AMD Radeon RX 6700 XT 12 GB to $479 from March 18th.

Hi.

Here is a tweet from Lisa Su CEO of AMD:

Quote: 'Getting ready for our launch watching “Big Navi” in action. Can’t wait to share
@Radeon RDNA2 architecture and RX 6000 series with everyone
- join us Wednesday, Oct 28th at 12pm ET #GameOnAMD'

From: https://twitter.com/LisaSu/status/13208 ... ve-here%2F

Regards Carsten.

So far NAVI hasn't been the success that many had hoped for and I doubt RDNA2 will be able to trump NVIDIA's new line of very cost-effective GPUs.

Of course all competition is welcomed and a good thing. I just wouldn't hold my breath for anything other than an announcement that keeps the AMD fanbois online.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

Navi 5700XT has performed about on par in Resolve with Vega 56 across platforms. That's with the resolution at 4K. At 8K, Vega reaps the benefits of high-bandwidth memory and leaves the Navi to rot. Going from that, Big Navi faces a steep cliff in reaching the power of Radeon VII (or high-end Turings). It may get there at 4K, but beyond that, we shall see.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

They already did, you won't run out of memory with an AMD card.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

They already did, you won't run out of memory with an AMD card.

Yes,but Misha, that's what I thought when I tried the Radeon 7. It's true but the system ram that's used to "add on" to the vram is a LOT slower, so yes you don't run out of GPU memory but the speed of GPU operations is, I don't want to say snail-paced, but it is much, much, slower. That said my RTX 3090 has 24GB of vram and is very fast indeed, and with the sale of my 2 1080ti's cost half its retail price.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

They already did, you won't run out of memory with an AMD card.

Yes,but Misha, that's what I thought when I tried the Radeon 7. It's true but the system ram that's used to "add on" to the vram is a LOT slower, so yes you don't run out of GPU memory but the speed of GPU operations is, I don't want to say snail-paced, but it is much, much, slower. That said my RTX 3090 has 24GB of vram and is very fast indeed, and with the sale of my 2 1080ti's cost half its retail price.

It doesn't use the system RAM (atleast not with 8k source material) and the radeon VII is fast with among others TNR (it has the highest RAM-speed).
This is about the biggest problem with NVidia and Resolve https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=124069. NVidia/CUDA has some kind of double buffering that eat's memory for breakfast, lunch, dinner and the rest of the day.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

They already did, you won't run out of memory with an AMD card.

Yes,but Misha, that's what I thought when I tried the Radeon 7. It's true but the system ram that's used to "add on" to the vram is a LOT slower, so yes you don't run out of GPU memory but the speed of GPU operations is, I don't want to say snail-paced, but it is much, much, slower. That said my RTX 3090 has 24GB of vram and is very fast indeed, and with the sale of my 2 1080ti's cost half its retail price.

sorry david, how the 390 compare to the two 1080ti?

sorry david, how the 390 compare to the two 1080ti?

I'm finding it between 50 and 100 percent faster on GPU intensive cache renders. Also about the same speed difference outputing an 4k nVidia encoded H265. As I'm doing a lot of 4k cache renders from my fixed node tree in the Color Page, the speed and lack of GPU Full errors is terrific.

But I do wonder if Resolve, the nVidia driver, and the EVGA bios have yet been fine tuned to take full advantage of the chip. That said the card (an EVGA XOC Ultra) runs at about 1950MHz in Resolve ops without OCing.

Hi.

Lisa Su just introduced the following AMD Graphics Cards:

Radeon RX 6900 XT Dec 8th at $999.

RX 6800 XT - Nov 18th at $649.

Also on November 18th RX 6800 at $579.

From: https://www.anandtech.com/show/16201/am ... n-1600-utc

And: https://www.tomshardware.com/news/amd-r ... s-revealed

Now we are just waiting for the Resolve Reviews.

Regards Carsten.

Would be interesting to find out if Resolve can take advantage of the Smart Memory Access too when paired with a new AMD system.

Navi 5700XT has performed about on par in Resolve with Vega 56 across platforms. That's with the resolution at 4K. At 8K, Vega reaps the benefits of high-bandwidth memory and leaves the Navi to rot. Going from that, Big Navi faces a steep cliff in reaching the power of Radeon VII (or high-end Turings). It may get there at 4K, but beyond that, we shall see.

My experience with the 5700XT in Resolve Studio has been... extremely lackluster.

It's certainly not competitive with Nvidia - even cheaper Nvidia GPUs (like the OG 2060) trounce it. And it Resolve has been a lot buggier with AMD than Nvidia in my experience.

I own machines with both platforms (Full AMD and Intel + Nvidia).

Even with a weaker CPU in the other machine, I am getting better performance with the weaker Nvidia GPU because it is just better than the 5700XT in this software.

I don't see why anyone using Resolve Studio should even consider Big Navi, considering the competitive pricing on the RTX 30 series.

My experience with the 5700XT in Resolve Studio has been... extremely lackluster.

It's certainly not competitive with Nvidia - even cheaper Nvidia GPUs (like the OG 2060) trounce it. And it Resolve has been a lot buggier with AMD than Nvidia in my experience.

I own machines with both platforms (Full AMD and Intel + Nvidia).

Even with a weaker CPU in the other machine, I am getting better performance with the weaker Nvidia GPU because it is just better than the 5700XT in this software.

Cuda tasks?

Navi 5700XT has performed about on par in Resolve with Vega 56 across platforms. That's with the resolution at 4K. At 8K, Vega reaps the benefits of high-bandwidth memory and leaves the Navi to rot. Going from that, Big Navi faces a steep cliff in reaching the power of Radeon VII (or high-end Turings). It may get there at 4K, but beyond that, we shall see.

My experience with the 5700XT in Resolve Studio has been... extremely lackluster.

It's certainly not competitive with Nvidia - even cheaper Nvidia GPUs (like the OG 2060) trounce it. And it Resolve has been a lot buggier with AMD than Nvidia in my experience.

I own machines with both platforms (Full AMD and Intel + Nvidia).

Even with a weaker CPU in the other machine, I am getting better performance with the weaker Nvidia GPU because it is just better than the 5700XT in this software.

I don't see why anyone using Resolve Studio should even consider Big Navi, considering the competitive pricing on the RTX 30 series.

I wait for the reviews and the comments on this forum.

Comments on this forum so far concerning the RTX 30xx series: they are fast and still run out of memory, just like the previous RTX and GTX series.
No reviews on the RX6000 series.

My experience with the 5700XT in Resolve Studio has been... extremely lackluster.

When AMD developed RDNA, it was seriously holding back because it couldn't afford not to. That was a very conservative design. I don't think that AMD was expecting to gain nearly as much market share as it did since, because no one in their right mind would have expected Intel to faceplant so hard on process tech. Add in the two console contracts, especially since both Sony and Microsoft decided that they want to go after serious gamers, and the resource situation changes a bit.

It's certainly not competitive with Nvidia - even cheaper Nvidia GPUs (like the OG 2060) trounce it. And it Resolve has been a lot buggier with AMD than Nvidia in my experience.

That's been my experience. Not the bugginess, the AMD GPU was reliable, but the nVidia GPU steamrolled it in perf.

I don't see why anyone using Resolve Studio should even consider Big Navi, considering the competitive pricing on the RTX 30 series.

The 3090 will probably still outperform the Big Navi, but that's also significantly more expensive. It sounds like preliminary testing of the 6900 is exceeding expectations, and may even be able to match the 3090 in some areas, at least in gaming.

The catch is that most of the 3000 series are tough to come because of the Black Magic Syndrome (TM) -- too many people want them, not enough exist for them to all have them yet.

And they cost more.

I've been predicting a major loss of market share for nVidia even if the Big Navi is only able to match the 3070, and I still thank that's correct, but it sounds like the 6900 is doing significantly better than expected.

I think the real issues will be the lack of CUDA support (OFX plug-ins etc) and the proven fragiity of AMD's drivers.

How long before they even sorted the thumbnail issue for the 5700 cards?

Those who want to play it safe (ie: professionals who can't afford to waste time chasing bugs) will stick with NVIDIA thanks to its proven track record. Those who are AMD fanbois or who just want to take a punt that AMD have got it right this time, will opt for the new RDNA2 cards.

Me? Well my current rig works just fine for the workflows I'm throwing at it so I'll sit on the sideline and watch until the smoke clears and we can tell for sure which way the winds are blowing.

AMD will really have to pull a rabbit out of the hat if they're going to

Hi.

Both AMD Radeon RX 6900 XT and 6800 XT have a Transistor Count of 26.8B

The RTX 3090 and RTX 3080 have a Transistor Count of 28B, but Samsung 8nm process node is actually a refined 10nm process. The new Radion Graphics Card are still 7nm, and will be able to run with higher clocks, so it will be difficult to tell which will be fastest.

Quote: 'Last, but not least, we have the mysterious Infinity Cache.....
it can be thought of as a local cache that buffers against reads and writes to the main memory, and provides a backstop for large operations that’s a lot faster than having to go out to VRAM.....
it would appear that AMD has dedicated a surprisingly large portion of their transistor budget to it. Doing some quick paper napkin math and assuming AMD is using standard 6T SRAM, Navi 21’s Infinity Cache would be at least 6 billion transistors in size, which is a significant number of transistors even on TSMC’s 7nm process.'

More in the Infinity Cache section of this link:

https://www.anandtech.com/show/16202/am ... ber-18th/2

What do you think of the Infinity Cache. And can it be the rabbit AMD have to pull out of the hat?

Regards Carsten.

What do you think of the Infinity Cache. And can it be the rabbit AMD have to pull out of the hat?

Regards Carsten.

It could be, but software will need to be optimized to use it properly.

What do you think of the Infinity Cache. And can it be the rabbit AMD have to pull out of the hat?

Regards Carsten.

It could be, but software will need to be optimized to use it properly.

No it doesn't, that is the beauty of Cache memory.

All current GPU's use Cache memory, the new RX6000 series just uses more and on another level compared to some other consumer GPU's (level 3 analog to CPU's).

Actual practical benefit of caching is highly dependent on work load type.

So far NAVI hasn't been the success that many had hoped for and I doubt RDNA2 will be able to trump NVIDIA's new line of very cost-effective GPUs.

Of course all competition is welcomed and a good thing. I just wouldn't hold my breath for anything other than an announcement that keeps the AMD fanbois online.

AMD will really have to pull a rabbit out of the hat if they're going to lure me away from the 3000 series as my next GPU.

The only reason consumers perceive the current lineup of Ampere cards as "cost-effective" is because NVIDIA jacked up Turing prices for no good reason except to line their coffers. Ampere offerings totally destroy the 2080ti in terms of price / performance, but that doesn't mean they are cheap. I still remember when the a flagship graphics card cost $500, which wasn't very long ago (1080 ti).

As far as pulling a rabbit out of the hat, it looks like they have done just that: a Turing-competitive card for less money that uses less energy. Whether it will be equally cost-effective for Resolve editing is another story.

I'm curious to see how they perform in Resolve. On paper, the memory bandwidth is unimpressive, but I have no idea how well the cache will work for us. I would love to be wrong.

Are there any parts of Resolve that are Nvidia RTX specifc and won't work with an RX 6800 ?

The neural engine might be, if it relies on tensor cores. There isn't an equivalent specific hardware in AMD cards afaik, but I might be wrong... Whan tensor core is is a specialized hardware for doing matrix calculations, which machine learning heavily relies on. If there isn't a hardware unit to do this, these calculations are moved to ordinary compute cores, so it works, but slower. The speed difference between doing matrix mults in tensor core vs gpgpu is pretty huge, in order of tens or even hundreds.

Regarding raytracing and specific rtx cores, there isn't anything in Resolve or Fusion that uses that.

What features use that neural engine?

The announcement back then said
"speed warp motion estimation for retiming, super scale for up-scaling footage, auto color and color matching, facial recognition and more"

I'm concerned about the "and more".
Of the named features I have absolutely no use for "facial recognition" (doesn't work with puppets)
but upscaling and sometimes auto color matching aren't exactly unimportant things.

The question is rather, what exactly does neural engine do. There is no tech info about it, just PR.

What features use that neural engine?

The first time Grant demoed the neural engine in action, it was for the object removal feature. I don't know whether any other features use the neural engine yet.

Sent from my H1T1000 using Tapatalk

Hi.

Quote: 'The 128 MB Infinity Cache boosts the standard 512 GB/s bandwidth by 3.25x, delivering an effective bandwidth of up to 1.664 TB/s across all Big Navi GPU based graphics cards.'

From: https://3dwarrior.net/amd-radeon-rx-690 ... hics-card/

I have today seen the same information on two different internet sites. I am not sure of its origin. It looks as it comes from a slide. So I have publish it as a Rumor. But I expect it to be for one Game only. And I expect it to be for the game, with the highest number?

Regards Carsten.

Hi.

I now found where the '128MB Infinity Cache times 3.25x gives 1664 GB/s Effective Bandwidth' come from. It come from an AMD webpage. Here is a link:

https://www.amd.com/en/technologies/rdna-2

'Acting as a massive bandwidth amplifier, AMD Infinity Cache delivers up to 3.25x the effective bandwidth of 256-bit 16 Gbps GDDR6'

Let hope that Resolve can take full advantage of it.

Regards Carsten.

It looks like that Infinity Cache is basically a high speed embeded DRAM cache with DMA support, and it's also a power saver as well as bandwidth booster. The ray tracing accelerators will be using it for BVH storage, so they won't share bandwidth with the compute cores.

Also interesting, and potentially actually MORE interesting for color grading software, is the Direct Storage support and Smart Access Memory. The first is a Microsoft API, so widely usable, the second is an AMD only thing -- and it requires both a Zen3 and an RDNA2. It provides the CPU fine grained access to the GPU memory. That's a potential ace in the hole for AMD, because nVidia can do nothing about it, while Intel is the GPU underdog and Zen3 is going to pretty easily outperform Intel's CPUs when it's released into the wild.

Wonder if this works with Threadripper as well as the 5000 series.

render glitches on amd 5700.. had to switch over to 2060 super of Nvidia.. at least no glitches, but GPU memory full errors galore...

Wonder if this works with Threadripper as well as the 5000 series.

The Infinity Cache yes, but the new smart cache AFAIK requires a Zen3, so 5000 series onwards.

Maybe when the Zen 3 Threadrippers come out.

Regarding the 3080 having only 10GB VRAM, can I ask, what sort of Davinci, particularly Fusion work, would run up against the 10GB VRAM limit?

For my videos, it will be entirely 4K, heavily reliant on 2D Fusion effects start to finish through the entire hour-long videos - just basic color correction under studio lighting, with no color-grading, and no noise reduction.

While I'm attracted to the AMD's higher 16GB VRAM because of Fusion, nevertheless I hear so many comments about Nvidia being better for Davinci. Also, I really want to use Nvidia's Omniverse, and so I want to know how bad the 10GB VRAM limit of the 3080 will be.

Regarding the 3080 having only 10GB VRAM, can I ask, what sort of Davinci, particularly Fusion work, would run up against the 10GB VRAM limit?

For my videos, it will be entirely 4K, heavily reliant on 2D Fusion effects start to finish through the entire hour-long videos - just basic color correction under studio lighting, with no color-grading, and no noise reduction.

Well I do exactly what you describe (albeit *WITH* some temporal noise reduction, stabilization and I use VFX-connect to Fusion 9 Studio for my Fusion work) and I get by very easily with 6GB of VRAM under Resolve 15.3. I believe Resolve 16 is a lot more demanding of VRAM but there's nothing in verison 16 that I "must have" so I stick with what works and benefit accordingly.

I think I'll wait for the higher-VRAM versions of the 3000 series cards before investing in either a 3070 or 3080 and will upgrade to Resolve 16 only then.

I think I'll wait for the higher-VRAM versions of the 3000 series cards before investing in either a 3070 or 3080 and will upgrade to Resolve 16 only then.

They were cancelled, according to recent leaks.

I think the real issues will be the lack of CUDA support (OFX plug-ins etc) and the proven fragiity of AMD's drivers.

How long before they even sorted the thumbnail issue for the 5700 cards?

Those who want to play it safe (ie: professionals who can't afford to waste time chasing bugs) will stick with NVIDIA thanks to its proven track record. Those who are AMD fanbois or who just want to take a punt that AMD have got it right this time, will opt for the new RDNA2 cards.

Me? Well my current rig works just fine for the workflows I'm throwing at it so I'll sit on the sideline and watch until the smoke clears and we can tell for sure which way the winds are blowing.

Yea, I'm going Nvidia. I'm not willing to risk more hundreds on AMD.

render glitches on amd 5700.. had to switch over to 2060 super of Nvidia.. at least no glitches, but GPU memory full errors galore...

Did the prevalence of glitches increase with the complexity of the video, such as complex Fusion compositions - or did the glitches happen irrespective of the simplicity or complexity of the task?

I also got my Strix 3090 OC and tested it with resolve. The performance uplift over the 2080 ti is insane...even for non 4K or 8K work!

I wonder how much that huge cache (or any cache for that matter) actually affects Resolve type of processing where most of the "action" is related to loading image data into gpu and then throwing it away. I think all the cuda/opencl kernels are loaded into gpu memory and not discarded, so what kind of data would actually benefit in this context from the 128MB cache? Noise reduction and other temporal effects maybe...?

I think all the cuda/opencl kernels are loaded into gpu memory and not discarded

Hi.

First time some cuda/opencl code is used, is it loaded from vRam and into the GPU and executed.
All the following times Resolve want to use the same code, it is loaded from the 128 MB Infinity Cache,
if it still is stored in the Cache. Larger cache, will give higher possibilities for recent cuda/opencl
code, to still be in the Cache.

Regards Carsten.

My question was, isn't this the case always, aren't kernels stored in their own dedicated cache? In Resolve, I think there is no kernel swapping during the processing...

What features use that neural engine?

The announcement back then said
"speed warp motion estimation for retiming, super scale for up-scaling footage, auto color and color matching, facial recognition and more"

I'm concerned about the "and more".
Of the named features I have absolutely no use for "facial recognition" (doesn't work with puppets)
but upscaling and sometimes auto color matching aren't exactly unimportant things.

So.. is the conclusion then that a 10GB Nvidia GPU will perform better than a comparable 16GB AMD GPU due to the lack of Tensor Cores if using superscaling/retiming and the above features, and also NR?

Can the new AMD FidelityFX tech be used for superscaling?

So.. is the conclusion then that a 10GB Nvidia GPU will perform better than a comparable 16GB AMD GPU due to the lack of Tensor Cores if using superscaling/retiming and the above features, and also NR?

Generally the nVidia cards outperform the AMD cards because they've been faster. AMD just hasn't been competitive, so a lot of companies have been optimizing their CUDA implementations.

The Tensor cores do enhance performance for a lot of AI functionality without using the general purpose compute cores. The Zen3 + RDNA2 combination looks very promising though.

Can the new AMD FidelityFX tech be used for superscaling?

I don't think that enough developers have gotten their hands on it to answer that yet.

Can the new AMD FidelityFX tech be used for superscaling?

I don't think that enough developers have gotten their hands on it to answer that yet.

But AMD would never add such crappy features that nVidia uses to its cards, like superscaling or variable rate shading... oh wait..

I don't think that enough developers have gotten their hands on it to answer that yet.

The AMD ProRender libraries have been out for ages now https://gpuopen.com/radeon-prorender-suite/. But I am not sure how many of the features are useful only for gaming only or rendering apps like Blender versus being actually useful for a video editing app like DVR. Maybe Fusion could find some use for it? There appears to be some API/library for Denoising as well which non-Fusion features of DVR could use I think. Unless I missed it, I did not hear any announcement from BMD about optimizing DVR for AMD technologies in the past.

The new FidelityFX also has its libraries https://www.amd.com/en/technologies/radeon-software-fidelityfx which appear to include VRS and Image Upscaling.

I guess the developers at BMD will need to move from their CUDA skillset and expand it to the AMD set of technologies in order to optimize it for the RDNA GPUs.

The 16GB AMD GPUs are compelling value at a lesser price point and I hope BMD will work to make the software more performant universally across both Nvidia and AMD platforms.

I did not hear any announcement from BMD about optimizing DVR for AMD technologies in the past.

Well, the Blackmagic eGPUs all had AMD cards in it.

NVidia is backed by a lot of money raised by gamers, AMD is now backed by Microsoft and Sony for their GPU's. The money needed for this support is also raised by gamers.
AMD has never been in a better situation than it is now.
AMD GPU's on MacOS where already pretty good and since they got hugh support from Microsoft the gap with NVidia will be closed in the near future.

The AMD ProRender libraries have been out for ages now https://gpuopen.com/radeon-prorender-suite/.

ProRender is designed for 3D rendering, so I doubt that much of it is directly applicable here.

But I am not sure how many of the features are useful only for gaming only or rendering apps like Blender versus being actually useful for a video editing app like DVR.

Gaming? ProRender? Wrong tool for the job; it's not designed for realtime graphics. Its interactive renderer is fast, but not even close to acceptable for gaming. It's designed for production rendering.

Maybe Fusion could find some use for it? There appears to be some API/library for Denoising as well which non-Fusion features of DVR could use I think. Unless I missed it, I did not hear any announcement from BMD about optimizing DVR for AMD technologies in the past.

BMD did optimize for Metal. The de-noising libraries might start getting wider use as the libraries become more cross platform.

[qoute]
The new FidelityFX also has its libraries https://www.amd.com/en/technologies/radeon-software-fidelityfx which appear to include VRS and Image Upscaling.

It's brand new. Downloading a library doesn't make it work, it still takes time to learn how to use it. I do that kind of thing for a living on a regular basis, though for a different market.

I guess the developers at BMD will need to move from their CUDA skillset and expand it to the AMD set of technologies in order to optimize it for the RDNA GPUs.

The 16GB AMD GPUs are compelling value at a lesser price point and I hope BMD will work to make the software more performant universally across both Nvidia and AMD platforms.

BMD has been, but it can't decide what everyone else does, and "everyone else" includes a lot of plugin developers, many of which have much smaller teams than BMD and can't afford to spread themselves as thinly as BMD can. (Except Adobe, which is actually a large company, but it has other... inefficiencies.)

Hi.

Here are the first AMD Radeon RX 6800 XT and RX 6800 Reviews:

https://www.tomshardware.com/news/the-a ... 800-review

https://www.theverge.com/21571414/amd-r ... -xt-review

Not all forum members will prefer the AMD reference design. Here is a link with pictures of the first versions, from AMD Graphics Cards partners. More cards will follow.

ASRock and PowerColor Roll Out Custom RX 6800 and RX 6800XT Designs:

https://www.tomshardware.com/news/asroc ... xt-designs

Regards Carsten.