Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[moinmoin]

Nvidia investing more into the gaming community is less about helping the community and more about driving sales for future products. Nvidia's biggest competitor isn't AMD; it's themselves. With such a dominant position in the market, to sustain long-term revenue they need to convince the majority of their user base to upgrade as often as possible. If RT didn't exist, we'd all eventually settle for a GPU that does 4K comfortably and then Nvidia annual sales would fall because we'd hit the point of diminishing returns, not unlike how smartphones have gotten so competent and expensive that it's silly to upgrade on an annual basis. RT and DLSS were created to solve this exact long-term problem that Nvidia faces. Nvidia will continue to fund and sponsor game developers to use their proprietary technologies because it guarantees a market for their GPUs.

Exactly.

And I'm pretty happy by the approach AMD takes in this environment. AMD doesn't really compete heads on but instead carves out different niches that Nvidia with its demand for control doesn't serve. Be it open source software like drivers and implementations of popular algorithms, be it hardware for consoles. The combination of both allowed for Steam Deck, a PC running Windows games under Linux without much issues.

AMD clearly is a follower, but the current setup is not bad: Nvidia popularizes new tech, AMD follows with open implementations of them. The latter will be interesting to watch in server space as well with AMD, Intel and others all pushing for a smooth migration away from CUDA.

The biggest challenge for AMD in the consumer graphics space is finding an implementation of RT that actually scales and is also usable on handhelds and other form factors like consoles clearly not suitable to triple digit watts monster size cards. Nvidia's focus for RT on the other hand is purely brute force performance with RT a selling proportion for premium cards.

 

[eek2121]

Cloning is a bit overstated. AMD is forced to have a different approach. They have to be 'open'. There is almost 0 reason to add features because they will not be adopted by developers. There is maybe 15% of the market with AMD GPUs and only 10% of those who have the cool new GPU that can use the feature. What developer will implement that feature without a payoff? It only pays if the feature works on competitor's products too. A second player at RTG's inconsequential size does not get to add new features that require developer support. That is always a losing battle.

(Except on consoles - where Radeon did do some weird things like the Xbox 360's tessellator before Nvidia)

I disagree that they have to be open.

NVIDIA RTX Remix is the perfect example of what I'm talking about. It is not likely that RTX Remix will be open, but because any user will be able to create mods using it, it doesn't need to be.

 

[Kepler_L2]

https://twitter.com/x/status/1601036292380250112

Profiling results from RDNA2 running Portal RTX. TL;DR Only about 3% of the HW is being utilized properly for some reason (insert GameWorks 2.0 conspiracy here).

 

[GodisanAtheist]

https://twitter.com/x/status/1601036292380250112

Profiling results from RDNA2 running Portal RTX. TL;DR Only about 3% of the HW is being utilized properly for some reason (insert GameWorks 2.0 conspiracy here).

Absolutely would not be shocked at all if non-RTX cards (or just AMD cards) are defaulting to "software" mode and trying to compute the rays on their shader units.

 

[Saylick]

https://twitter.com/x/status/1601036292380250112

Profiling results from RDNA2 running Portal RTX. TL;DR Only about 3% of the HW is being utilized properly for some reason (insert GameWorks 2.0 conspiracy here).

Welp, knowing Nvidia's history, I wouldn't put it past them for making Portal RTX a repeat of Crysis 2. It doesn't matter if it's an unoptimized mess as long as it makes Nvidia's latest GPUs look amazing in comparison to what came before it.

 

[Stuka87]

Welp, knowing Nvidia's history, I wouldn't put it past them for making Portal RTX a repeat of Crysis 2. It doesn't matter if it's an unoptimized mess as long as it makes Nvidia's latest GPUs look amazing in comparison to what came before it.

The weird thing is RTX doesn't look amazing. I mean, barely getting 60fps with DLSS3 (meaning 30real fps) on a video card that cost $1600 is a joke.

Valve should be ashamed of this.

Also, the reviews on this are hilarious: https://store.steampowered.com/app/2012840/Portal_with_RTX/

One of the reviewers had ChatGPT write a song in the style of "Still Alive" but with bad framerates, its fantastic.

 

[coercitiv]

Also, the reviews on this are hilarious: https://store.steampowered.com/app/2012840/Portal_with_RTX/

Top review delivers everything with one short punch.

 

[Ranulf]

"It's like Portal, but brighter, requires a driver update, and with smoke coming out of your GPU. Outstanding."

 

[psolord]

View attachment 72499

Dang. I knew that Lovelace coolers were big, but this really puts it into perspective. Why the heck does a 4080 need 450W of heat dissipation??

They need to justify the irresponsible amounts of pricing!

 

[Leeea]

One of the reviewers had ChatGPT write a song in the style of "Still Alive" but with bad framerates, its fantastic.

Steam Community :: chimera :: Review for Portal with RTX

@edit I asked ChatGPT to create a song in the style of "still alive" but with "horrible framerates". Here is the amazing result (unedited!) Verse 1: I'm not the same as I was before I used to run smooth, but now I'm a bore My graphics are choppy, my gameplay is slow My framerate is dropping and...

steamcommunity.com

That is epic.

 

[psolord]

Steam Community :: chimera :: Review for Portal with RTX

@edit I asked ChatGPT to create a song in the style of "still alive" but with "horrible framerates". Here is the amazing result (unedited!) Verse 1: I'm not the same as I was before I used to run smooth, but now I'm a bore My graphics are choppy, my gameplay is slow My framerate is dropping and...

steamcommunity.com

That is epic.

No, it's steam mate! xD

 

[Dribble]

Cloning is a bit overstated. AMD is forced to have a different approach. They have to be 'open'. There is almost 0 reason to add features because they will not be adopted by developers. There is maybe 15% of the market with AMD GPUs and only 10% of those who have the cool new GPU that can use the feature. What developer will implement that feature without a payoff? It only pays if the feature works on competitor's products too. A second player at RTG's inconsequential size does not get to add new features that require developer support. That is always a losing battle.

(Except on consoles - where Radeon did do some weird things like the Xbox 360's tessellator before Nvidia)

Most of the time "open" means "We are not going to put the effort into developing this properly", and "open" without strong investment from some company tends to fail. If AMD were actually to put the serious investment required into doing some new technique properly it probably wouldn't be open because they need to sell more AMD stuff off the back of it to pay for that investment, so it's got to run best on AMD hardware. AMD are not a charity after all.

That said it's not like AMD have never done this successfully - remember eyefinity (multi-monitor gaming). That wasn't exactly new tech but up to then it had been something you had to buy a pro card to use. Anyway they gave it to normal gamers and developed some proprietary tech to make it work better, and AMD fans loved it. Nvidia was forced to follow and release their version of the tech afterwards.

As to why AMD don't do more of this, well obviously it costs money to invest in new tech, but AMD have money now. The bigger problem is most of these things require as much software developement as hardware. AMD can do the hardware but they simply never invested in software - they always let people like Microsoft do it for them. That works fine in some cases - e.g. for consoles AMD provide the hardware and MS or Sony write the software, but relying on it has been more and more limiting as gpu's have evolved. Advanced gpu functionality, gpu compute for server farms, AI, all require a lot of software that no one else is going to write (other then a direct competitor who isn't going to give you fair access).

AMD really need to fix this and become advanced software as well as hardware experts.

 

[Kaluan]

Enable DLSS and enjoy 16 FPS experience on the rtx3080.


Well, 16 FPS is not single digit, so that is a massive improvement.


. . .
Even more interesting, on the 4000 series is that DLSS3 they are turning on? With its fake frames and latency issues? In a First person shooter game?


edit:
It appears that it is DLSS3! In a first person game!


It did not take long for journalists to abandon all pretension of integrity and jump onto nvidia's fake frame bandwagon. What a pathetic way to destroy TPUs reputation. Nvidia shills, nothing more.

Well, they also have the 8GB(!) 3070 as a "ideal" card for 4K up on their GPU database. But the 16GB 6800 somehow is just "acceptable" at 4K.

We're also getting closer and closer to 2 years since FSR was launched, but they still can't be bothered to add a FSR sub-section to their ray tracing section, as they have with DLSS when they review nVidia cards... since pretty much ever since DLSS was a thing.

Oh and don't get me started, they bent themselves backwards to give DLSS3 it's own section with every 4000 series review (so far) now.

Pretty obvious nVidia has a say in their review methodology.

 

[Ranulf]

Firestrike leaks via Vidcardz:

First AMD Radeon RX 7900 XTX/7900 XT 3DMark TimeSpy/FireStrike scores are in - VideoCardz.com

AMD Radeon RX 7900 in 3DMark We now have the very first confirmed 3DMark results of AMD’s upcoming Radeon 7900 series. The RX 7900 series cards perform just as expected and akin to what we know from RDNA2 series. This means better 3DMark FireStrike performance, which is a DirectX11 benchmark...

videocardz.com

7900XTX is 7% or so faster at 4k than a 4080. 7900XT is 1% under at 4080.

 

[TESKATLIPOKA]

Firestrike leaks via Vidcardz:

First AMD Radeon RX 7900 XTX/7900 XT 3DMark TimeSpy/FireStrike scores are in - VideoCardz.com

AMD Radeon RX 7900 in 3DMark We now have the very first confirmed 3DMark results of AMD’s upcoming Radeon 7900 series. The RX 7900 series cards perform just as expected and akin to what we know from RDNA2 series. This means better 3DMark FireStrike performance, which is a DirectX11 benchmark...

videocardz.com

7900XTX is 7% or so faster at 4k than a 4080. 7900XT is 1% under at 4080.

I will add the graphs from your link.

30% higher Time Spy score with Extreme(4K) Preset than RX 6950XT. Performance(1440p) preset is only 17% better. Both N31 are too close to each other in Time Spy.


23% higher Fire Strike score with Ultra(4K) Preset than RX 6950XT. Extreme (1440p) preset is only 24.5% better.

 

[Karnak]

Looks like either maybe some sort of driver/fw issue, way higher gaming than 3dmark perfomance gains or AMDs own benchmarks were not... that accurate.

 

[leoneazzurro]

Looks like either maybe some sort of driver/fw issue, way higher gaming than 3dmark perfomance gains or AMDs own benchmarks were not... that accurate.

Timespy score always favored NV architectures, however the fact that the two AMD 7900 cards are so close to each other in TS point to some sort of benchmark issue. XTX has practically 20% more than XT in everything except maybe frontend so being less than 10% more performant is strange (HW bugs or not).

 

[SteinFG]

Current rumor is that N32 will be released in late 2023, and the performance will be similar to RX 7900 XT.

https://twitter.com/x/status/1601218456883724289

RX 6000 is here to stay for quite a while it seems

 

[lixlax]

From easily matching/beating fully enabled AD102 we've now come down to this. Lots of people should have fully functional drivers now as well.

3 days to go for reviews.

 

[jpiniero]

Looks like either maybe some sort of driver/fw issue, way higher gaming than 3dmark perfomance gains or AMDs own benchmarks were not... that accurate.

The pricing makes sense now though if it ends up being about what it does in games. A tad faster in raster and worse in RT for $200 cheaper.

 

[PJVol]

XTX has practically 20% more than XT in everything except maybe frontend so being less than 10% more performant is strange

Vega 64 and 56 2.0 ?
TS scores ..., basically what heavily OC-ed 6800xt and 6900xt can do.
Pathetic... and nothing to eager this year.

 

[insertcarehere]

The pricing makes sense now though if it ends up being about what it does in games. A tad faster in raster and worse in RT for $200 cheaper.

If the firestrike numbers compared to 6950XT mean anything it's in question if this is faster than 4080 at all..

 

[eek2121]

The weird thing is RTX doesn't look amazing. I mean, barely getting 60fps with DLSS3 (meaning 30real fps) on a video card that cost $1600 is a joke.

Valve should be ashamed of this.

Also, the reviews on this are hilarious: https://store.steampowered.com/app/2012840/Portal_with_RTX/

One of the reviewers had ChatGPT write a song in the style of "Still Alive" but with bad framerates, its fantastic.

I mean, it is using a mod framework NVIDIA is developing and will eventually release. It isn’t a rewrite or anything.

I get around 35-40fps on ultra settings on my 3090.

There is definitely a visual fidelity improvement over the base game. The real time lighting and stuff is a definite improvement.

EDIT: you will almost never be able to tell how great RT looks from a picture since RT is all about the accuracy of lighting, reflections, shadows, etc. Down the road, having proper RT acceleration and native RT games will allow the rest of the GPU to be freed up for other things.

 

[leoneazzurro]

Vega 64 and 56 2.0 ?

Vega had issues with scaling due to register size and its Wave 64 mode unable to get enough instruction parallelism, plus some issues in the frontend, in fact even when "fixed" in the Radeon VII it had still gaming performance not up to par. But RDNA/RDNA2 had way less issues, and L0/L1 is even bigger in RDNA3 and Wave 64 should be single cycle so... it's quite strange.

 

[eek2121]

Vega 64 and 56 2.0 ?
TS scores ..., basically what heavily OC-ed 6800xt and 6900xt can do.
Pathetic... and nothing to eager this year.

Are you really taking a leaked result and making that claim? How about we wait for reviews instead.