Question Speculation: RDNA3 + CDNA2 Architectures Thread

[H T C]

Right now, nVidia has A TON of 3000 series inventory to get rid of due to crypto crash, which is why they priced 4090 as they did. Don't kid yourselves: nVidia's plan was to price the upcoming 4090 Ti @ LEAST in the $2000 range.

By pricing the 7900 cards aggressively like this, they make it A LOT MORE difficult for nVidia to get rid of those cards, which is why stuff like this is already happening.

AMD is in a position to make it VERY HARD for nVidia to get rid of that inventory: if they make their 6000 series cards EVEN CHEAPER than they are now and NOT BY A LITTLE, they MAY be able to clear out THEIR OWN inventory of 6000 series cards which they too have (though nowhere near to the extent of nVidia's).

Should AMD opt to go this route, it won't be cheap for them and would MOST CERTAINLY greatly reduce their profits BUT, it would also substantially increase their market share in the process WHILE forcing nVidia to resort to SERIOUS price cuts themselves in order to get rid of their current inventories, but this would cost A LOT MORE to nVidia than it would to AMD (my guess would be around 10 times more).

An approach like this would be MUCH MORE EFFECTIVE if the 7900 series were to match (or more) the 4090 @ least in raster but that appears not to be the case and, as expected, they still trail in RT.

 

[RnR_au]

I'm still a devout rasterfarian.

I can get behind this new fangled religion

The evil unbelievers in rays and tracing needs a cool name as well... otherwise how will we know them? How will we be able to cast stones at them?

The GCD is too small to compete at the top.
If you take out interconnect area the GCD is around ~265mm2 only. Not sure what is the thought process there.

They have the capability to go big, but they did not. Why do chiplet at all with such a small die, if monolithic it could even be around 450mm2 when all that interconnect logic is cut out

What could be done with 2 GCD's (say 2 x N32?) at 450W TBP?

 

[gdansk]

The evil unbelievers in rays and tracing needs a cool name as well... otherwise how will we know them? How will we be able to cast stones at them?

Oh we shall not demonize them.

(Because some day RT will be pretty much always 4K 144Hz+ without any upscaling and we will convert but that's not this generation)

 

[amenx]

I think I would have no problem going for better raster (at 4k) than better RT this time around. Next gen RT would be more capable for 4k which would make it more appealing then.

 

[jpiniero]

What the heck happened to 7900xt price? Cut down 15%, 12.5% lower clock, but priced only 10% less?

I am not paying $1k for a GPU, and the cut down part seems to have terrible perf/$ ratio. I guess I wait for N32 or get some N2x part.

Unless 7900xt is really only 10% slower? That would be terrible scaling on the top die...

Maybe that was the one 'concession' on pricing AMD was able to get. Make the XTX look better in comparison so people spend the extra 100 bucks.

 

[RnR_au]

For RT AMD is still pushing for an alternate approach to Nvidia's brute force one. The latter has little chance to really become mainstream anytime soon, with mainstream meaning including consoles, iGPUs and handhelds. So realistically there has to be more room for scaling with RT rendering beyond silly tech like upscaling and frame interpolation. This is still a pretty major unsolved problem if the gaming industry is supposed to be able to use RT across all possible gaming systems eventually.

I think Lumen from Epic's Unreal Engine 5 could show the way.

Could.

Its a smart fast software implementation that is twice as fast as their implementation using RT cores. More RT hardware would obvious make things faster, but what hardware primitives on the cpu could accelerate Lumen?

I tried googling to see if there was any indications that additional L3 cache ie 5800X3D might run Lumen faster, but my google-fu is either faulty or there just isn't any data atm. Could AVX-512 help? Its not in the current consoles, but its likely to be in the next-gen versions.

And if cpu's get FPGA's?

 

[exquisitechar]

At least a silver lining for RDNA 3 being not very good is that Navi33 will probably be pretty cheap.

What could be done with 2 GCD's (say 2 x N32?) at 450W TBP?

Rather than 2 GCDs, I think the possibility for this generation was more so a larger GCD. I don't think such a part would have been successful, AMD was right not to make anything higher end than Navi31.

 

[moinmoin]

I think Lumen from Epic's Unreal Engine 5 could show the way.

It does. Starting with a fast software based implementation that can make use of optional hardware acceleration realistically is the only way to offer (back and forward) compatibility with a lot of hardware. If I were AMD this is the kind of RT solution I'd push with technical support and through hardware acceleration.

I hope other engines like Unity will take that approach as well. E.g. currently RT in Unity appears to require hardware acceleration ensuring it will never fly as a mainstream feature.

 

[Timorous]

I saw comments expecting RXT4080 16GB to be discounted, because It's "unsellable" for the current price.
I have a question for you all.
If Nvidia discounts It significantly to $999, then which one would you choose?

I would take 20% more raster performance at the same price.

The Techspot 4090 review showed that the 4090 still takes a 54% hit to FPS when turning on RT at 4K and it is still sub 60 fps in many titles at native 4K. Even though it is the best RT card you can buy it still does not do for RT what the 9700XT did for AA and AF.

 

[DAPUNISHER]

Ray tracing is a useless eye candy? Because It eats up performance?
Will you set everything to low quality, because higher settings also eat up performance?
It looks like even that useless 4080 16GB will have higher RT performance than 7900XTX.

That is only one ray traced title, and not indicative of how other engines will perform. Cyberpunk'd is poorly optimized for AMD hardware, CPU and GPU. With Spiderman the bleeding isn't nearly as bad.

The game engines AMD announced yesterday, are going to improve AMD's approach to RT even more. Consequently, I don't think 2077 can be used as the poster child for future RT performance in every title. I am not suggesting AMD will outright win any benchmarks with RT on, but that beatdown in 2077 is by no means a good magic 8 ball for what to expect.

 

[Kocicak]

Even with 2x N32 you get weak RT performance. ...

So what? Some people can live without RT.

These new GPU dies are pretty small and putting more of them on one card seems like an obvious option, provided they can communicate with each other.

 

[exquisitechar]

That is only one ray traced title, and not indicative of how other engines will perform. Cyberpunk'd is poorly optimized for AMD hardware, CPU and GPU. With Spiderman the bleeding isn't nearly as bad.

The game engines AMD announced yesterday, are going to improve AMD's approach to RT even more. Consequently, I don't think 2077 can be used as the poster child for future RT performance in every title. I am not suggesting AMD will outright win any benchmarks with RT on, but that beatdown in 2077 is by no means a good magic 8 ball for what to expect.

The bleeding isn't nearly as bad because Cyberpunk makes relatively heavy use of RT, while Spider-Man is a PS5 game that doesn't. RDNA 3 probably won't even match Ampere in any game that isn't limited to console-like RT. Ada's performance advantage will be simply massive.

We still need to see far more benchmarks, though. AMD's presentation was really poor in that regard, and they further obfuscated things by using FSR as well. They really do not seem confident about these cards.

 

[zinfamous]

Some people are saying Ray tracing is uselss dam. Just wait for next gen graphic engines, which will have built in Ray Tracing.

Then I guess I'll wait for next gen GPUs to be released.

 

[zinfamous]

Don't be angry mate, wait for RDNA4.
RDNA 4 I promise will be awesome.😊

Jokes aside,
The GCD is too small to compete at the top.
If you take out interconnect area the GCD is around ~265mm2 only. Not sure what is the thought process there.

They have the capability to go big, but they did not. Why do chiplet at all with such a small die, if monolithic it could even be around 450mm2 when all that interconnect logic is cut out

my guess is that they plan to glue two sets of those chiplet GCDs together at some point and create a serious frankengpu in the next 2 years.

 

[DAPUNISHER]

The bleeding isn't nearly as bad because Cyberpunk makes relatively heavy use of RT, while Spider-Man is a PS5 game that doesn't.

Cyberpunk was on console too, so I don't see what binging PS5 into it does for your point. It is a tech showcase game that implemented more RT features for PC. Perhaps it is important to recall what an absolute debacle it was at launch. The talking point keeps using 2077, but based on how long it took to get that game sorted out, I think it is somewhat hilarious. Hey guyz! Look at RTX in this badly broken game! Not the flex people think it is. And again, it is not indicative of future performance in all titles, as you just remarked yourself. And people can't simply wish away the game engines that AMD is partnering with.

BTW, could you point to a few posts where you type anything good about AMD? I haven't seen one. It seems like at best you might give them a backhand compliment. Considering this an AMD specific thread, I am not saying you have an agenda, but you are serious negative Ned, why? You can answer that in PM if you like, don't want to derail the thread. Or don't answer it, that's an answer too.

 

[Carfax83]

I would take 20% more raster performance at the same price.

The Techspot 4090 review showed that the 4090 still takes a 54% hit to FPS when turning on RT at 4K and it is still sub 60 fps in many titles at native 4K. Even though it is the best RT card you can buy it still does not do for RT what the 9700XT did for AA and AF.

Comparing RT to AA and AF is ridiculous. RT is the most intensive (and realistic) computational graphics methods available to gamers and requires multiple teraflops of processing power to pull off.

And even then, we have to use A.I and machine learning algorithms to accelerate it so that the games are actually playable.

RT is probably the pinnacle of contemporary game technology. AA and AF are so 90s.

 

[exquisitechar]

Cyberpunk was on console too, so I don't see what binging PS5 into it does for your point. It is a tech showcase game that implemented more RT features for PC. Perhaps it is important to recall what an absolute debacle it was at launch. The talking point keeps using 2077, but based on how long it took to get that game sorted out, I think it is somewhat hilarious. Hey guyz! Look at RTX in this badly broken game! Not the flex people think it is. And again, it is not indicative of future performance in all titles, as you just remarked yourself. And people can't simply wish away the game engines that AMD is partnering with.

Cyberpunk was on consoles too, yes, but it wasn't an exclusive and it made full use of the capabilities of high end PCs. Spider-Man is a PC version of a PS5 version of a PS4 game that makes little use of RT. We will see what performance will be like in future titles, but based on AMD's data, it doesn't look like RDNA 3 will be remotely competitive.

BTW, could you point to a few posts where you type anything good about AMD? I haven't seen one. It seems like at best you might give them a backhand compliment. Considering this an AMD specific thread, I am not saying you have an agenda, but you are serious negative Ned, why? You can answer that in PM if you like, don't want to derail the thread. Or don't answer it, that's an answer too.

Type anything good? I was very optimistic about RDNA 3 right up until yesterday, in this very thread, and I got called out on it after the hype train derailed, lol. I also generally like AMD's hardware. They haven't been doing so well recently in the desktop market, though, so my recent posts are naturally a bit negative.

 

[Dribble]

Initial thoughts:
$1000 and no DDR6X?

Being as they've gone to all that effort with the separate cache chiplets I would have expected more cache? I mean you gotta assume on die would be faster and simpler as it's on the same chip, so if you go off die with all the extra logic and complexity then you've gotta be aiming for a higher quantity but this has less then the 6900XT?

XT only $100 cheaper, who's gonna buy that if the XTX is available?

They have AI cores meaning FSR 2 (using shaders) is likely to be quietly forgotten and left without support once FSR 3 (using AI cores) appears.

 

[Timorous]

Comparing RT to AA and AF is ridiculous. RT is the most intensive (and realistic) computational graphics methods available to gamers and requires multiple teraflops of processing power to pull off.

And even then, we have to use A.I and machine learning algorithms to accelerate it so that the games are actually playable.

RT is probably the pinnacle of contemporary game technology. AA and AF are so 90s.

You miss my point. Before the 9700XT turning on AA or AF tanked performance to a point where it was unusable. After the 9700Pro you could use various AF and AA levels and remain fully playable. It was a large change and turned AF / AA to being default on technologies.

RT has taken longer to go from unaplayable to its current state because it is a lot more complex but it is still not a default on option because in many games turning it on makes the game unplayable, even on a 4090 at native.

However. I do expect the 5090 to actually manage to make RT default on for 4K60 in virtually everything, I think the 5090 will be the card that does for RT what the 9700Pro did for AA/AF and at that point if AMD are still behind in RT performance they are going to need to offer deep discounts to make the trade off worth it.

So right now, today I do not think RT is common enough or performant enough to be a primary purchasing consideration. By next gen I expect the shift will have happened and then RT performance does become a primary purchasing consideration.

 

[exquisitechar]

I mean you gotta assume on die would be faster and simpler as it's on the same chip, so if you go off die with all the extra logic and complexity then you've gotta be aiming for a higher quantity but this has less then the 6900XT?

It's cheaper because N5 is ridiculously expensive.

They have AI cores meaning FSR 2 (using shaders) is likely to be quietly forgotten and left without support once FSR 3 (using AI cores) appears.

I don't think they're "AI cores", judging by the performance. Just new WMMA instructions that run on the ALUs. They may be usable for upscaling, but I doubt that FSR2 is being abandoned.

 

[Dribble]

It's cheaper because N5 is ridiculously expensive.

Yes but I am guessing a 16mb cache chip is super tiny - most of the cost will be the i/o between the cache chips and the main gpu chip (requiring logic on the cache chip, the interposer and the logic on the main gpu chip). So surely if you made that chip have 32mb or 64mb which is just more cache, all the i/o logic is the same then it would still be very small and only cost slightly more to make but it multiplies your cache.

 

[leoneazzurro]

Yes but I am guessing a 16mb cache chip is super tiny - most of the cost will be the i/o between the cache chips and the main gpu chip (requiring logic on the cache chip, the interposer and the logic on the main gpu chip). So surely if you made that chip have 32mb or 64mb which is just more cache, all the i/o logic is the same then it would still be very small and only cost slightly more to make but it multiplies your cache.

It depends on the level of scalability and return in terms of performance/cost of one solution against the other. With this solution AMD can scale their memory bus by multiples of 64 bit from 384bits to 192bits by using the same MCD module. With more cache the MCD would have been bigger, not 2x but certainly not by a negligible amount, and AMD says that improvement in usage patterns offset the reduction in size. Even if N6 is way cheaper than N5, it is not exactly cheap by any means. And what the performance increase would have been, we don't know, but AMD profiling team does and if they decided not to use that solution, they have probably their reasons.

 

[DAPUNISHER]

Cyberpunk was on consoles too, yes, but it wasn't an exclusive and it made full use of the capabilities of high end PCs. Spider-Man is a PC version of a PS5 version of a PS4 game that makes little use of RT. We will see what performance will be like in future titles, but based on AMD's data, it doesn't look like RDNA 3 will be remotely competitive.

Type anything good? I was very optimistic about RDNA 3 right up until yesterday, in this very thread, and I got called out on it after the hype train derailed, lol. I also generally like AMD's hardware. They haven't been doing so well recently in the desktop market, though, so my recent posts are naturally a bit negative.

I thank you for the reply.

From my POV, a bit negative is an understatement. Why are you even contending my post about Spiderman then? Why not be positive on the idea that all PS5 ports will do well on AMD cards, instead of hand waving it away as moot or worse? Why move the goal post so that only certain RT titles and effects matter? If you are suggesting that titles that don't use every possible effect should be discounted, why? Is Spiderman not fun? Have reviewers and players not remarked about how cool and value added the reflections are? I rarely see anyone mention the better shadows, or proper/correct lighting in games they play. Yeah, a couple of IQ aficionados but that's about it. That's your cue you silly outlier people that seem to disproportionately post in tech forums. If games use those RT effects, and many of us don't turn them on, what did we lose? I only use reflections, the rest cost me way too much performance on my 3060ti for what I consider a terrible trade off.

I hope you can see where I am going with this. Even your remark suggesting AMD themselves doesn't expect to be remotely competitive is negative hyperbole. They just announced partnerships with game engines. I drew a very different conclusion than you did from that info. What I took away, is that their future games will optimize for the ray tracing AMD hardware uses. That can only help performance in those titles.

And let me close my remarks with saying again, I think it is absurd to use 2077 as the poster child. The game is so badly optimized, they are still releasing major updates all this time later. For a single player game! SMDH. Not what I'd call a great standard bearer. And everyone knows it runs poorly on AMD hardware, period. It makes it look like an agenda driven talking point to constantly hold it up. Because it reads like sloganeering when it is used and other examples are discounted. No one is denying that AMD is behind on one of the most important graphics techs. However, like all graphics tech in the past, like TressFX, hairworks, PhysX, etc. it is entirely possible for a particular game to implement and favor one vendor over the other. In this case it is likely to close the gap in those titles, as opposed to leading. Not remotely competitive is negativity for its own sake. There is no way we can know what the performance in those titles will be like yet.

That's just like, my opinion man.

 

[eek2121]

for older people with bad eye sight with corrective lenses it doesn't make a huge difference. My eyes aren't the same as a 20 year olds or even a 30 year olds or even a 40 year old's. I can make out major details in rt scenes, anything other than that my eyes won't pick up on. I want good graphics with good frames. I'll leave stellar graphics and less frames to the young.

FYI Eye sight should have nothing to do with RT. RT is more about lighting and reflectivity than details. Using RT means you get realistic lighting, accurate reflections off from mirrors, water, etc.

All of that hyper realistic CGI you see in movies uses RT.

You forget that RTX 4080 Ti and RTX 4090 Ti will launch next year . RTX 4090 is not even the full die that is the issue and RDNA RX 7900XTX is full die so it means that Nvidia has not even launch the real flagship and AMD has clearly shown the white flag for this generation.

Do. you really not think AMD has an answer to the refresh?

At least a silver lining for RDNA 3 being not very good is that Navi33 will probably be pretty cheap.

Rather than 2 GCDs, I think the possibility for this generation was more so a larger GCD. I don't think such a part would have been successful, AMD was right not to make anything higher end than Navi31.

I will be surprised if we don’t eventually get a dual CCD part from AMD. There were so many rumors of it early on, and then one day it got cut. Either multiple people were fed misinformation or AMD needed additional time to get it to work. Note that if AMD added a second GCD, raised TGP to 450W, and dropped clocks a bit, that 1.5-1.7x would likely be 2.5-3.0x faster (assuming no scaling overhead), and would be significantly faster than the 4090 in raster performance.

AMD would likely also have to charge almost $2,000 for it.

I feel like what we got is an early preview of what is to come and AMD will refresh RDNA3 with their original vision for it later on, either when NVIDIA drops the Ti models or at some other time.

EDIT: Oh and the clocks/TGP/TDP are suspiciously low as well. I find it very hard to believe they can’t scale to 450W like the competition.

Oh and I am leaving this relevant quote from Videocardz here for those upset about the “Stream Processors” number:

And since we are at it, AMD should clarify how we are supposed to list Steam Processor count for Radeon 7900 series. Both AMD and board partners are now showing 6144 cores for the top part, while almost every media outlet claims it is 12288, for a good reason.

AMD RDNA3 architecture has ‘Dual Issue’ design, which can now execute not one but two FP32 arithmetic commands at the same time. What this means is that each CU can now do 128 FP32 calculations instead of 64 (RDNA2). To reach the advertised 61 TFLOPs, one would have to multiply 6144 SP × 4 × 2.5 GHz ≅ 61 TFLOPs, or use the same method we use for every modern GPU: 12288 SP × 2 × 2.5 GHz. Obviously, the second option should be more readable to users.

 

[biostud]

Still a pricerange above my budget. Looking forward to seeing the ~$600 battle in spring.