Blackmagic Forum

First of all: I am totally aware that two 2080tis is still only 11GB of VRAM, while Titan RTX offers the whopping 24GB. But - 2x 2080ti cards will have more CUDA and RT cores than the Titan RTX, so the question is still valid (plus, the price of both solution is roughly the same again).

I also know what would BMD support call the winner - but I also know Puget actually benchmarked both, and the Titan is only marginally faster than a single 2080ti...

So?

Regards,
Piotr

So... it depends on whether you need CUDA horsepower or GPU memory, I suppose. Two RTX 2080ti should be faster based on your own observations (you reported a 40% speed increase having a second GPU IIRC) but you will need the GPU memory if you are processing large frame sizes and that is probably the best way to think of the trade-off since individual frames cannot be decomposed it means that even with faster (more) CUDA cores, if you are processing 9K material, memory may be more important than raw horsepower.

FWIW, My single 8 GB P4000 is not cutting it with Resolve Studio 16 but I will not upgrade my GPU until BMD is finished fixing what appears to be an over aggressive use of GPU memory... I've said this before, I know... broken record and all.

Steve Alexander wrote:So... it depends on whether you need CUDA horsepower or GPU memory, I suppose. Two RTX 2080ti should be faster based on your own observations (you reported a 40% speed increase having a second GPU IIRC) but you will need the GPU memory if you are processing large frame sizes and that is probably the best way to think of the trade-off since individual frames cannot be decomposed it means that even with faster (more) CUDA cores, if you are processing 9K material, memory may be more important than raw horsepower.

Steve - I don't and most probably will never do more than 4K (not in this lifetime, and I'm being serious). However, you got me thinking: when I do a Multi-cam T/L (off 4-6 UHD cameras), does it increase the VRAM capacity requirements?So far, I mostly did my M-C cutting and transcoding in Vegas Pro, as somehow I still find it much better suited for this kind of job - especially with music, when I do my A/V synchro with a millisecond "granularity". But sooner or later I would like to start doing everything in Resolve, so would like to know whether a nasty surprise is awaiting me with just 11GB of GPU VRAM!

Piotr

Piotr Wozniacki wrote:I also know what would BMD support call the winner - but I also know Puget actually benchmarked both, and the Titan is only marginally faster than a single 2080ti...

So the main group of people I'm politely addressing with this are actual users, who had an opportunity of comparing such two configurations... What is your real-life experience? TIA

Piotr

i have to deliver Helium 8k at source raster, useing an 11g card (or two) is painful and frustrating for this workflow - useing 2x 1080Ti's i can easly do the same at 5k no sweat, no drama, no caches.

i don't edit with Resolve much, don't use mutlicam, do use split screen/neighbours, so four streams running @ 4k, the machine has no big deal playing at 24 untill i run into a scene with NR, then it drops as the GPU's try to run NR on four streams of cam orig @ 5k at once - the disk sub system and CPU's are not bottlenecks in my machine

if you don't have a use case for delivering larger than 4k, i would go with the 2080Ti's

if you have a use case for delivering 8k, i would go with dual RTX Titans, assumeing the client is going to pay for the resources they are asking to use

Piotr Wozniacki wrote:.../...
So the main group of people I'm politely addressing with this are actual users, who had an opportunity of comparing such two configurations... What is your real-life experience? TIA

Piotr

Piotr,

You already said on the forum that you have 2 big big GPUs to get good results (RT) with your clips in (SONY FS7 if I remember correctly) at 4K to 50 FPS.

If it works well: what more do you hope by opening this thread?

I can not follow you.

Jean Claude wrote: what more do you hope by opening this thread?

I can not follow you.

J-C, thanks for chiming in. I did try sevral different card with my new PC while building it, but I never had enough money to actually purchase 2x Titan RTX (if this is what you mean) - simply because I don't have that much money

But you happen to be the right person to share your experience with me - considering strictly the choice I mentioned in the header: in my (now closed) thread, you said that even with 8K you can successfully use the RTX 2080ti if you are careful what number of frames you're using for TNR, and which will already be too much for the 11 GB of memory those cards have. You see - this is the kind of practical input I'm looking for from people, and this is why I "opened this thread": considering I never do 8K (4K or UHD is my max resolution), but using M-C often - will I or will I not be able to run TNR with each of say 4 to 6 streams, with a reasonable number of frames (like 3, for instance), but w/o the "GPU memory full" error?

Will you help m with some suggestions, J-C? I'm sure you will - TIA

Piotr

PS. Please believe me, J-C, that I only just started to think I'm finally grasping this terrific software's philosophy - and soon after, the "blind" recommendation from BMD to use a single GPU with just one criterion: the VRAM capacity.... It really ruined what I hoped was my maturing understanding of DR "philosophy". Isn't it strange confronted with the fact that it is the full Studio version supports multiple GPUs? In my now locked thread, I unleashed my imagination for awhile and proposed a "theory" of coming Resolve developments which would stay in a full accordance with this BMD's recommendation; I know it most probably was plain silly what I wrote, but at the same time, it's the only scenario which would fully support it? Yet even that has gone unnoticed by our BMD Friends here... Pity

Jean Claude wrote:

Piotr Wozniacki wrote:.../...
So the main group of people I'm politely addressing with this are actual users, who had an opportunity of comparing such two configurations... What is your real-life experience? TIA

Piotr

Piotr,

You already said on the forum that you have 2 big big GPUs to get good results (RT) with your clips in (SONY FS7 if I remember correctly) at 4K to 50 FPS.

If it works well: what more do you hope by opening this thread?

I can not follow you.

Just to make myself clear: when I was saying I was able to play UHD@50p in real time, I meant playback of either untouched material or - if I color-corrected it, applied NR and perhaps 1-2 OFXes - playing back the cached media. Otherwise - i.e. with the above 4-5 nodes, including NR and OFX - I could NEVER play back in real time! If you so understood, it simply was some sort of confusion.

But caching entire TL (even a short one, like 5 minutes long) takes ages with 2x Titan Xp cards (CPU only loaded at some 30%, BOTH Titans - 98-100%), and this is my reason I'm still looking to optimize it so that:

- the uncached but edited (nothing fancy, though) media plays back at a more acceptable speed (in the case of 50p this would mean at least 20-30 fps)

- caching (in order to achieve RT p/b) is faster than the current 30-40 minutes for 5 mins TL (this is what it takes when both TNR and 1-2 OFX are present in addition to some base CC nodes)

So really, the answer which would give me what I want (single Titan RTX or two 2080tis) should be easy who actually tried both of the GPU configurations discussed here...

Piotr

DaVinci Resolve Studio doesn't support Nvlink then you can't have 11+11GB. Nvlink also is a bottle neck for realtime performance, it's ok only for post-process rendering of big textures (many layers of 4-8K videos)
Also speed increase is not linear, for example adding a second gpu to your workstation you can't get 2x rendering performance.
Cuda 2D image rendering is not pathtracing, for 3D gpu based rendering engines like Octane or Red Shift you can have 2-3-4x rendering performance.
Because DaVinci needs to process everything in realtime, it's more important to have just one powerful gpu than many with lower performance.
I suggest you to buy the RTX Titan.

regards.

Gabriele Gelfo wrote:DaVinci Resolve Studio doesn't support Nvlink then you can't have 11+11GB. Nvlink also is a bottle neck for realtime performance, it's ok only for post-process rendering of big textures (many layers of 4-8K videos)
Also speed increase is not linear, for example adding a second gpu to your workstation you can't get 2x rendering performance.
Cuda 2D image rendering is not pathtracing, for 3D gpu based rendering engines like Octane or Red Shift you can have 2-3-4x rendering performance.
Because DaVinci needs to process everything in realtime, it's more important to have just one powerful gpu than many with lower performance.
I suggest you to buy the RTX Titan.

regards.

Hi Gabrielle (love that name as my only daughter is Gabrielle, as well - or are you actually a he? )

I do agree with most of your reasoning - and generally that there is no linear scalability; full point. That said, I can cleanly see some 35-40% caching speed increase when I use both Titans (compared to just one of them). So unfortunately - unless someone explains to me in which way the increased VRAM helps ("GPU memory full" errors aside - I never had them with 12GB Titans; chances are I wouldn't get them with 11GB 2080tis either, based on the actually used VRAM observed to never exceed 8GB) - I'm still leaning towards doubling the number of all types of cores the RTX architecture offers...

Anyone, please? TIA

Piotr

PS. I'm now wondering whether comparing the performance between single and double Titan Xp by simply checking/un-checking the "use the GUI GPU for computing" is a valid method, i.e. whether for reliable comparo one of the cards should physically be removed from my system?

Hello Piotr,

1) You post a link to a clip that lasts 5 minutes
2) you post a link to a project that uses this clip and effect BMD OFX or whatever you want
3) remember what is the first HDD declared in preferences => system => Media storage
4) Which Media: HDD? SSD? NVME?
5) go to the color tab and take a screenshot during a reading from your project
5a) time for render cache
6) I see you have AMD TR 2990WX CPU: 32-core, 64-thread Desktop Processor, I have always been for over 3 years my 6-core, 12-thread: it will be seen if it can reassure you on your 2 GPUs.

I have been considering a second RTX Titan to help with Sapphire effects. some are sooooo slow.

Careful (if you value your money). Some processing of Sapphire and its Mocha mask are not GPU accelerated and may actually be single threaded so you could spend a lot of money and see close to zero improvement.

I can't tell you how many times over the years I've upgraded my hardware only to find that my favorite NLE doesn't use (pick one) multiple cores, GPU, memory / disk caching, etc... or at least is not optimized for these technologies.

I'm not saying that a second Titan RTX ($2500 USD) won't help with some processing but it may not help speed up Sapphire specifically.

Wouldn't it be great if we could find some users who could do the benchmarks for us?

Hey - maybe you could be that person

Cheers

Jean Claude wrote:Hello Piotr,

1) You post a link to a clip that lasts 5 minutes
2) you post a link to a project that uses this clip and effect BMD OFX or whatever you want
3) remember what is the first HDD declared in preferences => system => Media storage
4) Which Media: HDD? SSD? NVME?
5) go to the color tab and take a screenshot during a reading from your project
5a) time for render cache
6) I see you have AMD TR 2990WX CPU: 32-core, 64-thread Desktop Processor, I have always been for over 3 years my 6-core, 12-thread: it will be seen if it can reassure you on your 2 GPUs.

Yes Sir, General Jean-Claude, Sir!

Seriously though - I will not prove to you, point by point, that what I described my system functions like is true; I can only add the missing info that I keep my media on the HDD RAID 0, and cache to NVMe RAID 0 - I need a faster drive for cache as my cached media are DNxHR 444 HDR, so are obviously higher bitrate that the source clips.

therefore, Private J-C:

1. Do you or do you not know the answer to my main question? I.e whether the VRAM the Titan RTX has in abundance does anything better to my system than the doubled number of CUDA or RT cores do in 2x 2080ti config?

2. Private J-C, dismissed!



Piotr

My experience is:

Two cards will allow better realtime playback performance but one card with more memory will allow more nodes before Resolve slows down. For me one bigger card is better. Only have Mac experience.

Peter Cave wrote:My experience is:

Two cards will allow better realtime playback performance but one card with more memory will allow more nodes before Resolve slows down. For me one bigger card is better. Only have Mac experience.

Peter,

Your observation is the most convincing so far - though it suggests Resolve isn't fully optimized yet. If it was, the "GPU memory full" error should never appear; with a card not having enough VRAM for a given task (number and complexity of nodes), lower fps would simply be available as it would wrangle smaller portions of data more frequently, but memory buffer overflow(s) shouldn't necessarily happen...

Who knows - as our BMD Friends visiting this forum know exactly where the optimization is at the moment and which direction it's going - what they are NOT saying is that Resolve will soon be able to use all the VRAM available on a card? If this is the case, larger portions of data would need to be shuffled less frequently, thus greatly increasing the performance of the card... Depending on where the "sweet spot" is planned in future Resolve versions, perhaps the entire 24 GB will used, thus making a single GPU with that much memory more efficient indeed than a couple of GPUs with half of that memory capacity, even though they might offer larger computing power in terms of CUDA cores and alike...

I'm deciding between a single Titan RTX and 2x RTX 2080ti as I type, so would appreciate some more "warm"/"cold" hints from BMD. If what I wrote above is the direction, the winner is clearly the 24GB Titan RTX - the more so that the total GPU computing power can be increased by the usual 40-50%, simply adding another one down the road (should the wallet allow) - while having bought 2x RTX 2080ti, one is hosed with their 11 GB forever! Hmmm....

Thanks in advance
Piotr

Piotr,

BMD have already advised you of their opinion in your previous (locked) topic. I would trust their advice. I would buy a single gpu card over two lesser cards. It will leave room for additional cards in future. Theoretically assessing performance is rarely useful in a real world situation.

Piotr Wozniacki wrote:
Yes Sir, General Jean-Claude, Sir!
.../...bla bla
2. Private J-C, dismissed!

Piotr

Hello Corporal Piotr,

You're so pitiful that I let you dig.


There are those who know and who dig
AND
6) I see you have AMD TR 2990WX CPU: 32-core, 64-thread Desktop Processor, I have always been for over 3 years my 6-core, 12-thread: it will be seen if it can reassure you on your 2 GPUs.

(freely inspired by a film by Sergio Leone)

Continue to spend your money as you wish. I would be you: I would ask for advice in hardware and software architectures to friends.

Did not you teach that to your polytechnic school?

Continue to spend your money as you wish. I would be you: I would ask advice in hardware and software architechtures to friends but not to me!

Did not you teach that to your polytechnic school?

If not learned: as you take me for an idiot: continue digging !!

Man do you guys know how to push each other's buttons.

If it were me and I had the extra $2500 to spend one way or the other, I might go with the Titan RTX just because it more closely matches the recommended Quadro RTX 6000 plus I could then purchase a second Titan RTX for even better real-time playback at some future date.

Peter's question comes down to one of best bang for the buck:

With $2500 USD to spend and knowing that 4K is the target resolution and assuming all else is equal (meaning the workstation specs are sufficient and that the only bottleneck is GPU), which will give better performance in both the color page and fusion page - the single Titan RTX or two 2080 Ti's...

Piotr Wozniacki wrote:
Yes Sir, General Jean-Claude, Sir!
.../...bla bla
2. Private J-C, dismissed!

Piotr

Hello Corporal Piotr,

You're so pitiful that I let you dig.


There are those who know and who dig
AND
6) I see you have AMD TR 2990WX CPU: 32-core, 64-thread Desktop Processor, I have always been for over 3 years my 6-core, 12-thread: it will be seen if it can reassure you on your 2 GPUs.

(freely inspired by a film by Sergio Leone)

Continue to spend your money as you wish. I would be you: I would ask for advice in hardware and software architectures to friends.

Did not you teach that to your polytechnic school?

Continue to spend your money as you wish. I would be you: I would ask advice in hardware and software architechtures to friends but not to me!

Did not you teach that to your polytechnic school?

If not learned: as you take me for an idiot: continue digging !!

Oups !!! and GPU parallel computing (I forgot) : A module a little special but very interesting !!

Piotr Wozniacki wrote:Who knows - as our BMD Friends visiting this forum know exactly where the optimization is at the moment and which direction it's going - what they are NOT saying is that Resolve will soon be able to use all the VRAM available on a card? If this is the case, larger portions of data would need to be shuffled less frequently, thus greatly increasing the performance of the card... Depending on where the "sweet spot" is planned in future Resolve versions, perhaps the entire 24 GB will used, thus making a single GPU with that much memory more efficient indeed than a couple of GPUs with half of that memory capacity, even though they might offer larger computing power in terms of CUDA cores and alike...

Honestly, I think you've been given more than enough anecdotal evidence about what to choose.

If all you care about is high FPS then multiple cards is better than one as long as you don't use more vram than an individual card has.

if you're going to be doing memory intensive tasks like TNR on 4/6/8k footage, then you need a GPU whith the largest amount of vram possible.

There isn't going to be one best answer, because it depends on what individual users are doing.

Dan Sherman wrote:There isn't going to be one best answer, because it depends on what individual users are doing.

With all due respect, Dan - I have admitted the criteria you mentioned are convincing. However, there is one more important factor in addition to the one I quoted from your post:

BMD's plans on the optimization of GPU usage.

And there must be some up BMD's sleeve judging from how many places/times they advise a single, powerful GPU with lots of VRAM rather than several smaller ones. They never differentiated this new "rule" according to the type of work or goals to be achieved, though... Leaving us with doubts like mine, I wish they had said nothing at all if they couldn't share the whole story.

Piotr

Piotr Wozniacki wrote:And there must be some up BMD's sleeve judging from how many places/times they advise a single, powerful GPU with lots of VRAM rather than several smaller ones. They never differentiated this new "rule" according to the type of work or goals to be achieved, though... Leaving us with doubts like mine, I wish they had said nothing at all if they couldn't share the whole story.

I don't think they have anything up their sleeve. As GPU decode, and gpu-accelerated effects become more and more common your average user is going to consume more vram.

Even if bm will optimize the GPU usage better, one big card will leave a slot free for a twin in the future, while having two already will force to replace both

waltervolpatto wrote:Even if bm will optimize the GPU usage better, one big card will leave a slot free for a twin in the future, while having two already will force to replace both

Sure - but if BMD does not teach GPUs to always use their entire VRAM, you will get exactly the same starting with a single RTX 2080Ti...Only for half initial investment cost.

Piotr

Piotr Wozniacki wrote:

Dan Sherman wrote:There isn't going to be one best answer, because it depends on what individual users are doing.

With all due respect, Dan - I have admitted the criteria you mentioned are convincing. However, there is one more important factor in addition to the one I quoted from your post:

BMD's plans on the optimization of GPU usage.

And there must be some up BMD's sleeve judging from how many places/times they advise a single, powerful GPU with lots of VRAM rather than several smaller ones. They never differentiated this new "rule" according to the type of work or goals to be achieved, though... Leaving us with doubts like mine, I wish they had said nothing at all if they couldn't share the whole story.

Piotr

Some things can't be further optimised- this is the problem. GPU is all about parallelism, but some processes (like motion blur) are difficult to be performed in parallel.

Andrew Kolakowski wrote:

Piotr Wozniacki wrote:

Dan Sherman wrote:There isn't going to be one best answer, because it depends on what individual users are doing.

With all due respect, Dan - I have admitted the criteria you mentioned are convincing. However, there is one more important factor in addition to the one I quoted from your post:

BMD's plans on the optimization of GPU usage.

And there must be some up BMD's sleeve judging from how many places/times they advise a single, powerful GPU with lots of VRAM rather than several smaller ones. They never differentiated this new "rule" according to the type of work or goals to be achieved, though... Leaving us with doubts like mine, I wish they had said nothing at all if they couldn't share the whole story.

Piotr

Some things can't be further optimised- this is the problem. GPU is all about parallelism, but some processes (like motion blur) are difficult to be performed in parallel.

Not to mention some tasks require you to get all info needed for the task in vram. If you have to go back and forth between the gpu and the rest of the system you might minimize vram usage, but you will murder throughput.

@ Piotr Wozniacki

You could try writing to Puget Systems and ask them to test 1 card vs 2. They might be interested to do this simply for the publicity.

Level1Techs and EposVox have also proven to be DaVinci fans. Level1Techs seems to appreciate good engineering.

Piotr Wozniacki wrote:Sure - but if BMD does not teach GPUs to always use their entire VRAM, you will get exactly the same starting with a single RTX 2080Ti...Only for half initial investment cost.

export Helium @ cam orig with deflicker OFX and get back to us on that theory....

Dermot Shane wrote:

Piotr Wozniacki wrote:Sure - but if BMD does not teach GPUs to always use their entire VRAM, you will get exactly the same starting with a single RTX 2080Ti...Only for half initial investment cost.

export Helium @ cam orig with deflicker OFX and get back to us on that theory....

But Dermot - I did say I will never be doing anything like this, and with what I am doing I never need more than 6-8GB of VRAM.

Below is my typical T/L - with RCM from S-log3/S-gamut3.cine to ST2084/rec2020, CC, TNR (with 4 frames), SNR (enhanced) and three OFXes, including deflicker (I added it for you ) - look at the VRAM being used during cache rendering (just 7.8 out of the 12 GB), and how much of a bottleneck my current GPUs are for this potent CPU (only allowing to tax it at 15%). Resolve was busy re-creating cached media I invalidated for this clip by adding the deflicker OFX when I took this snapshot; how slow this process is with the current GPUs can also be seen in Drive S: (Resolve cache on 3x 1TB NVMe RAID 0) usage; just 1% of its huge maximum throughput (this probably was a moment when it was closing previous/opening new file; normally my DNxHR 444 HDR it writes at 50-200 MBps)...


FYI: when the cached media are ready and I play the timeline back, it always does it at a solid 50 fps; the CPU/GPU loads change to: up to 100% CPU (at 50 fps) or 50% CPU (with UHD@25p projects) and some 15-25% GPU.

Piotr

PS. To reiterate and possibly finish this thread, let me tell you this: for my kind of work and Resolve as it works now, I'm sure 2x 2080Ti would be a better solution. What I fear though is that if I buy those, and within say one year Resolve can use more VRAM (and thus work much faster) even with my simple type of 4K work, I'd be furious with myself I didn't buy a single Titan RTX for roughly the same price.

Pity nobody from BMD cared to help further, but thanks to all of you guys who did for your invaluable opinion!

kinvermark wrote:@ Piotr Wozniacki

You could try writing to Puget Systems and ask them to test 1 card vs 2. They might be interested to do this simply for the publicity.

Dear Mark,

I did think of doing so long ago, believe me - but before communicating with them, one must create an account on the Puget Systems "portal" and log in. Unfortunately, when it recognizes my IP as located outside US, it blocks me (there is fine print on their pages saying that they only serve the US community - so much of globalization they endorse)... So - even though I do realize this might be too much to ask - all I can do is very kindly indeed ask you or any of our American Friends here to ask those guys at Puget Systems my legit question (perhaps even drawing their attention to this thread). Thanks in advance to anyone who can do it for me! But please - do it asap, as selling second-hand Titan Xps is becoming more and more difficult, and it so happens I just have a seriously interested and nice buyer (hence the intensification of my beating this dead horse lately)...Wouldn't like to lose this opportunity!

Cheers

Piotr

PS. I need to make up mind by Monday morning - it's obviously too late to get hard numbers from Puget (even asking them to run such an express benchmark would be outright insolent) - but to ask them their opinion is another matter...

"To reiterate and possibly finish this thread, let me tell you this: for my kind of work and Resolve as it works now, I'm sure 2x 2080Ti would be a better solution."

You have answered your own concerns very well. I don't think anyone knows for sure what the future holds regarding GPU's. I don't know of ANY software company that will make definitive comments regarding future CPU & GPU usage or what their development roadmap is.

Quite a few comments from users have pointed you towards a powerful single card solution, but you have determined otherwise for your particular usage. I hope it works well for you.

Personally I would ALWAYS choose a single card solution where possible, due to the extra VRAM available, and BMD's advice and my own experience.

@ Piotr Wozniacki

Hi Piotr,

I sent them an email today. Better late than never, I guess.

Regards,

Mark.

kinvermark wrote:@ Piotr Wozniacki

Hi Piotr,

I sent them an email today. Better late than never, I guess.

Regards,

Mark.

Thanks a lot , I will wait a day or two...

Piotr

Sent from my VTR-L29 using Tapatalk

Dear Mark,

As soon as you have any answer or hint, please let me know. Thanks so much!

Piotr

Hey everyone, Mark tossed me an email earlier to let me know about this thread. I think I can throw some light on things here! Sorry if I miss any details, I skimmed through the thread, but may have missed something.

In terms of raw performance, the RTX 2080 Ti and Titan RTX are pretty much identical in Resolve. This article has the latest public results if you want to dig into it: ww.pugetsystems.com/labs/articles/DaVinci-Resolve-15-NVIDIA-Titan-RTX-Performance-1328/#IstheNVIDIATitanRTX24GBGPUgoodforDaVinciResolve? . I'm also working on updating our benchmarks and re-did testing with Resolve 16 Beta and the results were pretty much the same - so nothing has changed in the newer versions.


If you are looking for the most performance, dual RTX 2080 Ti is generally going to be faster in Resolve than a single Titan RTX. This has the latest benchmarks looking at GPU scaling in Resolve: ww.pugetsystems.com/labs/articles/DaVinci-Resolve-15-NVIDIA-GeForce-RTX-2080-2080-Ti-Performance-1238/#GPUScaling-BenchmarkAnalysis , but I would expect dual RTX 2080 Ti to be up to about 20-30% faster than a single Titan RTX.

The exception to this is if you actually need the extra VRAM that is on the Titan RTX. This is a weird coincidence, but I was actually just looking at VRAM usage the other day while working on our new benchmark and stuck a quick chart up on Twitter: twitter.com/PugetMattBach/status/1143309444417110016


VRAM usage is a bit weird (just like system RAM) since there is some factor of "expand into the usable space", but the 11GB on the 2080 Ti should be just fine for anyone working with 4K media and timelines (or even 8K media in 4K timelines). The 2080 Ti is our go-to card for Resolve, and it is rare anyone has issues with the VRAM with 4K media unless they are using extra plugins that eat of VRAM. The 24GB of VRAM on the Titan RTX is really only going to be useful if you are going to work with 6K/8K media. Without that extra VRAM in that case, working in the edit/color tabs isn't going to be feasible unless you really like seeing that "out of memory" error pop up.

A couple other things to keep in mind: First, there are no "blower-style" Titan RTX cards, so the most you can put in a workstation (unless you go to extremes in terms of airflow) is going to be 2 cards. The 2080 Ti, on the other hand, we regularly use up to 4 cards in a single system. Useful if you might want to expand later.

Second, Resolve doesn't use NVLINK yet, but if/when they do, that should make dual RTX 2080 Ti even faster than a single RTX Titan. Depending on how they implement it, it may also mean we will get to "pool" together the VRAM from both the cards. I give that a 50/50 chance though, it is just as likely they focus on NVLINK for pure performance or ignore it completely.

Hope that helps answer your question! Basically, for you, get dual RTX 2080 Ti.

By the way, someone mentioned needing a US IP to register a full account at Puget Systems. That is true, but we did just add social media login that doesn't care where you are from. We also have open comments on all our articles and contact info on our About Us -> Contact Us page. Neither of those are behind a registration wall (although you need a Disqus account to post comments I believe).

Great discussion... but you guys do realize that the Raspberry Pi 4 is now out don't you?

Makes all this GPU/CPU two versus one kind of academic, right?

Puget_MattBach wrote:Hey everyone, Mark tossed me an email earlier to let me know about this thread. I think I can throw some light on things here! Sorry if I miss any details, I skimmed through the thread, but may have missed something.

Thanks so much, Matt (and Mark )

Matt - could you please tell us what sort of negative impact placing a 3rd card in the x8@x16 slot (the only one left on x399 and similar boards, after filling up the 2 full-speed, x16@x16 ones with GPU cards) has on entire GPU subsystem, and ultimately - a Resolve workstation efficiency as a whole? Do the other 2 cards (in the full x16 speed slots) have to "wait" for the slower card in the x8 slot in each cycle, or are they allowed by Resolve to work full speed, independently of the other one? Or - whenever you speak about 3 or even 4 cards - you're testing them on some other motherboard with enough full speed PCIe slots?

I perfectly understand there must be some negative impact of one of the 3 GPU cards, working in an x399 system like mine in a half-the-maximum-throughput PCIe slot - but based on your tests, is there still any point in adding a 3rd RTX 2080ti card to the existing two?

Piotr

Piotr Wozniacki wrote:

Puget_MattBach wrote:Hey everyone, Mark tossed me an email earlier to let me know about this thread. I think I can throw some light on things here! Sorry if I miss any details, I skimmed through the thread, but may have missed something.

Thanks so much, Matt (and Mark )

Matt - could you please tell us what sort of negative impact placing a 3rd card in the x8@x16 slot (the only one left on x399 and similar boards, after filling up the 2 full-speed, x16@x16 ones with GPU cards) has on entire GPU subsystem, and ultimately - a Resolve workstation efficiency as a whole? Do the other 2 cards (in the full x16 speed slots) have to "wait" for the slower card in the x8 slot in each cycle, or are they allowed by Resolve to work full speed, independently of the other one? Or - whenever you speak about 3 or even 4 cards - you're testing them on some other motherboard with enough full speed PCIe slots?

I perfectly understand there must be some negative impact of one of the 3 GPU cards, working in an x399 system like mine in a half-the-maximum-throughput PCIe slot - but based on your tests, is there still any point in adding a 3rd RTX 2080ti card to the existing two?

Piotr

part of the concept of having one beefy card instead of two ok ones, it is that you have another slot free for another beefy card....

part of the concept of having one beefy card instead of two ok ones, it is that you have another slot free for another beefy card....

So then what happens? You add another card, thus abandoning the "concept"?

IMHO, it is better to work within the CURRENT parameters of estimated performance requirements coupled with best price/performance. You can always pull the cards out later and resell them, so this boils down to economics in the end.

Seems to me the best choice at the moment was made : 2 x rtx 2080ti

Once you have two beefier cards, the bottleneck is somewhere else (cpu perhaps, or drives).

Ideally you should be able to saturate all the parts in motion without having one more loaded than another: if now you add another card and it underperformes, you're not distribute the load.

waltervolpatto wrote:Once you have two beefier cards, the bottleneck is somewhere else (cpu perhaps, or drives).

Ideally you should be able to saturate all the parts in motion without having one more loaded than another: if now you add another card and it underperformes, you're not distribute the load.

This is where skill and knowledge comes in when specing a custom machine.

Interestingly, with my previous 2x Titan Xp GPUs, I could ebale using memory above 4K by PCIe devices in the BIOS, which is recommended in Resolve Hardware Configuration Guide (in the context of Supermicro mainboard - but anyway). I have noticed that installing the two RTX 2080Ti I have now, my MSI motherboard-based machine won't even boot with this option enabled - I have to disable it (which BTW is the option's default setting).

Does anyone know what is the practical difference between using this setting enabled vs. disabled? All I can say is I remember that with the Titans, it happened randomly the PC couldn't load driver for both cards at boot-up, which was why I tried the "use memory above 4G" enabled - which helped with the problem once and for good... With the RTX 2080Ti I had to disable it, but a similar boot-up failure never happened so far...

So, what the implications are, do you think guys? Thanks

Piotr

Anyone, please?

I'm planning on getting a 3rd RTX 2080Ti card and am a bit concerned that w/o access to the memory above 4GB, it might be difficult if not impossible to boot my machine up in such a way that all 3 GPUs are granted their required resources...

BTW: for those questioning my sanity here's what Matt of Puget Systems wrote to me about pros & cons of running 3 such GPUs despite one of them only working in an x8@x16 PCIe slot (I hope Matt doesn't mind sharing the message with this great community):

"Hey Piotr,

In general, the difference between running the GPU in x8 vs x16 is very minimal. The theoretical bandwidth may be twice as much using x16, but in reality, it only equates to a few percent difference in performance in Resolve. It isn't until you get down to x4 that you would start to see a noticeable impact on performance.

I wouldn't really worry about it. It may be "best" to run all the cards at full x16, but that really isn't practical. Often, in order to do that with 3-4 GPUs means you have to use a platform that is sub-optimal for Resolve (like a dual Xeon), which results in overall lower performance than letting some of the GPUs run at x8 but on a better platform."

Hoping for an interesting discussion between those of you guys who are deeper into the PC/Windows arcane knowledge than I am .

Piotr

PS. Plus, I'm actually very satisfied with my current 2x 2080Ti performance - but better being enemy of good, why not make my system even more snappy? With the 1600W PSU and modded cooling of my PC case specifically around the PCIe slots area, the last empty one is tempting.

It seems no matter how much advice you are given, you are not satisfied. You have better hardware than 99% of the people here. Is it really preventing you from getting work done?

I am a musician and I have known people who have similar behavior. They are obsessed with new amps and guitars and it seems they rarely ever just make music.

bstaley wrote:It seems no matter how much advice you are given, you are not satisfied. You have better hardware than 99% of the people here. Is it really preventing you from getting work done?

I am a musician and I have known people who have similar behavior. They are obsessed with new amps and guitars and it seems they rarely ever just make music.

Where did I say I'm not satisfied with the advise I'm getting? From your post it sounds like trying to optimize the BIOS settings in the Resolve machine must prevent me (or anyone, for that matter) from editing and grading videos - why would you think so?

If the BIOS setting I inquired about has an optimum value and you know which it is (on or off - not too much choice here), why don't you just share your knowledge? And I'm definitely not looking for non-existing issues considering that this very setting is mentioned in the BMD official configuration guide...

Peace!

Piotr

If what you are looking for is the optimum on/off setting in the bios, why don't you just try it both ways and see which works best?

bstaley wrote:Iwhy don't you just try it both ways and see which works best?

Why don't you just read entire thread, or (if you did) try to understand what it is about?

Peace

Piotr

I understand. You'd prefer to have everybody spend their time explaining things to you, rather than spend your own time figuring it out. I wish you the best. Take care.