RTX 2080 Ti Resolve Performance (Mathmatical Theory)

From barefeats.com we have a comparison of the GTX 1080 Ti vs the GTX 1080 here: http://barefeats.com/mvc_pascal.html

NVIDIA released their comparison of the RTX 2080 vs the GTX 1080 https://www.theverge.com/circuitbreaker/2018/8/22/17769122/nvidia-geforce-rtx-2080-performance-benchmarks-games

I did the math based upon a 15% increase of CUDA cores resulting in a 47.9% increase in performance and applied this to the RTX 2080 Ti which has 21% more cores than the GTX 1080 Ti. I did not count the DLSS increase.

View my graphics, what do you think?? Do you believe it is worth the price to upgrade?

I did the math based upon a 15% increase of CUDA cores resulting in a 47.9% increase in performance

If everything else is the same a 15% increase in cores will have a 15% increase in performance.

Unless Resolve supports using the new ray tracing and tensor core functions for video processing, which is not likely, the increase will only be 15-20% with the extra cores and larger memory bandwidth.

What I think is more imporant is the NVLink functionality, I think it is essential for Resolve to add support for this so that the total available memory can be shared among multiple GPUs.

Unless Resolve supports using the new ray tracing and tensor core functions for video processing, which is not likely, the increase will only be 15-20% with the extra cores and larger memory bandwidth.

This would be correct except NVIDIA is not using the same CUDA Core technology. They claim the new Turing TU102 is 50% faster per core (which I do not give them - I found it overall 47.9%).

This new architecture SHOULD make each CUDA core more efficient and thus more powerful. If the RTX 2080 Ti is **only** 21% faster than the GTX 1080 Ti, the frame rate would be 62.92 in Resolve with a price of $18.28 per frame - DRASTICALLY not being worth it compared to the $12.11 per frame of the GTX 1080 Ti.

I would wait for an actual Resolve benchmark, not games benchmarks, before placing an order.

Hopefully Puget Systems gets their hand on one soon to test it!

By the way a Puget Systems rep is having a 20% as well:

"Until then, how much faster will these cards be compared to the GTX 10 series? That we don't know for sure yet, but looking at the CUDA core count, clock speed, and memory bandwidth numbers for the GeForce GTX 1080 Ti versus the new RTX 2080 Ti... I would hazard a guess at around a 20% increase in base performance. Specs for the 1080 vs 2080 and 1070 vs 2070 look to be in the same ballpark. Anything gained through RTX technology will be on top of that, but only when applications and games add support for it."

Source:
https://www.pugetsystems.com/blog/2018/08/20/What-Does-the-Launch-of-the-GeForce-RTX-20-Series-Mean-for-the-Future-1224/

I tend to think that is probably going to be accurate since my math is based off the game-play increase in frame-rates and trying to port that over to Davinci.

However, again, what I have shown is that anything less than a 50% increase in Davinci Resolve is a waste of money.

I tend to think that is probably going to be accurate since my math is based off the game-play increase in frame-rates and trying to port that over to Davinci.

However, again, what I have shown is that anything less than a 50% increase in Davinci Resolve is a waste of money.

Not necessarily, there is also the potential of growth with multiple GPUs, assuming Resolve supports NVLink.

This is a statement we got from a Redshift developer in regards to the RTX cards (I believe the 2080 Ti, but it was unspecified):

Now, regarding the new RTX cards… they are fast! Well, faster than a 1080Ti, but not quite as fast as a TitanV (but costing significantly less, which is important to note here) - and this is without utilizing the new RT Cores.

So raw CUDA compute performance should see a bit of a boost, the exact general amount we’ll have to see when we either get our hands on the cards or from compute benchmarks.

Cheers,
Mike

Thanks Mike for sharing. What I see, mathematically, is that each CUDA core is significantly more efficient that the previous Pascal series. So performance should be more significant than a mere ratio of the amount of CUDA cores.

I'm curious

I'm also curious if Davinci Resolve can utilize the RayTracing engine to do something in regard to rendering or who knows??

Lot's of bla bla again.

It's quite simple, the raw compute power(FP32) what we use in Resolve is:
For NVidia 0.002 x Clock Frequency x number of Cuda cores
Titan V = 0.002 x 1200 MHz x 5120 = 12.288 TFlops (fp32)
1080ti = 0.002 x 1480 MHz x 3584 = 10.609 TFlops

AMD
VEGA FE = 0.002 x 1382 MHz x 4096 cores = 11.321 TFlops

All at base clock

How Resolve handles it is a different story

Maybe, but not in this version, it would need some serious development.

As for the ray tracing unit it looks to be a fixed function unit that does, well, ray tracing. Kinda like the texture units just do mapping of textures to geometry and nothing else.

The tensor unit appear to be more general but your data and function need to match what they do. For what they can do they tear it a new one. If that is not what you need to do, well then, SOL.

The stream processors (CUDA cores) are the general purpose parallel compute platform. Any compute that is seriously parallel can work well there. Like many typical image processing effects.

As for the ray tracing unit it looks to be a fixed function unit that does, well, ray tracing. Kinda like the texture units just do mapping of textures to geometry and nothing else.

If some data goes in and something else comes out it can probably be used for general applications too. Tracing a ray needs to know about materials and MDL allows changing material definitions so creating necessary calculations on material/shader level should allow using ray tracing for computing.

Maybe, but not in this version, it would need some serious development.

This version also uses Cuda cores. Currently Resolve uses the compute cores(fp32) to calculate everything that is done in Resolve.

RTX 2080 Ti, 4352 cuda cores, 1350 MHz base clock, 1635(FE) boost clock and 616 GB/s Mem bandwidth.

Hence:

RTX2080Ti: 11,75 TFlops base clock and 14,23 TFlops at boost clock with a 616 GB/s mem bandwidth.
Titan V : 12,29 TFlops base clock and 14,9 TFlops at boost clock with a 652 GB/s Mem bandwidth.

Titan V has more TFlops, higher mem bandwidth and less mem-latency(but cost a lot more).

I don't see how the RTX2080ti can beat the Titan V in Resolve.

You would need to overclock a 1080ti to around 2000 MHz to get the same compute performance as the RTX2080ti at boost frequency, with a waterblock this in normally no problem, you still lack the memory speed of the RTX2080ti so this will be the bottle neck for the 1080ti in Resolve.

The remaining problems of the NVidia consumer cards is the amount of Memory (8..12 GB), the more you do in Resolve the more memory you need. So for me the go to card is still the VEGA FE.

Well, today we have official gaming benchmarks. The RTX 2080 Ti is 34% faster in 4k than the GTX 1080 Ti.

Remember, the GTX 1080 had a 71% performance uplift over the GTX 980.

If you see my original graphs from August, what I found was for the price point, the GTX 2080 Ti would need to be 64% faster. So we are at HALF that

I think part of the problem here is that they have perhaps unintentionally changed their numbering scheme in the lineup, because of the huge price hike. The price of the RTX 2080 aligns it with the GTX 1080 Ti, while the RTX 2080 Ti is more like a Titan Xp. Either way, the gouging is extremely evident.

The remaining problems of the NVidia consumer cards is the amount of Memory (8..12 GB), the more you do in Resolve the more memory you need. So for me the go to card is still the VEGA FE.

I am curious about this. In 4k content, is it easy to use 11GB of RAM? How many nodes is that? NR affects it or no?

I like the 16GB of the Vega FE but have been waiting to decide what to do.

The remaining problems of the NVidia consumer cards is the amount of Memory (8..12 GB), the more you do in Resolve the more memory you need. So for me the go to card is still the VEGA FE.

I am curious about this. In 4k content, is it easy to use 11GB of RAM? How many nodes is that? NR affects it or no?

I like the 16GB of the Vega FE but have been waiting to decide what to do.

11 GB is more than enough (normally) to handle 4k content.

My rule of thumb to not run out of memory to fast:
- 4k(8.85 Mpix) 3..4 GB We have 3.5 GB(eff) GTX970's that can handle 4k with a not to heavy grade.
- 6k(19.91 Mpix) 6..8 GB
- 8k(35.39 Mpix) 12..16 GB

I don't know what AMD will do but the VEGA FE is according to my knowledge out of production (to much value for money for content creators). It can do 10 bits color in OpenCL.(competes with the NVidia quadro P6000 $4,000).

I've never run out at 4K, but as Misha says, if I add TNR/OFC to 8K footage on an 8K clip beyond a tiny amount, instant out of memory error with 11GB.

Simple 4k is no problem for 11GB but I constantly run out of memory with clips on a 4096x1716 timeline because I'm working with material that is 4.6k to 6k and often has NR and ofx applied.

Simple 4k is no problem for 11GB but I constantly run out of memory with clips on a 4096x1716 timeline because I'm working with material that is 4.6k to 6k and often has NR and ofx applied.

Then the RTX is not the best option for your workflow.

Simple 4k is no problem for 11GB but I constantly run out of memory with clips on a 4096x1716 timeline because I'm working with material that is 4.6k to 6k and often has NR and ofx applied.

Then the RTX is not the best option for your workflow.

Yes. I know....sigh...and almost no new camera is shooting at less than 4.5k so looking forward, the RTX is going to create GPU memory errors for anyone working the mid to higher end. There's a hole in nVidia's line that's twice the cost to fill.

Yes. I know....sigh...and almost no new camera is shooting at less than 4.5k so looking forward, the RTX is going to create GPU memory errors for anyone working the mid to higher end. There's a hole in nVidia's line that's twice the cost to fill.

Nobody is forcing you to buy NVidia.
The quadro version of the RTX2080ti (well sort of) the RTX 6000 has 24 GB for only $6.300.

So please inform what does AMD offer that doesn't require water cooling that's a bargain in comparison to the Nvidia Quadros. BTW there's a much cheaper 5000 version of the Quadrro RTX cards $2300 that has 16GB of memory.

So please inform what does AMD offer that doesn't require water cooling that's a bargain in comparison to the Nvidia Quadros. BTW there's a much cheaper 5000 version of the Quadrro RTX cards $2300 that has 16GB of memory.

Yep that one is a bit faster than the RTX2080 and has slow memory 450 GB/s
About 11 TFlops fp32.

An undervolted VEGA FE will be around the same speed on air for around $1.000 at 1350 MHz (base speed is 1382 MHz, boost speed is 1600 MHz).

I recently purchased a second 1080 TI on eBay for $450, I can't say I regret it, it certainly beats spending $1200 for a 2080 TI.

If prices keep going down for the 1080 TI, I might want to consider a third one, after all, there is plenty of inventory.

I recently purchased a second 1080 TI on eBay for $450, I can't say I regret it, it certainly beats spending $1200 for a 2080 TI.

If prices keep going down for the 1080 TI, I might want to consider a third one, after all, there is plenty of inventory.

Purchased a third VEGA FE my self, now running in one of our older workstations(GTX970 broke down).

Here you can see what you can expect (speedwise).

https://www.pugetsystems.com/labs/articles/DaVinci-Resolve-14-GPU-Scaling-Core-i9-vs-Xeon-W-vs-Dual-Xeon-SP-1121/

RED 8k.R3D 9:1 25fps is a peace of cake with a threadripper at 4 GHz all-core
RED 8k.R3D 7:1 24 fps runs at around 22..24 fps CPU limited.

The more effects you use the more the memory bandwidth of the GTX1080ti will limit you.

Hi all,
Still kinda new to editing but have been building systems on and off for a long time. I recently started using the Davinci studio version for multi GPU capability.

Currently using a HP Omen X computer and I7 7820x CPU. I've upgraded up to 32 Gb ram and running dual EVGA GTX 1080 hybrid FTW water cooled GPUs. I have the both GPUs and the CPU max overclocked and 3D benchmarks that put me in the extreme class. So far I've had little problem doing 4k, though it sometime can still be painfully slow to render and compile. But now I've had a motherboard failure so have to upgrade that. But that got me to thinking about just building a whole new upgraded system and here's my thoughts.

Looking at the new 32 core, 64 thread AMD Threadripper 2 set on an EATX board with up to 128Gb ram. (Xeon W-3175X would be my first choice, but over $10K? Big nope!) Looking into the newly announced EVGA RTX 2080Ti FTW3 Copper GPUs, which I would start with a pair of and maybe eventually go up to 4 of them once prices come down. Of course it would all be on a dual custom loop cooling system so it'll look as spectacular as it runs.

But the question I have is with the NV link. I don't doubt Davinci will eventually capitalize on it, but I'm trying to figure out if there's plans to enable the ability of a 3 way or 4 way bridge like Nvidea they did with SLI. With the AMD chip, there isn't any issue with PCIe lanes even with 4 GPUs (64 lanes). Thoughts or rumors heard anyone?