Why CORONA Bench is so so so slow on a AMD Opteron sys ?

[tanguybod]

it is weird regarding the two benchs I ran at the office. one on a Dual Xeons 5670 and a Dual Opteron 6272.
The two machine are benching on FRY and CineBench same ish results, around 15 for CineBench.

Here with Corona, we got:

Dual Intel 5670: 3'15"
Dual AMD 6272: 15'01"

Any clue why it is this drastic gap between both system ? They are both running a W7 64 by the way and same specs beside the CPU and RAM.
And the AMD has way more RAM than the INTEL though

[Ondra]

Try running 2.0.bat on both computers: http://dl.dropbox.com/u/5426917/crown.zip
report the average FPS you get in he console without moving the object

[tanguybod]

So the results:

Dual Xeons is 6.5/7 fps and 150ms
Dual Opterons is 4.45 fps ish and 246ms

[Ondra]

Interesting, I though the problem was in embree, but it obviously is not. I will have to get myself some AMD computer and try it for myself

[tanguybod]

yep look like something fishy here. You still can use my ws for tests, just let me know what you want me to run. i can give you the specs of it.
i have others AMD node based.
I can run the same crown test on one of them if you want.

Tang

[juang3d]

My AMD 1090t are performing well, there is no important differences between intel and amd computers here, the performance variation is teh same than with mental ray engine, more or less same ratio.

Cheers.

[tanguybod]

yep me to in others render engine or bench, that's why i was quite surprise.

[Juraj]

Scaling between regular AMDs and i7 is identical, apart from AMD being sadly too poor performers, but I know Opterons struggle heavily in most rendering engines due to some architecture reason, which I can't remember but you can search for Vlado's take on this (which might be biased, but I've seen the same score in Maxwell and other likewise). The suggestion was to run multiple spawners on single computer to harness the power of the but even than (despite horrible drawbacks like multi version of scene allocated on ram) still provided meager performance.
Top Quad-Opteron will be killer for folding and with incredible price (you can build it under funny 3-4k euros) will be beaten by regular i7 sadly in rendering.

One time I almost build one, thankfully, I did my research.

[tanguybod]

Hey,
yeah usually the Opterons are not the best fighter, but they have their advantage too. This station is my simulation platform for Houdini and it runs quite really well. Plus the 128 gig of RAm serving it, the 32 cares can breath.

I ran the bench without the hacking running. I could do it again while hacking the core, which more or less pumps up the clock full speed ALL THE TIME.
Here the Opeterons always take a shit load of time to gear full speed while the Xeons crunch asap full speed.

Now the Opterons are really good CPU. No matter what at the end, not the GIGAFLOP/s are what counts but the ratio PRICE/CORE.
AMD is unbeaten on that. the same benchmark on CineBENCH between this rig and a 32GIG RAM Dual Xeons 5670 at the time I built them was quite eloquent... 10Gs for the Dual INTEL against Gs for the Dual AMD. You push AMD at 10k in money spent I would have a 75Bench.

AMD Opterons has to be socked in a very carefully picked motherbord I guess, that is the trick.

Well,
enough for tonight !

Cheers

[fh3d]

The 6200 Opteron processors (and all 3200-6300 series Opterons) share one FPU between each pair of CPU cores.  The Phenom II (AMD 1090) have one FPU per CPU which is why they scale better against the Intel chips. 

[Juraj]

No matter what at the end, not the GIGAFLOP/s are what counts but the ratio PRICE/CORE.
AMD is unbeaten on that.

That's imho faulty logic, since per actual performance, 1 Sandy Bridge Core (2 threads with HT), correlates to 3-4 Cores (4 threads, no HT similar tech for AMD) of 63xx Opteron family.
What that meant in previous generation,  was that 4x150W ( 600W) Opterons (64cores) still fell behind in multi-threaded performance compared to 2x115W ( 230W ) SandyBridge 26xx Octacores (16cores total only, 32 threads with HT.

This isn't so well, since you need 4-socket mob, 4 times the memory to utilize identical performance due to individual access, need for Server based OS (even Windows Server), which disqulifies it from being Workstation friendly,
 2 to 3 times higher TDP and associated electricity costs and zero single-thread power.

I like AMD a lot, it's much better company on moral grounds than Intel, but they simply vanished from high-end market, even on Server grounds.

But I wouldn't suggest it for CGI, since so many renderer struggle to use it to its potential.

[tanguybod]

it depends your logic indeed.
And again, I compared chips which are 3 years old dude, both of the INTEL or AMD.
And no matter what all your calculus about cores and RAM, at the end, 2 opteron 6272 brings same power as two XEON 5670. What do you have to say about that ? Both run on same power supply and run at the least same huge minimal RAM, 48GIG, again those two rigs are not render node, but ws.
So at the end Juraj, you put the price of two Xeons and two Opterons, INTEL cost 10 ground at that time for the ws while the AMD cost 3500. For a same benchmark.
My renderfarm is made of single chip only, I never run multiple cpu render nodes, to quick depreciated and expensive to deploy + when you lost one, you losty a pretty decent amount of rendering power. We have around 35 of them and yes, it is cheaper. But you can't only take in consideration wattage or cores. Running single CPU nodes has a price, specially i7, they run HOT, which at the end is killing you, heat and wattage.
BUT again, we made the math, even let say a renderfarm running ARNOLD or RENDER MAN will cost less on a single socket node + cost of the license per node rather than multi socket rig for a node.
I don't want to argue, don't get me wrong, what I gave you guys are real facts, not spread sheet before byuing the dream machine everybody thinks he is going to assemble. ;) me including of course :)

No I didn't ran a dual E5 Xeons, maybe they are amazing, but still, they cost an arm and one leg.

Cheers

[Juraj]

what I gave you guys are real facts    ///// "10Gs for the Dual INTEL against Gs for the Dual AMD. You push AMD at 10k in money spent I would have a 75Bench."

I disagree here. I don't see it as factual  what you wrote in above, it's very misleading in my opinion and far from truth I feel I have to comment on it because it could sway someone to even buy these (happened many times on Chaos forum, only to people desperately regret doing it), which would be pretty bad mistake. I don't want people to throw money into trash because they see apparent miracle solution on forum being very cheap now. Lot's of people buy poor performance old server systems from eBay thinking they are getting great deals)
Notwithstanding that even the most expensive E7-46xx Quad Xeon setup consting 16 000+ euros released only month ago would have trouble reaching 75+ Cinebench. That's just crazy.

(I am also confused why you explain me the negatives/positives of running multiple i7s as render nodes. I have 5x 4930k workstations, I use 4x 2680 v2 E5 Xeons as my true nodes (140 logical cores total). I might buy 4 another in summer, I value my comfort level and all the benefits much more than saving every dollar to end up with full room of boxes. But that's sideway. )

Your math doesn't add up. You hyperbolize numbers like "10Grand vs almost nothing" with little facts supporting this extravenous claims.

[Absolute comparison of 6272 vs 5650 in Maxwell, Cinebench, Blender, and all associated ] http://www.anandtech.com/show/5058/amds-opteron-interlagos-6200/14

Opteron 6272 was direct competitor for Xeon 5650 not 5670 (despite your FryBench score) and at price difference (at start, with Interlagos Opterons being much much newer) of 800 dollars vs 1000 dollars (now both cost +/- 400 dollars), which is only 20 perc. more expensive if I do my math right ?
You compared absolutely more high-end CPU to justify flawed argument.
And at this similar price level, the Xeon ranges from 5perc. faster, to 44 perc. faster in softwares limited by Opteron Architecture {Shared FPU, CMT, floating point limitations}.
And this is comparison for pure multi-threaded performance, for Workstation purpose (as you state), it would be undisputed in favour of Xeons, again, the above benchmarks showcase this very clear with big difference. Workstation processes
are largely still based around single-threaded performance, something that very much AMD lacked in (with exception of great old Athlon 64, which was amazing at its time), and these Opterons would be beat by low-clocked i3/i5 (first gen, Westemere, to be at the same year of 2011)

Xeons don't cost an "arm and one leg" when compared to Opteron system. Both systems cost "arm and leg" when compared to regular consumer market units (i7, Bulldozers), but compared to each other, the older Xeon systems used to be more costlier by no more than 20-30perc. at similar performance level, which doesn't apply anymore today, since E5/E7 Ivy Bridge Xeon family is top performance tier, while recent Opteron generations (Delhi, Seould), focused on micro-server architecture (not performance!) in 2012, before absolutely vanishing from market. Opteron is dead now.

Please don't assume I write this on conflicting manner, it's how I discuss HW :- ) It's all good, and I hope they serve you well. I just don't want people to buy them now. Enjoy your day !