 770 (and also try to fit my GTX 570) to see what the temps are like, but I have a feeling I may need to consider some cooling options.

I have a Half-X Full tower case with two 120mm fans on top and also the case's large side fan.

Before I consider liquid cooling (I am currently looking at some XSPC kit systems) could anyone offer advise on using a riser to reposition at least one of my cards so it is not directly on top of another card? Are there any particular brands or types of risers I should look at - NewEgg has a range with prices from $6 - $40. Do I need a powered riser?

Any risers in the below list from NewEgg good and safe to use? Or does any one have suggestions from other vendors / outlets other than NewEgg?

http://www.newegg.com/Product/ProductLi ... d=1&Page=2

[acc24ex]

I would recommend powered risers - the normal one worked OK with a 580gtx, but with 3 590 gtx it doesn't - also 590gtx don't work at all without powered risers /
- I got a few of them over ebay - they work great

those ones: tested and working on 3 590gtx asus mainboard

http://www.ebay.com/itm/PCI-E-16X-to-16 ... 27dc8de19a

[slepy8]

Don't go without powered riser. System will crash each time it will try to comunicate with a card.
Use the one from post above or, for more possibilities on placing the card, use the one with USB:

https://www.google.pl/search?q=usb+rise ... B600%3B432

I'm using a few of them, no problems at all. They are powered from the PSU with molex conectors.

[acc24ex]

I think you should keep in mind all USB risers convert to 1x speed
- did you check for loading times - for a scene that fills up at least 1gb in VRAM? - I would really like to see those
- I specifically got the ribbon ones to keep the loading speeds going - although the USB ones look much more flexible, and cheaper?

- I fitted 3 gpus on ribbon risers and two on the boards - it's a mess - those gpus get pretty heavy when stacked, and the crazy heat - I got two powerful fans on top of that - barely keeps the temperatures under 90 - but it works stable

- makes you wonder at the complete failure of SLI - there is no chance of you hooking up three gpus without overheating over 90 degrees C, without a lot of additional cooling, 97C is when the system shuts down I think..

[slepy8]

Judge for yourself:

Scene uses 1077MB
lots of generators and meshes.

780Ti - x16 slot - 22647.821ms
2 X 660 - both x1 slots - 25364,921ms
all 3 cards at the same time - 22664,749ms

[mate]

Maybe interesting here - actual measured throughput is 6011MB/s on PCIe3.0 16x vs 383MB/s on PCIe2.0 1x

via https://twitter.com/hashtag/7GPU?src=hash

[slepy8]

cool.
Now please give us some real rendering export times and differences betwen x16 and x1.
That will show us what those numbers you said really mean.

In my case 3sec. difference doesn't match your throughput measurements.

Oh - one moment - forgot to mention. In my case the x16 slot is PCIe 2.0 !!
This might mean the difference!

[acc24ex]

- that is a good point - we actually do not know what is the throughput of octane - how much bandwidth it uses to push memory onto the cards - and maybe most of the time it does is voxelizing or some crap like that.. - so it may take 3 second to load a gigabyte ..

- it's practical data not the theoretical that counts obviously - and according to yours there is almost no difference between 16x and 1x

6011MB/s on PCIe3.0 16x vs 383MB/s on PCIe2.0 1x

- lets do some math - 1gb is 1000 mb, 383mb * 3 sec is more than 1GB

- so yes it takes 3 seconds to load 1 GB into vram with 1x, and third of a second on 16x - is this true?

..what the F - OK - the 1x speed is like 3 seconds per GB - you might get 15 sec per 5gb - if you have the GPU to fill it with .. hmm, not that much

[mate]

true

[prehabitat]

Interesting.
While 9 seconds isn't a long time compared to a fully baked render, it does feel like a lot... Hardly going to break the time bank unless you're doing animations and yo become constrained by the 'upload' bottleneck...

Octane team: does the voxelizing occur as one singular task then send to gpu as one singular packet? (ie. 5 seconds* voxelizing work then 9 seconds upload afterwards) OR, does the upload begin as soon as some data is voxelised by the cpu...