AMD RX 7900 XTX, 24 GB, 61 teraflops for $999 on December 13

More info:

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

Regards Carsten.

How many CUDA cores?

How many CUDA cores?

What?

CUDA cores is the name nvidia chose for their shaders. AMD uses stream processors. And the XTX has 12288 FP32 cores (6144sps).

What matters in the end is the FP32 TFLOPs number which is 61. Considering a 6800XT with its 21Tflops and much lower bandwidth is better or at worst equivalent to a 30Tflops 3080, the XTX should roughly match or beat the 4090 in Resolve at almost half the street price.

Bruce's point, one that I agree with, is that CUDA is the more mature and better technology.

So the specs of a new AMD card are simply...irrelevant.

Here are some slides related to video editing:

Bruce's point, one that I agree with, is that CUDA is the more mature and better technology.

So the specs of a new AMD card are simply...irrelevant.

In the context of Resolve, you have no idea what you're talking about. Might be irrelevant to you, but don't generalize this.

I, for one, am really looking forward to seeing the performance reports for Resolve and Fusion on this card!

I just hope there aren't the same driver issues that plaged the 5700 and other recent AMD GPU releases.

Considering a 6800XT with its 21Tflops and much lower bandwidth is better or at worst equivalent to a 30Tflops 3080

That's not what Puget says: https://www.pugetsystems.com/labs/artic ... ance-1990/

Except if you compare in an Mac environment with Metal?

Considering a 6800XT with its 21Tflops and much lower bandwidth is better or at worst equivalent to a 30Tflops 3080

That's not what Puget says: https://www.pugetsystems.com/labs/artic ... ance-1990/

Except if you compare in an Mac environment with Metal?

Puget tests include Fusion which drops the score. Even they admitted that your workflow dictates the performance and their overall score is not indicative of this.

In RAW/ProRes/H265 its better / equivalent. In some other cases (Fusion, H264) its slower. So it really depends.

And my own testing supports this. I get 112fps using their materials (puget files) when transcoding 4K H264 to 4KH264 vs their 60 result. I get 20+fps when using triple NR nodes while they get 4...

So their testing is ...irrelevant.

I just hope there aren't the same driver issues that plaged the 5700 and other recent AMD GPU releases.

The 6000 series drivers are solid.

Most of the 5000 series issues stemmed from people refusing to turn off XMP or lower their aggressive RAM OC which caused system instability and the driver happened to be the first to drop / black screen. People that knew what they're doing had no problems with 5700 cards.

They announced AV1 encode/Decode hardware support. Don't really care about that. I would LOVE if they added H265 422 decode support though. I didn't see this confirmed/denied, so I still pray.


Nvidia couldn't be bothered with the 40 series.

Strange to consider buying an Intel GPU to pair with an Nvidia GPU just to get good timeline performance without proxies.

This Spanish site says yes without elaborating or linking to the statement.
https://news.eseuro.com/technology/1123555.html

AV1 encoding is interesting but does Youtube accept such files without re-encoding?

How’s AMDs current state with machine learning application? They have their own analogy for Tensor cores I believe but anyone knows if it actually works with all the common frameworks like Torch, Tensorflow etc? And how does it do in Resolve with ML tools?

Good question. I'll post if i find anything.

AV1 encoding is interesting but does Youtube accept such files without re-encoding?

No, if you send a 100mb AV1 file to YT it is going to compress it down to save bandwidth.

AV1 uploads may make sense for streamers who can get a better quality video with the limited bandwidth (8mb) that twitch allows.

Bruce's point, one that I agree with, is that CUDA is the more mature and better technology.

So the specs of a new AMD card are simply...irrelevant.

What utter nonsense.
An RTX 3060 has CUDA. A 7900XTX has 'less mature and not as good technology". But blows the pants off a silly 3060.
Core-to-core it's a different matter of course but the specs are most certainly not "irrelevant". If the AMD card had 5x the cores, it would be the one to get. Those, my friend, are specs.

What utter nonsense.
An RTX 3060 has CUDA. A 7900XTX has 'less mature and not as good technology". But blows the pants off a silly 3060.
Core-to-core it's a different matter of course but the specs are most certainly not "irrelevant". If the AMD card had 5x the cores, it would be the one to get. Those, my friend, are specs.

Don't mind them. Some of the people here have a deep lack of understanding how Resolve works. On the stuff that matters (raw, h264, h265) AMD cards work as good or better than the nvidia counterparts. Also no GPU memory full errors...

The 7900XTX is also the fastest h265 card as of yesterday (handbrake test where it beats a 4090 by quite the margin, link was posted somewhere on the amd reddit).

The other niche cases (like some bugs with fusion or some obscure plugin that was made for cuda only ) are the irrelevant ones...

When more software will implement HIP the perception will change.

Buy what suits you best, ignore the naysayers.

Well maybe...

But keep in mind there is a history of poor AMD performance/reliability explicitly with Resolve.

I am willing to keep an open mind, but have been burned before so will be very cautious before going red again. The price/performance seems like a better deal, but I don't want a giant technical hassle just to save a couple hundred bucks.

You guys seem like committed fans, so please buy and report you findings. Cheers

Also no GPU memory full errors...

The 7900XTX is also the fastest h265 card as of yesterday (handbrake test where it beats a 4090 by quite the margin, link was posted somewhere on the amd reddit)..

I get those memory full errors even on my 3090.. I thought something was buggy with the card haha.

I really wish somewhere there was a test done with the SPherical Stabiliser tool because *that* would determine in an instant what suits me best.

AV1 encoding is interesting but does Youtube accept such files without re-encoding?

No, if you send a 100mb AV1 file to YT it is going to compress it down to save bandwidth.

AV1 uploads may make sense for streamers who can get a better quality video with the limited bandwidth (8mb) that twitch allows.

AV1 doesn't automatically guarantee you best quality (specially one from GPU), specially when you do just HD.
Mature software h265 (or even x264) may end up even better. It depends how much effort they put into those AV1 encoders. They are very early versions, so I doubt they are that well polished.

Anyone have a link to the decoder/encoder specs?

I am looking for decode support for HEVC 10 bit 422

I think you'll need a modern Intel iGPU or ARC card to get 422 support in HVENC. As far as I'm aware, NVIDIA doesn't offer that yet.

[ On the stuff that matters (raw, h264, h265) AMD cards work as good or better than the nvidia counterparts. Also no GPU memory full errors...

The 7900XTX is also the fastest h265 card as of yesterday (handbrake test where it beats a 4090 by quite the margin, link was posted somewhere on the amd reddit).

Cool

But keep in mind there is a history of poor AMD performance/reliability explicitly with Resolve.

I heard that this was an issue that was reasonably problematic with the 5-series AMD cards (specifically the drivers) but that the 6-series AMD cards were much more stable with only rare problems.

Can any AMD users confirm?

I'd love a 7900XTX but realistically -

I'm looking to upgrade my 5700XT to something in the late 30 series nVidia (ie 3080 or above) or 6-series AMD. I'd welcome advice as to which would be the better option for those slightly older cards.

I'm just a 'home user' not pro so don't NEED a top of the line GPU but I figure as people upgrade I might be able to pick up an 'older' card for a decent price.

I never had a problem with my two Radeon 580 cards, but that was under MacOS.

Everything I've watched and read has said that NVidia is better than AMD for hardware encoding quality, for driver quality and for software support, has this changed? I have no skin in this game and will use whichever brand is best but all my research has pointed to NVidia being better so that's what I've been using over the years. Likewise I just watched a video from Eposvox on Youtube that mentions how in the Netflix quality test, NVidia (and even Intel Arc) is still somewhat better for encoding than AMD's latest.

Everything I've watched and read has said that NVidia is better than AMD for hardware encoding quality, for driver quality and for software support, has this changed?

Not really.

Good morning.
I saw on the net that the first bench related to daVinci Resolve was leaked, until now I had only noticed tests for games.



To see the first graphs the 4080 is doing quite well despite 16 gB of Vram , evidently it is not only the amount of memory that makes the difference.
As for the price as of today's date (Italy) the Radeon 7900 XTX sells for about 1300 euros versus 1400 for the Geforce RTX 4080.




Translated with http://www.DeepL.com/Translator (free version)

Amount of memory only makes any difference if you run out of it.

As it has been stated over and over again regarding AMD cards (even from puget members on reddit)...the puget bench overall extended score is not a good indication of performance.

(Also the XTX loses to the XT...yeah no...)

You need to look at your own workflow and then look at the individual scores in their tests.(even though their tests are only valid for like 1 timeline 1 node export type of deal)
.
Eg: If your workflow is H265 4:2:0 - 1 node - export - then the 7900XTX might be the fastest card there...Same with ProRes and some flavors of RAW...

If you use Fusion, then the AMD cards are weaker there...

Also the whole driver issues thing is nonsense. Both companies have aprox. the same quality in drivers.
Either great or crap (with nvidia on the more serious , cards melting side of things)...

And so on and so forth...

As it has been stated over and over again regarding AMD cards (even from puget members on reddit)...the puget bench overall extended score is not a good indication of performance.

GPU effects is the only thing I look at. Many activities on Resolve will use all your GPU, and that's where the slow down will be. Here's another test

As it has been stated over and over again regarding AMD cards (even from puget members on reddit)...the puget bench overall extended score is not a good indication of performance.

GPU effects is the only thing I look at. Many activities on Resolve will use all your GPU, and that's where the slow down will be. Here's another test

I've gotten a lot better results with my 6800XT than their tests. Especially with NR...There need to be some other people out there that can make a different test other than ...put this H264 file on this timeline, 1 node ->export...Put this same file , 3 NR nodes and export...

Their test, for the most part, seems to be basically testing the asics on the GPUs and not the GPU power...

The candlelight test people used before this puget bench appeared seemed to be more accurate for performance...

My 2 cents.

Everything I've watched and read has said that NVidia is better than AMD for hardware encoding quality, for driver quality and for software support, has this changed?

Not really.

You base this on what?

H265 is higher quality than nvidia by a mile, especially when high bitrate is involved.
H264 with 7900 series might be on PAR.
And AV1 is the same on all three brands.

Driver quality - is the same on both brands...Both have ok and bad drivers...People that keep advancing the narrative that AMD drivers are bad have no idea what they're talking about...

Software support - what does this even mean?

If we're talking about features, AMD have the same equivalent features as nvidia (adrenalin)
If you're referring to support for eg 3D rendering - yes nvidia wins there, but things are moving along with HIP (blender for example and redshift).
If you're talking about Resolve - the cards work fine in Resolve...

As it has been stated over and over again regarding AMD cards (even from puget members on reddit)...the puget bench overall extended score is not a good indication of performance.

GPU effects is the only thing I look at. Many activities on Resolve will use all your GPU, and that's where the slow down will be. Here's another test

I've gotten a lot better results with my 6800XT than their tests. Especially with NR...There need to be some other people out there that can make a different test other than ...put this H264 file on this timeline, 1 node ->export...Put this same file , 3 NR nodes and export...

Their test, for the most part, seems to be basically testing the asics on the GPUs and not the GPU power...

The candlelight test people used before this puget bench appeared seemed to be more accurate for performance...

My 2 cents.

Their transcoding tests are interesting, shows efficiency of decoder and encoder, and intel amd and Nvidia give very different results.

These are the tests it does, and seems like a good test for the GPU effects score.. H.264 is worse on new AMD cards, AV1 is worse. Cuda is the gold standard for 'productivity work' , AMD drivers are terrible for productivity. In my region, 3090 was selling for $US150 less than 7900 XTX, but prices have crept up again, when prices reduce again I can't see anyone being interested in the AMD, not for Resolve.

Everything I've watched and read has said that NVidia is better than AMD for hardware encoding quality, for driver quality and for software support, has this changed?

Not really.

You base this on what?

H265 is higher quality than nvidia by a mile, especially when high bitrate is involved.
H264 with 7900 series might be on PAR.
And AV1 is the same on all three brands.

Driver quality - is the same on both brands...Both have ok and bad drivers...People that keep advancing the narrative that AMD drivers are bad have no idea what they're talking about...

Software support - what does this even mean?

If we're talking about features, AMD have the same equivalent features as nvidia (adrenalin)
If you're referring to support for eg 3D rendering - yes nvidia wins there, but things are moving along with HIP (blender for example and redshift).
If you're talking about Resolve - the cards work fine in Resolve...

Software support meaning how many pieces of software work well with it. For example I also use OBS which works somewhat better with NVidia and I also use it for games where NVidia is far ahead. If you only use your gpu with Resolve then none of this may apply to you, but it's important to me.

Every review I watched on the new AMD video cards has mentioned driver issues and even boot issues in some cases, NVidia still seems somewhat better in this regard. I have two older Ryzen pc's that I use for streaming and both just had the AMD Adrenalin software suddenly fail and claim they aren't compatible, uninstalling/manually installing the latest versions gives errors so I gave up and just removed Adrenalin completely. For comparison those same pc's each have old NVidia (1050 and 1650) cards as well which have never had a driver/app issue.

h264 encoding, again just from the reviews that I've seen and read, show NVEnc to be ahead of AMD for many years now and to still be ahead even with latest AMD cards. I could post endless links showing this but I don't want to clutter my post.

h265 encoding is relatively new to me and I've only tested NVEnc to a Mac Mini M1 were I found NVEnc to be better. I don't know how the new AMD cards compare for HEVC encoding and I can't seem to find any comparisons on that.

EDIT: Linus Tech Tips will apparently be doing an image quality comparison for the video encoders on the new AMD cards so I look forward to that. They did test h265 encode speed on them and they were slower than the new NVidia 40xx cards but not by much. But they also had driver issues and crashes on the new AMD cards.

Their transcoding tests are interesting, shows efficiency of decoder and encoder, and intel amd and Nvidia give very different results.

These are the tests it does, and seems like a good test for the GPU effects score.. H.264 is worse on new AMD cards, AV1 is worse. Cuda is the gold standard for 'productivity work' , AMD drivers are terrible for productivity.

Resolve uses the GPU not only the encoders and decoders. And even in your pasted screenshot the XT is above the 3090. The XTX should be closer to the 4080. In what world is this terrible?

Also AMD drivers are not terrible, stop with this nonsense.

This has been regurgitated by clueless youtubers or clueless redditors...Most of the "driver problems" are unstable ram oc (XMP is overclock). Just because the video driver is the first to go when a memory errors occurs doesn't mean the drivers are bad...Just means the OC is unstable...

Software support meaning how many pieces of software work well with it. For example I also use OBS which works somewhat better with NVidia and I also use it for games where NVidia is far ahead. If you only use your gpu with Resolve then none of this may apply to you, but it's important to me.

Only the 4090 is ahead. The 7900XTX is faster in games than the 4080...And probably in OBS too.

Every review I watched on the new AMD video cards has mentioned driver issues and even boot issues in some cases, NVidia still seems somewhat better in this regard. I have two older Ryzen pc's that I use for streaming and both just had the AMD Adrenalin software suddenly fail and claim they aren't compatible, uninstalling/manually installing the latest versions gives errors so I gave up and just removed Adrenalin completely. For comparison those same pc's each have old NVidia (1050 and 1650) cards as well which have never had a driver/app issue.

This sounds like the famous Windows Update issue where Microsoft replaces AMD GPU drivers forcefully with older versions and the Adrenalin Control Panel stops working. Its fine if you're annoyed by this, but its not an AMD issue...its a Windows issue.

h264 encoding, again just from the reviews that I've seen and read, show NVEnc to be ahead of AMD for many years now and to still be ahead even with latest AMD cards. I could post endless links showing this but I don't want to clutter my post.

The h264 encoder has been updated recently to support B frames. Its now on par with nvenc...

h265 encoding is relatively new to me and I've only tested NVEnc to a Mac Mini M1 were I found NVEnc to be better. I don't know how the new AMD cards compare for HEVC encoding and I can't seem to find any comparisons on that.

H265 was always higher quality than nvenc. And its now faster too...

EDIT: Linus Tech Tips will apparently be doing an image quality comparison for the video encoders on the new AMD cards so I look forward to that. They did test h265 encode speed on them and they were slower than the new NVidia 40xx cards but not by much. But they also had driver issues and crashes on the new AMD cards.

Here are some more tests (Premiere, but until we have others they should do):

https://techgage.com/wp-content/uploads ... nd-XTX.jpg

https://techgage.com/viewimg/?img=https ... nd-XTX.jpg

https://techgage.com/viewimg/?img=https ... nd-XTX.jpg

https://techgage.com/wp-content/uploads ... nd-XTX.jpg

https://techgage.com/viewimg/?img=https ... nd-XTX.jpg

https://techgage.com/viewimg/?img=https ... nd-XTX.jpg

There are more tests on that page where they switch places a few times but overall it is what it is...aka NOT TERRIBLE like some people claim...3D rendering is catching up to nvidia thanks to HIP and soon HIP RT (OptiX equivalent).

Also AMD drivers are not terrible, stop with this nonsense.

You should give them a try on Linux, you'd see. So many troubles reported (I suffered myself).

By the way, some postproductions software on Windows require nVidia cards only.

Do the new AI features in resolve (voice isolation, magic mask etc) work as well with AMD cards?

Yes.

Only the 4090 is ahead. The 7900XTX is faster in games than the 4080...And probably in OBS too.

Only in pure rasterization, once you include ray tracing (which on games is glorious) AMD falls far behind, which is unfortunate because NVidia really needs some competition in the games space. OBS has been tested on the new AMD cards by various people and they are better than the old AMD cards, but still behind NVidia. EposVox on Youtube has a very recent video on this but there's others.

The h264 encoder has been updated recently to support B frames. Its now on par with nvenc...

I read about that but apparently it's still noticeably behind NVenc, again according to the various reviews and tests I've watched.

H265 was always higher quality than nvenc. And its now faster too...

I'd love to find an h265 quality comparison review of this on Resolve (I don't use Premiere Pro) but haven't been able to find one yet, only speed tests which aren't as important to me. For example my Mac Mini M1 while I briefly had it encoded faster than NVEnc but in the end it was moot because I didn't like the quality.

I find one flaw of many reviews is they only time encode speeds but they don't actually look at the final product. Reviews like that to me aren't useful because what is the point in saving time if I don't like the final product as much? It's the exact problem I had with the Mac Mini M1 and why I returned it.

Now that the AMD cards are out hopefully people will do quality comparisons. I've seen h264 comparisons and based on the videos I've watched, h264 quality on the new AMD cards is still not as good as NVenc. I've yet to see a quality comparison on h265, hopefully someone tackles that soon.

Most YT "tests" are going to use stock settings for things like encode. That may not match what you're doing.

Bigger question is are the stock settings for the various encoders equivalent for quality? An easy way to get better test results is to have lower quality stock settings.

Personally I wouldn't put too much faith into any of the gaming YT channels to have the foggiest idea of encoding quality.

The h264 encoder has been updated recently to support B frames. Its now on par with nvenc...

I read about that but apparently it's still noticeably behind NVenc, again according to the various reviews and tests I've watched.

H265 was always higher quality than nvenc. And its now faster too...

I'd love to find an h265 quality comparison review of this on Resolve (I don't use Premiere Pro) but haven't been able to find one yet

Here's a streaming bitrate comparison, that is why it's only 6mbit, h.264 is better but not in the same league as NVENC.

2022-12 creators_! - YouTube.png (609.87 KiB) Viewed 3129 times

The AMD H.265 encoder of the 6000 series is much faster than the 4090, but that's only impressive if the quality is the same or better

I just love these charts where graphically you are led to believe that NVENC is twice the speed of VCE (not RDNA2) when actually there is 15-20 % difference.

Yes, my puny Quadro T600 is more than twice the speed of RX580 VCE, but that's comparing several years old architecture.

Oh, and yes, the moment Quadro runs out of memory, say when you apply speedwarp on 3K footage, millions of little "GPU out of memory" windows pop up and Resolve crashes. Never ever did this happen with the RX580 for proportionally higher demand...it just soldiers on, however slowly...my well be that shared GPU memory technology works really better on AMD.

I just love these charts where graphically you are led to believe that NVENC is twice the speed of VCE (not RDNA2) when actually there is 20 % difference.

.

The graph above is not a measure of speed, but a measure of Quality, using Netflix VMAF

I just love these charts where graphically you are led to believe that NVENC is twice the speed of VCE (not RDNA2) when actually there is 20 % difference.

.

The graph above is not a measure of speed, but a measure of Quality, using Netflix VMAF

Heh got me there, but still, the chart plot should be normalized

The h264 encoder has been updated recently to support B frames. Its now on par with nvenc...

H265 was always higher quality than nvenc. And its now faster too...

Are either of them yet anywhere close to the efficiency of FFMpeg though? I've never used the nvenc in my content because file sizes were always between 50-100% larger than Handbrake.

I find one flaw of many reviews is they only time encode speeds but they don't actually look at the final product. Reviews like that to me aren't useful because what is the point in saving time if I don't like the final product as much? It's the exact problem I had with the Mac Mini M1 and why I returned it.

Exactly. There are 3 factors in fact (relative importance of each is down to your use-case of course):

- Speed
- Image quality
- Resulting filesize

In addition, NVENC in Resolve is limited to certain resolutions. e..g h264 won't go above 5K/2880s, and H265 won't go above 8K. Kinda useless (for me - VR).

The h264 encoder has been updated recently to support B frames. Its now on par with nvenc...

H265 was always higher quality than nvenc. And its now faster too...

Are either of them yet anywhere close to the efficiency of FFMpeg though? I've never used the nvenc in my content because file sizes were always between 50-100% larger than Handbrake.

Yeah that's generally how it works, hardware encoding is not as efficient, but if the bitrate is high enough it can equal FFMPEG quality wise. Therefore, if you're uploading to YouTube or other sites where the file is re-encoded, the size of the encode is not very relevant if your internet connection is fast, but if you're encoding files for streaming/download you want to software encode for maximum efficiency. At 1080P the bitrate where hardware encodes from Intel/Amd/Nvidia(current generation) and software encodes are all reasonably equal quality wise is about 40Mbit/s VBR