Quadro Cards, Not Dead Yet?

[Peter Moretti]

The word was you had to use a Quadro card for reliable performance with post production apps. Avid helped to promote this belief, as only Quadro cards were certified, and non-Quadro cards were disabled for GPU acceleration.

That started to change as Premiere made serious inroads into post, and people were finding better performance in their home gaming rig than their office workstation.

Enter more heavily GPU reliant programs like Resolve, and the race was to get the fastest GPU for your $, and that's always a GeForce.

Quadros were explained away by being the result of savvy NVidia marketing, good openGL performance with CAD apps with old code bases, and superior double floating precision performance. None of these should help modern post production software (which use at the most 32-bit (single precision) float).

But have we been deceived a bit? SGO seems to think so.

Here's a summary of the points they make:

1. Mistika is highly optimized to take advantage of the benefits of Quadro cards.

2. Quadro cards allow for simultaneous bi-directional communication between the GPU and RAM, while GeForce cards do not.

3. Quadro cards are more accurate than GeForce cards. So GeForce cards could add noticeable image degradation when working with high precision file formats. I'm assuming R3D, 12-bit versions of codecs like ProRes and DNxHR, Arri RAW, Cinema DNG, Canon RAW Light, etc. fall into this category.

4. Quadro cards are more reliable.

Point 3. has me the most intrigued/concerned.

Point 2. is also interesting, as most benchmarks are not performed with a Decklink card and external display attached. And it's something I've never heard before about Quadro cards.

Here's a link to the SGO whitepaper on the this topic:

https://support.sgo.es/support/solutions/articles/1000247927-nvidia-quadro-or-g-force-

 

[Jeff Kilgroe]

The Quadro cards do have some additional capabilities, unfortunately very few apps actually take advantage. And if you do run apps that can utilize the benefits of a Quadro, is there enough to justify the cost differential and occasional performance gap?

To address the points made...

1> Mystika is well optimized, yes. I would most likely choose a Quadro or Quadro+Tesla configuration for a workstation built to predominantly run Mystika.

2> Yes on the memory bandwidth, but that is only one of several factors that determine overall performance. Quadro cards are often clocked slower and are far less powerful per dollar than their GeForce counterparts. So there are trade-offs here and we can’t just go to memory bandwidth as a great reason to buy them. Memory bandwidth is rarely a bottleneck in computational settings at this time.

3> They’re reaching a bit here. Quadro cards have greater precision for floating point operations. The GeForce Titan GPUs had the same ability until this was removed with the Maxwell based Titan X and Pascal Titan Xp. However this functionality is back with the Titan V. Additionally, increased FP64 instruction precision (essentially double precision vs. single precision) has no bearing on many of the apps we in this community use. Premiere and Resolve do not benefit from additional precision. Your codecs like ProRes, DNG, Canon RAW, etc.. don’t benefit here. Mystika might, I’m not read up on SGO’s current processing abilities.

4 > Yes Quadro cards are a bit more reliable, supposedly. I haven’t burned up any GeForce cards in the past few years either so who knows... And I guess I don’t buy the no-name brands. In a nutshell the Quadro cards are clocked a bit slower and have better memory architecture. They tend to plug along a bit slower than a GeForce but also provide much more even performance as the GeForce cards can be more erratic as they’re built for raw speed. 5 year standard warranty on Quadro cards. I haven’t had to use a warranty on a GPU um, like ever. And if you rip the factory cooler off the card and connect a water block, you just voided your warranty anyway.

For most uses we encounter here on these forums, I would recommend GeForce cards. There are some apps or uses where I would definitely recommend Quadro cards. For years the nVidia marketing machine has been trying to push professionals into the Quadro line, even if it’s not the best fit for their applications and workflows. And many software companies have followed suit with that. From their perspective it’s far easier to recommend the “professional” product line that they can easily test and certify where they only have a few models from one supplier. Many of these software companies have also steered away from recommending the consumer gaming cards as they have encountered backlash from the user communities in the past. For a long time many 3D / CAD professionals refused to believe that a “gaming” card was up to the task of running the latest versions of their $25K Alias or SoftImage license...

 

[Misha Engel]

1. Mystika is highly optimized to take advantage of the benefits of Quadro cards.

2. Quadro cards allow for simultaneous bi-directional communication between the GPU and RAM, while GeForce cards do not.

3. Quadro cards are more accurate than GeForce cards. So GeForce cards could add noticeable image degradation when working with high precision file formats. I'm assuming R3D, 12-bit versions of codecs like ProRes and DNxHR, Arri RAW, Cinema DNG, Canon RAW Light, etc. fall into this category.

4. Quadro cards are more reliable.
https://support.sgo.es/support/solutions/articles/1000247927-nvidia-quadro-or-g-force-

1. So they must use 10 bit openGL overlay and/or FP64 bits, maybe I don't know.
2. Hahaha
3. BS, ask Blackmagic Design, FP32 is FP32.
4. Maybe, they are a lot more expensive.

NVidia hates this https://www.servethehome.com/deeple...rds-with-10x-nvidia-geforce-gtx-1080-ti-gpus/ because it's cheap and it works.

 

[Michael Lindsay]

interesting find..

We have found over the years that Quadra cards are more reliable and stable.. (but I accept this is very anecdotal)

We also only use certified cards in our Avids

BUT for our other applications Resolve and Redcine (where we need speed) Nvidia do not offer what we need.. They should consider something inbetwen a P6000 and a 1080ti for video.. happy they hobble it a bit but nor re raw core count or speed or memory amount..

I'd jump to buy a £2k card that is basically a Titan Xp with 16-20gb of Vram and quadra drivers (happy if they pull our double floating precision etc...)

we use Quadras anyway for our GUI so in that instance I would happily do 2 or 3 of these fictional cards in a Red processing machine..

11 and 12gb of Vram can limit nosie reduction on large images in resolve but I can't justify a P6000 ever and I wont touch the underdog AMD..

 

[Misha Engel]

and I wont touch the underdog AMD..

Why not?
VEGA FE is still the sweet spot in my opinion.
16GB VRAM, 10 bits openGL/CL overlay, professional certified drivers and only $ 1.000 a piece.

http://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128.html

I must admitt I thought the same till threadripper and VEGA FE came out.

 

[Michael Lindsay]

Hi Misha

I started to respond with a technical response but if I am honest it is because I have natural tendency to support the underdog .. this tendency has caused me to pick the less practical (though often better) option so many times... and I simply have to stop... as I have tracked down too many dead ends and wasted simply too much time... I'm the guy that backed satori fx over photoshop and Incite and discreet edit over avid;-)

Why not?
VEGA FE is still the sweet spot in my opinion.
16GB VRAM, 10 bits openGL/CL overlay, professional certified drivers and only $ 1.000 a piece.

http://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128.html

I must admitt I thought the same till threadripper and VEGA FE came out.

 

[Misha Engel]

Hi Misha

I started to respond with a technical response but if I am honest it is because I have natural tendency to support the underdog .. this tendency has caused me to pick the less practical (though often better) option so many times... and I simply have to stop... as I have tracked down too many dead ends and wasted simply too much time... I'm the guy that backed satori fx over photoshop and Incite and discreet edit over avid;-)

Well I just switched over from NVidia to 2xVEGA FE because of the price/perf for Resolve, Fusion, PremierPro, AE and PhotoShop and are pretty happy with it.
PP en AE won't be renewed in 2018 and Photoshop will be replaced by GIMP 3.0.
We have 4 intel workstations with NVidia running and now one new Threadripper with double VEGA FE, guess which one is always in use...

 

[Michael Lindsay]

Ok Misha... I will look into AMD..;-)

In a workstation with a single slot GUI card and a double slot processing option what would be the BEST amd option?

Thanks

 

[Peter Moretti]

Michael, FWIU the whole adding a dedicated GUI card to your setup is obsolete in Resolve 14.

 

[Misha Engel]

Michael, FWIU the whole adding a dedicated GUI card to your setup is obsolete in Resolve 14.

+1

That's why I use 2 GUI cards or 2 processing cards, depends how you want to look at it.
2xVEGA FE (so support for 10 bit in programs like adobe photoshop) with EKWB VEGA water block(Just hate noise).

 

[Michael Lindsay]

Hi Peter and Misha

My understanding form BM is that , outside temporal effects, when using 2+ cards to process footage they take turns frame to frame. Therefore it is only ever as fast as your slowest card. Now what I don't know is how much a card is doing when handling GUI functions because the GUI card (if the cards are identical) will be the slowest card... I am told it is not as simple as looking at GPU usage for that card as there are other factors beyond processing (bandwidth blocks, memory usage). Obvosuly it part depends on resolution of the screen but what about stuff like running waveforms etc?

Need to test but would be happy to hear of any other tests done.?..

One other note about the card you recommended .Radeon Vega FE . it really pulls a little ot of power? Is that never an issue for you guys?

Also what is the comparable speed card for resolve from Nvidia? Is it as good as a Titan Xp but with a bit more memory...

 

[Misha Engel]

Hi Peter and Misha

My understanding form BM is that , outside temporal effects, when using 2+ cards to process footage they take turns frame to frame. Therefore it is only ever as fast as your slowest card. Now what I don't know is how much a card is doing when handling GUI functions because the GUI card (if the cards are identical) will be the slowest card... I am told it is not as simple as looking at GPU usage for that card as there are other factors beyond processing (bandwidth blocks, memory usage). Obvosuly it part depends on resolution of the screen but what about stuff like running waveforms etc?

Need to test but would be happy to hear of any other tests done.?..

One other note about the card you recommended .Radeon Vega FE . it really pulls a little ot of power? Is that never an issue for you guys?

Also what is the comparable speed card for resolve from Nvidia? Is it as good as a Titan Xp but with a bit more memory...

GUI doesn't demand much in Resolve (even a simple IGP should be good enough for that).
Yes Vega does require a lot of power (300 Watt vs. 250 for a GTX 1080ti), I undervolted my VEGA's for increased speed but they still average around 300 Watt each. Speed wise you can compare them with a GTX1080ti https://www.slashcam.de/artikel/Tes...ingts--63--Vega56-unter-Resolve-14.html#Was_b haven't had the opportunity to test it myself.

The reasons I choose for VEGA FE was the 10 bit overlay, 16 GB and price. It's not the best gaming card but I don't care about that part.

There is a lot to like about the Quadro P6000, but not the price.

 

[Michael Lindsay]

Thanks Misha useful info

10bit GUI is not interesting to me as I actually like seeing a 8bit crap display at the same time as I look at our 10bit Flanders.. the 16gb is very interesting as I have choked our current setup more than a few time with NR due to insuficant memory...but at the moment we only have 8gb vram and it is difficult to work out if 11 or 12 is enough..

GUI doesn't demand much in Resolve (even a simple IGP should be good enough for that).
Yes Vega does require a lot of power (300 Watt vs. 250 for a GTX 1080ti), I undervolted my VEGA's for increased speed but they still average around 300 Watt each. Speed wise you can compare them with a GTX1080ti https://www.slashcam.de/artikel/Tes...ingts--63--Vega56-unter-Resolve-14.html#Was_b haven't had the opportunity to test it myself.

The reasons I choose for VEGA FE was the 10 bit overlay, 16 GB and price. It's not the best gaming card but I don't care about that part.

There is a lot to like about the Quadro P6000, but not the price.

 

[Misha Engel]

When it's just resolve 11 GB should be enough for 8k, when you also do a lot of blender, 3ds max, v-ray etc... you can never have enough memory, a Kingston DCP1000 with 1,6 TB might help you out (or a different sollution with the same space and speed).

 

[Dermot Shane]

i've ran Resolve projects in the last two weeks with 8,11,12,16g vram in diffrent machines.

in a Dci4k timeline there was no noticable diffrence, the machine with a single 1080 the 8g vram was not a weaklink, no issues.. so no surprise that the machine with 4 x 12g Xp's or the machine with 2x Ti's @ 11g or the machine with single aP6000 were not breathing hard, none of them threw up a gpu out of ram error, all of them played back @ 23.98 without burping

i run timelines that match my prime deliverable, and match my monitoring to that, i can see little pratical use for more than 8g

 

[Michael Lindsay]

Good info Dermot

I run timelines that match my prime deliverable, and match my monitoring to that, i can see little pratical use for more than 8g

The only process that def causes out of memory for me is.. 6k or higher clip - split into individual colour channels - different NR per channel ...NR is both spatial and temporal using the most frames and highest quality settings - also qualified per channel via a soft key to the part of the image that needs it... all on 4K Timeline...

Just wondering if 11gb VRAM would def always handle this?