Rumor: RTX 3080 Ti with 20 GB vRam, at Feb or later.

For me the possibility of running dual Cards has a high priority. I only have one PC so during export i need to wait, i do not need the 3090, but love the thought that i can go one up. So any tests with resolve in a dual card setup in windows is interesting. my dream setup was the dual 16gb radeons VII because the speed and headroom was more than enough if i could get them running. Dual 10gb 3080 would be perfect but will i work? any one seen anyone try?

In my very simple work the max is always reached on the nvidia 2080 super 8gb (not a problem yet) and on the radeon viis the max i saw was 12gb used on a single card. It is almost only 4k 59,94 gopro media.


Sent from my iPad using Tapatalk

In my very simple work the max is always reached on the nvidia 2080 super 8gb (not a problem yet)
Sent from my iPad using Tapatalk

Why is maxing out the VRAM on the 2080 Super "not a problem yet"?

The 3090 doesn't look that good compared to the 3080.

https://videocardz.com/newz/nvidia-geforce-rtx-3090-gaming-performance-review-leaks-out

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

Yes, but wouldn't the extra vram be interesting (enough) for us resolve users?

When you use CUDA with 4k+ export and are willing to spend ~€1.500..€1.700, yes...

It does not fail and speed is better than the Radeon VIIs in my work flow. So even though varm is Mac’er out it still perform ok.

The 3090 doesn't look that good compared to the 3080.

https://videocardz.com/newz/nvidia-geforce-rtx-3090-gaming-performance-review-leaks-out

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

Yes, but wouldn't the extra vram be interesting (enough) for us resolve users?

Of course the extra vram is important for 4k and above. And the leaks of the '3090' performance are to be taken with a grain of salt.
1. It's not identified that it's a 3090.
2. It's for games, not Resolve.
3. It's hard to believe that even nVidia could think they can charge 2x the price for 10% more performance.

Best to wait and see. The Puget review should drop in days and the early adopters soon after.

The 3090 doesn't look that good compared to the 3080.

https://videocardz.com/newz/nvidia-geforce-rtx-3090-gaming-performance-review-leaks-out

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

Interesting. I'm curious why they are only using a 10900K for the CPU. I wonder if the CPU is choking the benchmarks.

The 3090 doesn't look that good compared to the 3080.

https://videocardz.com/newz/nvidia-geforce-rtx-3090-gaming-performance-review-leaks-out

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

My guess is that the 3080 is priced in anticipation of competing with RDNA2, while the 3090 is priced based on having 24 GB of display memory.

Even so, 10% more performance for 10% more power consumption is much better power/perf scaling than normal in this industry...

It's for games, not Resolve.

Important to note. The memory bandwidth on the 3090 is much faster, and that should be reflected in some operations in Resolve (TNR).

Interesting. I'm curious why they are only using a 10900K for the CPU. I wonder if the CPU is choking the benchmarks.

Intel is still ahead in gaming performance, and in many workloads that are strictly GPU bound, the CPU work is done on a single thread, and benefits from being as fast as possible, despite the fact that it won't saturate a core.

In all the benchmarks I scanned, the 4K reviews were not bottlenecked but the 1080p benchmarks were all described as bottlenecked by the CPU - I am guessing the GPU is able to churn out computations too fast for the CPU to process at these lower resolutions?

[/url]

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

Even so, 10% more performance for 10% more power consumption is much better power/perf scaling than normal in this industry...

Hi.

Here is a link, where they decreased the power consumption for the RTX 3080, to the same level as a RTX 2080 Ti. They decreeased the power consumption from 320 Watt to 270 watt.



As you can see, did the performance of the RTX 3080 only decrease 4,2%.

Here is a Google translate of the link to English:

https://translate.google.com/translate? ... est%2F6%2F

And if somebody prefer, to read the link in German:

https://www.computerbase.de/2020-09/gef ... 80-test/6/


I think that nVidea, this time have increased the power consumption, far more than they did for the previous Pascal or Turing Graphics Cards. I wonder why?

Regards Carsten.

Intel is still ahead in gaming performance, and in many workloads that are strictly GPU bound, the CPU work is done on a single thread, and benefits from being as fast as possible, despite the fact that it won't saturate a core.

Older games tend to be single or dual threaded but for modern games, it is not uncommon for it to be multi-threaded. While they don't use 32 cores they can utilize 4 or more cores.
The 3090 could be just 10% faster but just need more details on the bechmark and other possible specs.

In all the benchmarks I scanned, the 4K reviews were not bottlenecked but the 1080p benchmarks were all described as bottlenecked by the CPU - I am guessing the GPU is able to churn out computations too fast for the CPU to process at these lower resolutions?

Yes at 4K, GPU is the usual bottleneck in terms of performance for gaming (raster-based rendering).

[/url]

It's the same silicon as the 3080 (and also the same drivers), it uses around 10% more power and is around 10% faster.

Even so, 10% more performance for 10% more power consumption is much better power/perf scaling than normal in this industry...

Hi.

Here is a link, where they decreased the power consumption for the RTX 3080, to the same level as a RTX 2080 Ti. They decreeased the power consumption from 320 Watt to 270 watt.



As you can see, did the performance of the RTX 3080 only decrease 4,2%.

Here is a Google translate of the link to English:

https://translate.google.com/translate? ... est%2F6%2F

And if somebody prefer, to read the link in German:

https://www.computerbase.de/2020-09/gef ... 80-test/6/


I think that nVidea, this time have increased the power consumption, far more than they did for the previous Pascal or Turing Graphics Cards. I wonder why?

Regards Carsten.

Each transistor uses fewer watt, but there are many more. So the technology 8n was not efficient enough to offset the amount of transistors at that performance.


Sent from my iPad using Tapatalk

Did I miss the discussion here on the long-awaited testing, or did everyone else miss the Puget article?

Here it is.

In short:

Overall, the new RTX 3080 does extremely well in DaVinci Resolve. Performance while editing and doing basic grades may only be 10-25% faster than the previous generation RTX 20-series cards, but when doing noise reduction or using OpenFX, the performance gap widens from a minimum of 30% faster compared to a Titan RTX to almost 2x faster compared to an RTX 2060 SUPER!

RTX 3080 - 20 GB confirmed: https://wccftech.com/nvidia-geforce-rtx ... confirmed/

I'll buy two.

RTX 3080 - 20 GB confirmed: https://wccftech.com/nvidia-geforce-rtx ... confirmed/

I'll buy two.

Hi.

The wccftech site is known as a place to read the latest rumors.

But I personal also expect the NVIDIA GeForce RTX 3080 20 GB to arrive. The only questen is when?

I don't expect more information, until after the AMD Big Navi event on the 28th October.

Regards Carsten.

Did I miss the discussion here on the long-awaited testing, or did everyone else miss the Puget article?

Here it is.

In short:

Overall, the new RTX 3080 does extremely well in DaVinci Resolve. Performance while editing and doing basic grades may only be 10-25% faster than the previous generation RTX 20-series cards, but when doing noise reduction or using OpenFX, the performance gap widens from a minimum of 30% faster compared to a Titan RTX to almost 2x faster compared to an RTX 2060 SUPER!

That is big! For us, OpenFX and NR performance have always been extremely slow.

Hi.

Quote a nVidea Blog: 'For 4K gaming, the GeForce RTX 3090 is about 10-15% faster on average than the GeForce RTX 3080, and up to 50% faster than the TITAN RTX.'

From: https://www.nvidia.com/en-gb/geforce/ne ... 9x7Q7Y2XFQ

We have to wait and see, how mush faster it is for Resolve.

Regards Carsten.

Hi.

Now it is time for the GeForce RTX 3090 Reviews.


Nvidia GeForce RTX 3090 Founders Edition Review:

https://www.tomshardware.com/news/nvidi ... 090-review

Please click on 'The largest GeForce for a small target group':

https://translate.google.com/translate? ... &sandbox=1

Regards Carsten.

Puget updated their results to include the 3090.
https://www.pugetsystems.com/labs/artic ... ance-1903/


to me it seems like if the rumors about a future 20GB 3080 are true, that might be the sweet spot in regards to performance per dollar and maybe even outright performance.

it's an educated guess, but going by the detailed benchmark results it seems like the 3080 can self boost higher than the 3090. It also looks like the 3090 only pulls ahead in use cases where its extra vram can make a difference, but even then its not by much.

It also looks like the 3090 only pulls ahead in use cases where its extra vram can make a difference, but even then its not by much.

It only makes sense if you're doing something that requires the higher VRAM capacity (for now), or the faster memory bandwidth, or will use the tensor cores. Underwhelming.

i found funny that if blackmagic modify the software in order to take full hardware advantage of that card, it could make all the other card impossible to work with: that will be a flurry of "my computer does not open/crash/out of gpu memory" angry users....

while at the same times there are plethora of users begging for a CPU only resolve....

i found funny that if blackmagic modify the software in order to take full hardware advantage of that card, it could make all the other card impossible to work with: that will be a flurry of "my computer does not open/crash/out of gpu memory" angry users....

That would be piss poor development work. Decent developers know how make their application scale relative to the hardware its running on. While I don't understand or agree with every design decision BM devs have made, I haven't seen anything from them that suggests a lack of skill.

Puget updated their results to include the 3090.
https://www.pugetsystems.com/labs/artic ... ance-1903/

Hi.

Quote: ' If you are considering a multi-GPU setup, however, we have you covered with our DaVinci Resolve Studio - RTX 3080 & 3090 Multi-GPU Performance Scaling article which looks at the performance with a triple RTX 3080 and dual RTX 3090 setup. We will spoil it a little bit, however: there are significant problems with these cards in multi-GPU configurations.'

From above link.

And here is the first version of 'DaVinci Resolve Studio - RTX 3080 & 3090 Multi-GPU Performance Scaling', even that Puget Systems had 'significant problems with these cards in multi-GPU configurations.'

https://www.pugetsystems.com/labs/artic ... ling-1904/

Regards Carsten.

i found funny that if blackmagic modify the software in order to take full hardware advantage of that card, it could make all the other card impossible to work with: that will be a flurry of "my computer does not open/crash/out of gpu memory" angry users....

while at the same times there are plethora of users begging for a CPU only resolve....

For me the pure software edition is just to have at second option if I can`t solve an issue with the gpu. Gpu drivers have some history of compatability issues, a pure C++ implementation on the CPU will be slower by fare like in the 10 times slower.

Of note is the optical flow test, add a 2nd 2080ti or 3080 or 3090 , no performance gain at all - add a 3rd and performance doubles. thats just poor coding for a very heavily used function.

And here is the first version of 'DaVinci Resolve Studio - RTX 3080 & 3090 Multi-GPU Performance Scaling'

Probably want to wait to see how this capacitor design flaw works out, or buy from somewhere with a good return policy.


https://www.igorslab.de/en/what-real-what-can-be-investigative-within-the-crashes-and-instabilities-of-the-force-rtx-3080-andrtx-3090/

Probably want to wait to see how this capacitor design flaw works out, or buy from somewhere with a good return policy.


https://www.igorslab.de/en/what-real-what-can-be-investigative-within-the-crashes-and-instabilities-of-the-force-rtx-3080-andrtx-3090/

Code: Select all

An interesting pattern of behavior emerged that did not affect all cards or manufacturers and the problems only occurred at certain boost clock rates above or just around 2 GHz

are they trying to OC the card and complaining that it does not work?

are they trying to OC the card and complaining that it does not work?

It sure sounds like someone trying to push a card well beyond the redline and not knowing what they are doing.


Edit: I did some digging, and it looks like some of the manufacturers are doing the bare minimum when it comes to clean power delivery, others have gone off the deep end and have very clean power delivery.

Hi.

Here is a link with more explanation of the Igorslab idea:

https://www.techpowerup.com/272591/rtx- ... tor-choice

And here is a good 21:35 min long YouTube, where he takes 4 RTX 3080 apart and explain the idea behind the Igorslab posting. Please start it twice:

Regards Carsten.

I did some digging, and it looks like some of the manufacturers are doing the bare minimum when it comes to clean power delivery, others have gone off the deep end and have very clean power delivery.

I'd be grateful if you can please share your findings

I'd be grateful if you can please share your findings

Take a look at the Jayztwocents video above, it will give a good high level overview.

To go into detail you need a not trivial amount of overclocking and electronics knowledge. Mainly because the last several generations of gpu self overclock.


as an aside, I'm shocked igo didn't go at the cards with an oscilloscope, as he could probably see the voltage ripple pretty easily.

Evga has come out with what for them would be an official statement.

https://forums.evga.com/Message-about-E ... 95238.aspx

Hi all,

Recently there has been some discussion about the EVGA GeForce RTX 3080 series.

During our mass production QC testing we discovered a full 6 POSCAPs solution cannot pass the real world applications testing. It took almost a week of R&D effort to find the cause and reduce the POSCAPs to 4 and add 20 MLCC caps prior to shipping production boards, this is why the EVGA GeForce RTX 3080 FTW3 series was delayed at launch. There were no 6 POSCAP production EVGA GeForce RTX 3080 FTW3 boards shipped.

But, due to the time crunch, some of the reviewers were sent a pre-production version with 6 POSCAP’s, we are working with those reviewers directly to replace their boards with production versions.
EVGA GeForce RTX 3080 XC3 series with 5 POSCAPs + 10 MLCC solution is matched with the XC3 spec without issues.

From a leak the nvidia rtx a6000 has 2x the tensor cores. So for AI it looks very interesting.
https://wccftech.com/nvidia-next-gen-qu ... connector/

Also AMD Seem to be shipping RDNA workstation cards close to rtx a6000 late this year at a 250 w TDP. I will skip this card and wait for real 2x performance pr. watt.


Sent from my iPhone using Tapatalk

From a leak the nvidia rtx a6000 has 2x the tensor cores. So for AI it looks very interesting.
https://wccftech.com/nvidia-next-gen-qu ... connector/

Hi.

Quote above link: 'NVIDIA Quadro RTX (Ampere) A6000 Rumored Specifications
The specifications list down 10752 CUDA Cores which are arranged in 84 SMs. This is the full-fat Ampere GA102 GPU we are looking at which should be faster than the GeForce RTX 3090 since it packs 2% more cores however it would mostly depend on the clock speeds the card runs at. The Quadro RTX A6000 operates at a boost clock of 1860 MHz which should equal around 40 TFLOPs of Compute horsepower.'

The NVIDIA GA102 die have 336 Tensor Cores, and the NVIDIA GA100 die for datacenters 512 Tensor Cores.

From: https://www.techpowerup.com/gpu-specs/nvidia-ga100.g931

So I will be surprised, if the A6000 Ampere will get 656 Tensor Cores?


About the 40 TFLOPs of Compute horsepower on Ampere. nVidea changed the ALUs on Ampere compared to Turing.

There are still 64 pure FP32-ALUs per SM, but instead of the 64 pure INT32-ALUs, there are now 64 new ALUs, which can perform floating point and integer calculations.
- but not in parallel.

A Turing SM can therefore perform a maximum of 64 FP32 and 64 INT32 calculations at the same time. An SM of Ampere, on the other hand, either 128 FP32 or 64 FP32 and 64 INT32 calculations
- depending on the workload, the application distributes to the graphics card.

Regards Carsten.

are they trying to OC the card and complaining that it does not work?

No, a bit of a rushed launch by Nvidia, I guess to be in the market before new ps/xbox consoles and amd. 3rd parties given a couple choices of capacitor options, looking like those who went with the cheapo option, and didn't have time to test thoroughly before launch, because of non fully functional driver, are most likely to recall cards.

The 2-slot blower and aio water cards, most useful for workstations, not out until next month anyway when this should be resolved.

Pretty awesome that the scalper bots now have to deal with all the defective cards.

and you can get some free mooncakes for your trouble!

Is it possible/doable to water cooling that card?

I look with interest at these new cards as I was thinking to get a myself a new card for Christmas present !! Currently have a 1080Ti MSI 11G but am confused as to how Resolve compares with CUDA for NVIDIA or OpenCL for AMD. There are rumours of 3070 with 16G etc and no idea how Big Navi will turn out. It is the performance difference on Resolve with decode and encode that is of interest to me.

Is it possible/doable to water cooling that card?

Hi.

In the last 16 hours haveI read about the RTX 3080 and RTX 3090 Problem on 5 or 6 different sites. The best summary and current status I have read is from Wccftech.
Wccftech is a site known for its early rumors, but as I say I want to share the information from their site. Here is a link:

https://wccftech.com/nvidia-official-re ... groupings/

And

'EK Is Racing To Deliver GeForce RTX 30 Ampere Water Blocks In Time For Launch'

https://hothardware.com/news/ek-geforce ... cks-launch

https://www.techpowerup.com/271431/alph ... ter-blocks

Regards Carsten.

but am confused as to how Resolve compares with CUDA for NVIDIA or OpenCL for AMD.

Hi.

nVidea changed the ALUs ( Calculator circuit ) on the RTX 3080 Serie compared to the RTX 2080 Ti Serie and cards before.

There are still 64 pure Floating Point FP32 ALUs per SM, but instead of the 64 pure Integer INT32 ALUs, there are now 64 NEW ALUs, which can perform Floating Point and Integer calculations. But not at the same time.

A RTX 2080 Ti SM can therefore perform a maximum of 64 FP32 and 64 INT32 calculations at the same time. An RTX 3080 SM, on the other hand, either 128 FP32 or 64 FP32 and 64 INT32 calculations
depending on the workload, Resolve is sending to the graphics card.

If a PC Application only is using its RTX 3080 Graphics Card for Floating Point FP32 calculations using CUDA. Is the performance 29768 GFLOPS.

But if a PC Application only is using its RTX 3080 Graphics Card for 50% Floating Point FP32 and 50% of Interger calculation using CUDA. Is the performance 14.884 GFLOPS.

I don't know how the Internal of Resolve is working, but I expect the right answer to be somewhere in between. The best way to find the actual performance, is to check the Resolve Benchmarks for the different Graphics Cards.


In the paid STUDIO Version of Resolve, can certain combination of Codec, Resolutions, Bit width and Chroma subsampling be hardware decoded/encoded on either a AMD or a nVidea Graphics card.
Then you can use a little less powerfull CPU.

Regards Carsten.

Still a little confused. I have the paid studio version. To clarify, I may be incorrectly assuming that both AMD and NVIDIA have hardware accelerated decode and encode for various codecs that Resolve can access. For Resolve to perform other image processing it would use either CUDA or OpenCl to perform these calculations so would be dependent on what is being asked of the GPU. Am I correct in these assumptions? In my case I use very little image processing other than fine tune converting from V log to rec 709 or HLG. Source is from GH5 UHD 60P. Decode on the timeline is important and also then encode of finished project. At the moment 2 tracks of UHD GH5 and GH5S both V log play fine in multicam so I am happy at the moment anyway. I am just interested to understand how the GPU's and their performance will effect Resolve performance.

As a second question assuming the hardware encoders are used for export is NVIDIA or AMD the better performer.

Hi.

In Resolve the CPU is used to run the app, disk I/O, fusion, compression and decompression of codecs.
Resolve does all its image processing in the GPU on the graphics card. More CUDA/OpenCL Cores are better.

In the paid STUDIO Version of Resolve, can certain combination of Codec, Resolutions, Bit width and Chroma subsampling be hardware decoded/encoded on either a AMD/nVidea Graphics card.

nVidea have the best documentation, so let us begin with it. Here is a link:

https://developer.nvidia.com/nvidia-vid ... ECFeatures

If you click on 'Supported Format Details (Click to learn more)', then you see which Codec, Resolutions, Bit width and Chroma subsampling, that can be hardware decoded/encoded.

And if you again click on the 'Consult the GPU Support Matrix for a full list of GPU's, encode formats and ..... ' Icon, can you see a list for the different nVidea Graphic Cards.

And finally the newest. The AMD's Graphics Card had only support for hardware accelerated decode and encode since resolve version 16.0:

https://en.wikipedia.org/wiki/Unified_Video_Decoder

Regards Carsten.

Is it possible/doable to water cooling that card?

Sure but waterblocks are not really suitable for a work environment.

EVGA among others will have hyrbid cards - factory liquid and 240/360mm radiator, maybe acceptable risk, not until probably november.

https://www.evga.com/articles/01434/evg ... 30-series/

Gigabyte has a 2-slot 3090 blower card, should be out soon. Thats the most practical if your moving you system around at all.

https://www.gigabyte.com/Graphics-Card/ ... BO-24GD#kf

but am confused as to how Resolve compares with CUDA for NVIDIA or OpenCL for AMD.

Hi.

nVidea changed the ALUs ( Calculator circuit ) on the RTX 3080 Serie compared to the RTX 2080 Ti Serie and cards before.

There are still 64 pure Floating Point FP32 ALUs per SM, but instead of the 64 pure Integer INT32 ALUs, there are now 64 NEW ALUs, which can perform Floating Point and Integer calculations. But not at the same time.

A RTX 2080 Ti SM can therefore perform a maximum of 64 FP32 and 64 INT32 calculations at the same time. An RTX 3080 SM, on the other hand, either 128 FP32 or 64 FP32 and 64 INT32 calculations
depending on the workload, Resolve is sending to the graphics card.

If a PC Application only is using its RTX 3080 Graphics Card for Floating Point FP32 calculations using CUDA. Is the performance 29768 GFLOPS.

But if a PC Application only is using its RTX 3080 Graphics Card for 50% Floating Point FP32 and 50% of Interger calculation using CUDA. Is the performance 14.884 GFLOPS.

I don't know how the Internal of Resolve is working, but I expect the right answer to be somewhere in between. The best way to find the actual performance, is to check the Resolve Benchmarks for the different Graphics Cards.


In the paid STUDIO Version of Resolve, can certain combination of Codec, Resolutions, Bit width and Chroma subsampling be hardware decoded/encoded on either a AMD or a nVidea Graphics card.
Then you can use a little less powerfull CPU.

Regards Carsten.

So much for the theory.
Test show a different outcome https://www.pugetsystems.com/labs/articles/DaVinci-Resolve-Studio---NVIDIA-GeForce-RTX-3080-3090-Performance-1903/ with Resolve which only uses fp32.

Is it possible/doable to water cooling that card?

Sure but waterblocks are not really suitable for a work environment.

I wouldn't fully agree with that. watercooling can be fine in a production environment assuming you follow proper procedures. it's kind of like driving a very high performance car, you need to monitor it more closely, and you need to perform maintenance more often.

if you are the type of person that just wants to throw a bunch of components in a box and then not touch it for a couple years, then no you don't want to water cool a machine in a production environment.

So compression and decompression of h.264/265 codecs is not the same as h/w decode support in GPUs? a GPU with h/w decode/encode would still use CPU for compress/decompress?

So compression and decompression of h.264/265 codecs is not the same as h/w decode support in GPUs? a GPU with h/w decode/encode would still use CPU for compress/decompress?

Hi.

I am sorry. I am using different words for the same thing.

The word 'codec decompression and compression' Is Peter Chamberlain way to describe it.

Quote: 'In general, the CPU hosts the application, UI etc, manages disk I/O and most codec decompression and compression. There are exceptions to this for some OS and some GPUs and some codecs .....'

From this link, his last post: viewtopic.php?f=21&t=119899&p=658985&hilit=compression+decompression+of+codecs#p658985


I expect that some time ago, I was using 'hardware accelerated decode and encode for various codecs', but lately changed over to use 'hardware decoded/encoded'.

But all 3 wordings describe the same thing.

Regards Carsten.

At least with my Studio version I can turn on or off hardware decode by the GPU in settings. Not sure if that is the same for the free version. With my Threadripper and 1080Ti with GPU hardware decode ON it will play UHD 60P from my GH5 with CPU at 10% GPU at 35%. With hardware decode turned off the % are reversed. The Threadripper by itself will play 2 tracks of the GH5 in multicam too but now CPU is at 55%. GPU is still doing something but not sure if that is just the display.

I think that nVidea, this time have increased the power consumption, far more than they did for the previous Pascal or Turing Graphics Cards. I wonder why?

Regards Carsten.

It seems pretty obvious to me.

The new process that nVidia is using offers a huge improvement in power efficiency as well as transistor density. That left nVidia with two options. One was to reduce power consumption at the same performance level, the other was to fill that up with more resources.

nVidia chose more resources, so while most of the line has basically maintained parity as far as power consumption goes when compared to its predecessor, they're all also packed with twice as many cores.

So it's a theoretical peak of double the compute throughput with around a 15% increase in power consumption.

As for why, I think that's a tad less obvious. It's actually the Xbox Series X + PlayStation 5 launches. They're being targeted at serious gamers rather than the usual budget gamers, both designed to deliver 4K gaming. That's a major problem for nVidia because it's a market nVidia can't compete in at all, and it could consume a huge part of the gaming market, which would be brutal for nVidia, but a sweeping win for AMD. RDNA2 could be an abysmal failure and AMD would still win.

That's not that I expect RDNA2 to be a lame duck, it seems like it's going to compete pretty well, but the consoles are still a big ace in the hole for AMD.

That's probably also why nVidia rushed the Ampere launch a bit, leading to the supply problems...