New WS CPU+GPU help

[alessandro.eu]

Hi,

 

I decided to upgrade my machine:

 

Q9550 @ 2.83GHz

Quadro fx3800

8 GB ram

 

3DsMax + Vray + Photoshop

 

after reading Dimitri's blog I think that what he suggests as CG Workstation "the pro" could meet my needs: improve render time and explore gpu render (especially vray RT for a fast feedback but with an eye to octane too) so I could go for:

 

i7 3930K

quadro 4000

2 x gtx 670

32 Gb ram

 

this is probably best price/performance day by day machine instead a 2 x Xeon configuration.

 

My doubt now is if I should keep my fx3800 instead of upgrading for a new quadro 4000.

I don't know if a new generation Quadro4000 could considerably improve fluency on the viewport or maybe I can opt for a cheaper card Quadro or ATI but honestly I don't know which are the best options price/performance at the moment.

 

I work on architectural visuals and try to keep models light but I would like to explore more detailed models especially on the interiors to enanche realism (this is why I'm thinking to upgrade the quadro).

 

I would really appreciate suggestions on the configuration and quadro/ati choice.

 

thank you and sorry for the long post

[Dimitris Tolios]

I believe you have a nice base to work with...

Since you already have an ok Quadro (which I believe would beat the 670s in viewport easily), I would recommend getting your pieces opting the 4000 out, and using your FX 3800 for the time being. If you see that the card is struggling, you can start thinking about the 4000.

 

By that point the K4000 might have been through a couple of tests / reviews and you might opt for that one.

 

- Sidenote: it appears to be the case that the up-coming IB-E s2011 processors won't come with a 8-core version.

We will still have 4xxx s2011 CPUs being a low-end Quad and mid/high end Hex cores.

The 3930K will be slower ofc, but logically around 10% from its similarly priced successor, much like the 2700K vs. 3770K case.

[alessandro.eu]

Dear Dimitris thank you for your reply.

 

I found some benchmark here and thought that comparing a fx3800 to a new quadro 600 (more or less same specs) the quadro 4000 would be much better, let's say almost the double in performance (and that could justify and upgrade for me). But of course these charts are nice but don't know if those reflect the behavior in 3ds Max.

 

What I didn't understand is also the behavior of the system cpu+gpu. Let's say that I will opt for a quadro k4000 + one gtx670 or 680. The cuda core in the quadro (768) would help the render in vrayRT or not? If yes I can still handle the viewport and render at the same time or the two cards quadro+gtx will be busy and I will be stuck?

[Dimitris Tolios]

Alessandro,

I have some benchmarks in PCfoo based on SPECviewperf® 11, where the FX3800 struggles to keep up with the 4000, while in others it keeps up decently.

In all cases the FX 3800 is much faster than the 600, which cannot really top even the FX 1800.

 

Yes, the 4000 is a faster card, no doubt. But the FX 3800 holds ok for its age.

Again, the 4000 is also a card more than 2 years old, so if I was to spent that kind of money today, I would either get it used, or I would probably go for a new K4000.

[Sauger]

I would consider a GTX Titan card for GPU rendering, it´s very silent/cool with loads om mem for large scenes.

Speed is same as 2xGTX 680, and use much less power. Double precision performance if needed is way faster too.

[alessandro.eu]

Dear Sauger thank you for your reply.

 

Yes the Titan could be a good solution instead of two gtx 680 and maybe add one more next year if needed.

 

Is still not clear to me if the cuda on the quadro could be added to the ones on the gtx for rendering in RT or if the quadro can manage only the viewport meanwhile the gtx is working on the render

[Sauger]

Octane Render use all cuda devices in my workstation (2), with priority set to low realtime navigation is smooth also even if none are dedicated to viewport.

I had a GTX 570 for viewport, but the noice was irritateing.

In Vray RT you might be able to do the same in latest ver.

[Dimitris Tolios]

Some quick points:

* A single titan is NOT faster than a pair of 680s.

* It has more RAM than any available 680, true,

* You can pack more processing power in multiple titans than the same number of 680s, true

* As far as raw speed with VRay RT GPU goes: 2x GTX 680 > 1x GTX 690 > 1x GTX Titan.

* Also, idling or plain desktop consumption is not a real issue with Kepler cards. Total consumption a 690 or two 680s will be within 5-6W of a single titan, simply because GKxxx cores slow down memory and core extensively and have excellent power management.

Edited April 9, 2013 by dtolios

[Sauger]

I think power consumption under use is the important factor, and 2x680 use far more then 1 Titan.

Nobody has mentioned that 1xTitan is faster then 2x680, speed is the same in Octane Render.

[alessandro.eu]

but

 

Isn't better to go for one Titan now than 2 x gtx 680 and maybe next year get another Titan?

price of two gtx680 is more or less one Titan

[Sauger]

I would not buy 2x680 even if it was half the price, different strokes for different folks

[Dimitris Tolios]

but

 

Isn't better to go for one Titan now than 2 x gtx 680 and maybe next year get another Titan?

price of two gtx680 is more or less one Titan

 

If you are cost driven and you believe you can work with less than 6GB of VRam, then 3x GTX 670 4GB are the best choice in a scenario with motherboard with 4x pci-e slots, properly spaced to allow for 3x "accelerator" cards + a Quadro if needed + an 8-9 slot case that can accomodate them with proper clearances. 2x 670s are right with a single Titan as far as raw performance goes in VRay RT and probably all the equiv. renderers, and 3x are doable for someone who can realistically afford a Titan.

 

If cost is not a real issue and you are positive (as with the above scenario ofc) that VRay RT suits your workflow thus worths the investment, there is no contest: a Titan now, more later - unless Chaosgroup decides to properly implement OpenCL to work with AMD cards where 79xx and upcoming 89xx cards could be compleling.

 

nVidia GTX 7xx series, will be an improvement over the GK104 that 670/680 use, but I doubt it will even come close to the GK110 - even the low-binned chips that find their way in the Titan.

 

A real issue to consider though, is Titan availability. Getting one is not easy - at least known vendors are reporting it out of stock within hours of new batches becoming available, so planning on multiple of it is not something that can be realized on the spot.

[alessandro.eu]

the seller, that I usually use, sells gtx670 469 EUR and a Titan 1099 EUR ...is for that I thought 1 Titan is better than 2 x 670 thinking to a possible more Titan in a year.

 

It is true that the availability is poor but I hope that in one month (when I will buy the pc) will be fine.

[Dimitris Tolios]

the seller, that I usually use, sells gtx670 469 EUR and a Titan 1099 EUR ...is for that I thought 1 Titan is better than 2 x 670 thinking to a possible more Titan in a year.

 

It is true that the availability is poor but I hope that in one month (when I will buy the pc) will be fine.

 

The Titan is marketed as a gaming card, not a computation card - despite being based on one.

They upmark the price for a number of reasons:

 

* It is currently the fastest single GPU in most games. SLI has quite a few issues, and many gamers dislike it / don't want to bother with it, so the Titan is much more desirable for a large player base than the 690 (it is 2x cards in SLI on the same board), despite the latter being faster.

 

* Gamers after absolute e-peen through SLI, can go 3-4 Titans that will be faster than 2 690s...remember, 2x 690s = already quad SLI, the max possible.

 

* The GK110 is a massive, very expensive and very complicated chip. The number of chips that fail QC to make it in K20s should be high, and so would the number of those that won't even qualify for a Titan (a Titan LE with extra shader/core modules missing over the Titan is already rumored = an excuse to sell chips that would be otherwise discarded). The yields on the much smaller and simpler GK104, along with the higher demand makes it a much more profitable chip even at much lower prices.

 

All of the above make it more expensive, and drive the $/performance down.

 

Just like a $75,000 car is not "3 times better" or in many ways not even "2x better" than a $25,000 car, there is not a straight relationship between CPU or GPU pricing and their actual performance.

Think 3970X being more than 2x more expensive than a 3930K, while not even 10% faster, and feel lucky about the titan

[alessandro.eu]

Thank you Dimitris for your reply,

 

Would be nice to hear if someone has experience with a gtx in vrayRT...

[Dimitris Tolios]

Thank you Dimitris for your reply,

 

Would be nice to hear if someone has experience with a gtx in vrayRT...

 

What do you mean? There are multiple people with experience with GTX and VRay RT - myself included.

 

On the Chaosgroup forum there is a standardized test scene that people render with VRay RT and report the time needed for completing 512 samples. Keep in mind that the scene is VERY simple, and 512 samples are nothing. A complex interior GPU rendering for production might get hours.

 

Aprox. numbers reported:

 

GTX Titan: 1m 46s

GTX 580: 2m 39s

GTX 680: 3m 2s

GTX 690: 1m 31s

 

GTX Titan + GTX 580: 1m 10s

2x GTX 580: 1m 29s

2x GTX 680: 1m 34s

4x GTX 680: 49.5s

 

Variations are reported due to differences in drivers, and/or variations in the cards.

Ofc a factory overclocked GPU and/or problematic driver versions might cause weird results. The 1st 4 numbers are reported by vlado, the Chaosgroup employee that leads the forums on VRay RT GPU and were run on the same system with drivers 1-2 versions behind so I would call those directly comparable.

 

The rest are from other users, and ofc having a 690 being faster than 2x GTX 680 makes little sence, but the difference is so low that could be caused from different drivers etc.

 

I believe it is good enough to get a general idea.

Edited April 13, 2013 by dtolios

[alessandro.eu]

Thank you again Dimitris you're helping a lot to make things clear.

 

I meant to ear different people to get different experiences.

 

I'm looking at the chaosgroup forum and seems that the major problem could be the amount of the RAM to handle big scenes.

The 6 Gb of the Titan could be enough but maybe for very large scenes with trees don't. But again I'm not really interested on production with RT (maybe some simple exterior would be fine only in RT) and probably one gtx 690 or Titan would fit my needs.

 

Other question for you (if you don't mind):

 

How does the system handle the two cards? (let's say quadro + gtx) I would install the quadro with its driver and plug the monitor to it but the gtx will work with the quadro drivers? Will be automatically recongnize in Max + Vray?