Max GPU's with i9

[jasontv]

The company I work for has started converting some of our assets into 3D objects for VR. We use a combination of Artec and AgiSoft software to do this, however our three most powerful boxes have struggled to pump out more than two objects a day. This led my Boss to come ask me to put together the specs for a new computer that can crank out files faster. If you have never done this sort of work before you should know that divides the work between your GPU and CPU across four stages with the GPU having slightly more influence on processing time than the CPU. It can also eats through RAM without breaking a sweat. I threw this rig together to get an idea at cost, and the price was within budget. (We want to stay under $8000 if possible.) It's worth noting however that the software we use doesn't like it when you SLI the GPU's when rendering. As a result you can throw as many GPU's at it and see improvements in processing time. They don't even have to be the same model graphics card. SO...

 

1. What is the max number of GPU's supported on x299 including internal and external GPUs?

2. Please provide suggestions on the build. (We have spare Win10 keys, a wired connection, and no need for audio or perifs.)

3. Please share any other thoughts. 

[Technomancer__]

Get the amd epyc 32c 64th single socket one that is 200. So you have a very fast CPU with a shitload of pcie lanes for GPUs and power to feed them all

 

Max it out with 1080tis

[Technomancer__]

Just now, Ethocreeper said:

Get the amd epyc 32c 64th single socket one that is 200. So you have a very fast CPU with a shitload of pcie lanes for GPUs and power to feed them all

 

Max it out with 1080tis

You can also get threadripper instead epyc. It has more pcie lanes than x299

[Brooksie359]

Well for games usually 4 pcie lanes is enough bandwidth to not bottleneck modern gpus but that's for gaming so idk if it is any different for this type of workload

[Brooksie359]

1 minute ago, Ethocreeper said:

You can also get threadripper instead epyc. It has more pcie lanes than x299

Unless they need avx 2 or above. They should probably figure that out and based on that make a decision.

[jasontv]

5 minutes ago, Ethocreeper said:

You can also get threadripper instead epyc. It has more pcie lanes than x299

I wasn't thinking AMD because the i7 was outperforming the Ryzen 7. I don't know much about the Threadripper line.

[jasontv]

8 minutes ago, Brooksie359 said:

Unless they need avx 2 or above. They should probably figure that out and based on that make a decision.

Software is optimized for AVX. (Other users report AVX 2 improvements, but can't find official support.)

[Brooksie359]

7 minutes ago, jasontv said:

Software is optimized for AVX. (Other users report AVX 2 improvements, but can't find official support.)

Well amd does fine with AVX just not avx 2 and above

[Pesukarhu]

11 minutes ago, jasontv said:

I wasn't thinking AMD because the i7 was outperforming the Ryzen 7. I don't know much about the Threadripper line.

The threadripper 1950x will cost 1000$ and has 16 cores and 32 threads. It will have quad-channel memory support (up to 1 tb total memory) and 64 pci-e lanes.(i9 has 44) It is coming in "early august". The 1950x has a 180w tdp but it has a soldered heat spreader.
You could get the i9 now and start right away or wait for independent threadripper reviews before deciding.

[paddy-stone]

12 minutes ago, jasontv said:

I wasn't thinking AMD because the i7 was outperforming the Ryzen 7. I don't know much about the Threadripper line.

The only way an i7 beats ryzen is in gaming... even the intel 10 core models only just beat it for multi-thread workloads, this does depend on instruction set used, if AVX2 is used then intel wins big. The same will no doubt be said for the threadripper and EPYC CPUs,

Best thing to do is look at what instruction set the program uses, if anything other than AVX2, then AMD is the one for you. With that budget you should go for EPYC CPU I think, especially for the multitude of PCI-e lanes you can use too for multi GPUs.

[typographie]

1 hour ago, jasontv said:

I wasn't thinking AMD because the i7 was outperforming the Ryzen 7.

Core for core, you are correct. The benefit of Ryzen for the $100–400 consumer market is that it offers a relatively large number of nearly-as-fast cores for a really nice price, so a Ryzen 7 1700 is a really compelling product against a quad-core i7.

 

I don't think that's really applicable to someone with $8000 of the company's money to spend, though.  AMD's new EPYC server chips may be worth considering. They have eight-channel DDR4 and 128 PCIe lanes, and pricing still seems pretty reasonable. I don't think they're available yet though, and being server chips I'm not sure what the motherboard options are like.

[jasontv]

On 7/28/2017 at 10:11 AM, typographie said:

Core for core, you are correct. The benefit of Ryzen for the $100–400 consumer market is that it offers a relatively large number of nearly-as-fast cores for a really nice price, so a Ryzen 7 1700 is a really compelling product against a quad-core i7.

 

I don't think that's really applicable to someone with $8000 of the company's money to spend, though.  AMD's new EPYC server chips may be worth considering. They have eight-channel DDR4 and 128 PCIe lanes, and pricing still seems pretty reasonable. I don't think they're available yet though, and being server chips I'm not sure what the motherboard options are like.

I asked this question on another forum and it was suggested I look at a mining board so I could fit up to 8 GPU's in. As I look into it more I am not sure what the real world gains are after 4 GPU's. The third to fourth card seems to be about a 15% improvement. On our larger objects that cuts about six and half hours of our processing time. Likely worth the investment. However, at the diminishing returns we have seen so far the fifth card is expected to save just over 3 hours. We haven't tested it yet, but I am predicting about 24 minutes saved per object moving from seven to eight GPUs. Part of that might be due to the fact that the software disables one core for each GPU used. The only complete evaluation of benchmarks I have seen are the two from Puget Systems where they used dual Xeon E5-2687W with one to four 980's and one to four 1080's. Their (old) review suggests that 10 cores is the sweet spot for performance. Unfortunately this was done in 2015, and it used the old version of the software that only took advantage of the GPU for one stage of rendering. Even if that stage still accounts for most of the processing time. 

 

The converstaion here seems to be about the CPU only. Should I take it that there is consensus that the 1080 ti's are a good choice? I half expected the community to tell me to wait for AMD to release their cards first.