Maxwell Render | Official public forum

Hi everyone.

I don't know how Maxwell scales with multi-core processors.

For example:

There are a few Cinebench tests comparing Intel Xeons and AMD Opterons.
Would the tests of a scanline renderer such as Cinebench be useful to someone
using Maxwell render?

I'm interested because I encountered two render node solutions offered by the same company.

One with 96 Cores AMD 2.1GHZ (8 X 12-cores) which is considerably less expensive than
48 Cores Xeon 5645 (8 x 6-cores) running at 3.33 GHz.

What suits Maxwell best?
Faster processors or multiple cores?

El

Faster processors definitely... I saw a test with 48 cores that suggests that efficiency per core drops as you get higher in core count -- not by much but by the time you get to 48 cores it looks like you lose about 20% efficiency per core over a standard 6 core solution. Which makes sense with all the extra overhead of managing that many data pipes.

Obviously there are advantages to having all those cores in one system -- but I would take the option with the fastest processors.

My understanding is this efficiency issue may level out at higher sample levels (20+), but not many seem to be willing to let their scenes mature to those levels to find out.

Best,
Jason.

Thank you Jason.


The units are split into 4 render nodes so hopefully the
drop in performance will not be as bad as 20% on the 48 cores
as long as renders are in non-cooperative mode.

El

Yes, that would be true based on what I understand -- and of course this may be less of an issue when dealing with high sample level scenes anyway

Best,
Jason.

Cooperative mode will have little if any impact on performance, until it merges the mxi's.

Get the 8 x 6-cores. I'm not 100% sure, but I'm willing to bet that the Intel 6-core will outperform the AMD 12 core head to head with Maxwell.

What you guys haven't talked about yet, though, is the fact that the AMD setup has a total of 201.6 Ghz (96 cores x 2.1 Ghz/core) vs. the intel setup of 159.84 Ghz (48 cores x 3.33 Ghz/core) which is 26% more. From everything I've seen Maxwell acts pretty linearly, but I've never gotten into this many cores (if you have the time and inclination perhaps you could email a couple render farms for advice?), so even if there was a 20% hit which seems to be on the high end, the AMD would still be a bit faster (and quite a bit cheaper as was mentioned).

I do believe HT has been shown to increase speed, though. The xeons have HT, what about the AMDs?

-Brodie

What I should have said to make it clear is: " the more cores the more you will notice a loss" because the tested system was only 48 cores the results only showed up to that point but it very clearly showed a loss as you added more cores.

What I'm saying is I don't believe it "tops out" at 20% loss -- I think the more cores the more you lose in efficiency... so the larger core count may be moot (unless as I said the jobs are divided in more sensible divisions of core-count)

Nearly linearly is not the same thing as linearly -- and obviously this is a very extreme core count not available to most users at this time... I imagine Maxwell will adapt when those core counts are considered more "normal".

Best,
Jason.

brodie_geers wrote:What you guys haven't talked about yet, though, is the fact that the AMD setup has a total of 201.6 Ghz (96 cores x 2.1 Ghz/core) vs. the intel setup of 159.84 Ghz (48 cores x 3.33 Ghz/core) which is 26% more. From everything I've seen Maxwell acts pretty linearly, but I've never gotten into this many cores (if you have the time and inclination perhaps you could email a couple render farms for advice?), so even if there was a 20% hit which seems to be on the high end, the AMD would still be a bit faster (and quite a bit cheaper as was mentioned).

I do believe HT has been shown to increase speed, though. The xeons have HT, what about the AMDs?

-Brodie

no, AMD has no HT - a 12 core opteron can be compared with a 6 core xeon with ht
here are some cinebench scores:

http://www.anandtech.com/show/2978/amd- ... ore-xeon/6

in this comparison they used two Opteron 6174,12x 2.20GHz (~1000€ inkl. VAT) and two Xeon 5670, 6x2,93GHz (~1150€)


so the 2.10GHz Opteron 6172 (~850€) mentioned in this thread should be a bit slower and the xeon 5680 with 3,33GHz (~1350€) should be a bit faster. so the difference between these two systems should be more noticable.
A Xeon 5645 is only 2,4GHz (~450€) not 3,33GHz! so this one maybe could be at the same speed like the 2,1GHz opterons...

As far as i remember maxwell scales a bit better with HT-cores than cinebench - but i have no scores to prove it...

a new benchwell site would be helpful!!!! Where is the problem with it that it take 2 years to set up a new one?!?