OT: Turing Supposedly 6X Pascal Performance!
Ghosty12
Posts: 1,979
Was going to tack this onto my other thread but since the info included is not quite the same thought best to make a new thread and not muddy the waters with the other thread topic..
Anyway onto the topic we have some official info on the new Turing architecture limited as it is and it looks to be very impressive indeed video goes a bit more indepth.. So will be interesting to see what we get from this on the consumer side.. Also added supposedly in the title after what was said in the video..
Post edited by Ghosty12 on
Daz 3D is part of 
Connect
DAZ Productions, Inc.
224 S 200 W, Suite #250
Salt Lake City, UT 84101
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2024 Daz Productions Inc. All Rights Reserved.
Comments
The devil is always in the details.
Then there is how exactly has marketing hype affected the information that has been delivered.
...interesting comments concerning Volta.
The Volta Qaudro GV-100 has 32 GB of HBM2 and 640 Tensor cores while the RTX 6000 will still have 24GB (like the P6000) and 574 Tensor cores.
It makes me wonder more if there even will be a Turing "Titan" since the RTX 5000 will have better performance and more VRAM for 700$ less than the price of the current Titan V and pretty much fill that niche (while also having NVLink support for 32 GB of pooled VRAM which the Titan-V does not).
We'll see what happens in six days.
...yeah could build one heck of a workstation for the price of that single card, even if it does have 48 GB of VRAM (though that might handle 100% of my large format scenes).
C'mon Megabucks Lotto!
For giggles: https://wccftech.com/nvidia-pascal-gpu-gtc-2015/
So with my system with 64GB of system RAM, I've found that only when I load it to over 30GB with a Studio scene will it begin to max out the VRAM on my 1080ti (ie, 10+GB).
So Im curious how much system RAM do you have that allows you to load scenes so big that they overload your GPU?
What I want is a RTX server. A bunch of RTX goodness including 8 Quaddro RTX 8000 video cards installed in a 3U rack server chassis. Can you say production rendering with Global Illumination? I could actually render some if my Space scenes that bring my meager GTX 960 4 gig card to it's knees. Just have to win the Lottery first.
i just remember how long it took nvidia to deliver iray support for the pascal cards after they shipped.
the processing boosts sound good. but this new round of cards looks like they're built to advance some new raytracing methods. will iray support still be a priority for them? dunno.
Or you could have a general Next Gen GPU Rumor Thread...
Nothing surprising here at all. I've covered Tensor cores already as being the "secret sauce" that powers RTX. I also spoke that Tensor is shaping up to be the successor to CUDA, and Nvidia pretty much stated that in a round about way. The biggest innovation since the creation of the CUDA core, that is exactly what Tensor is about. Overall, the card only offers about 25% in "traditional tasks", no surprise here. Many games don't use real time ray tracing, and this wont change overnight. It will take some time before gamers actually see the real benefits of Tensor cores.
But Iray is nothing like gaming. Iray can and will use Tensor cores to their fullest advantage. And if Turing offers a 6X preformance increase in tasks because of its Tensor cores...well you don't need to think too hard about this. It will be a massive boost for us Iray users. Another old Pascal claim in 2015 was that Pascal was over 2.5X faster in VR gaming. This was because of how well optimized the software on the cards was. This claim was in fact largely true. It obviously did not translate to all games, but for VR it made a big difference. Fast forward to now, and the claim made here is also very niche, but dude, WE ARE IN THAT NICHE.
As I posted before, the Titan V makes a massive boost over any previous Pascal in testing. It more than doubles the Iray speed of Pascal Titan and 1080ti. I think it was 2.5 times or so. So if this new Quadro can run like 4 Titan V's, then I can actually believe the Quadro 8000 can push 6X over Pascal. It makes sense. Turing is huge. The dies are physically massive as they pack all those CUDA and Tensor cores in one die.
We have to remember that Tensor cores are still very new. It makes sense that they would see a big increase in efficiency quickly. CUDA has progressed over time as well. The CUDA cores in a 680 and nowhere near as effective as the CUDA cores in a 1080. In fact a 1050ti with exactly half the CUDA cores is faster than the 680, demonstrating just how much CUDA has progressed. That took like 4 or 5 years, but CUDA was already pretty mature, having been around since 2007. Tensor only just popped up last year. Apparently Tensor has made a big leap since the Titan V released with first gen Tensor cores.
I also spoke of the rise of gaming engines. We all were impressed with the Star Wars demo. And I said that in time, this tech would be coming to desktops, and sooner than we all think. It looks like that time really is coming sooner than I think, LOL. Sure, it is a $10000 GPU, but it is a SINGLE GPU. In time this will come down to us normal people. We wont even have to wait that long. It is now absolutely certain that this will be available to desktops in just about one or two years. TOPS. Remember, Nvidia revealed $9000 Quadros not that long ago, and now they are outclassed! HA! The Titan V was released just last year, and in just a few days it will be completely outclassed. At any rate, you be sure that Disney is going to be investing heavily in this. Note the relationship with Pixar. They have every intent on using this. It will drastically speed up production for movies, and it could change how movies are made.
So with everything blown open, the only questions left are over exact VRAM and pricing. I still think we will see multple SKUs available that offer different amounts of VRAM. I believe that the 2080 will start with 8GB, but also have another option for 12, maybe 16 if all wish very hard, LOL. It would be wonderful if the NVLink is present that allows you to double your cards and their VRAM.
...I already know that you as well as others use various methods to "optimise", compress, and reduce the load for the GPU that for the works as well as size I plan to render in, would compromise the final quality. This is why I am still working on the configuration for a 128 GB (possibly 256 GB) dual Xeon render box (the latter running 8.1 Pro).
The various methods do not always compromise the final quality, some of us can't afford several high powered cards and machines and therefore learnt how to work around their system limitations personally I found it a better way to deal with things than constantly bemoaning the problem.
This..!
So we would only need get Iray to 8% convergence, then the AI/Deep Learning/Denoiser will create the rest of the render instantly?!?!
The de-noiser does need enough of the render to give it a hint of what is noise and what isn't, before or after it actually starts working.
...graininess at large render resolution is an issue and I've seen where some "shortcuts" like Render Throttle and Scene Optimiser do have an effect on the final quality. Again I am looking at large resolution output for high quality art prints. Hence the high memory high core count system for the final rendering since increasing the render size increases the memory load. I also use a number of :"in render" effects solutions that further add to the memory load as I am terrible at postwork and digital painting.
Yes I know that CPU rendering takes longer than GPU rendering, for this system, that is not an issue. I also plan to use slightly older generation components to keep the costs more in line. The "latest and greatest" is nice if you can afford it, but the "tired and true" can handle the job.