Question Speculation: RDNA2 + CDNA Architectures thread

[soresu]

Interesting and important news, especially for running RT games on Linux or on Windows through Vulkan!

The vendor neutral RT extension for Vulkan just went gold in the latest revision of the Vulkan standard.

Link here.

We've been expecting this to come on the tails of AMD's first RT accelerating architecture, so it's good it's finally out there.

As to whether it will replace the nVidia specific RT extensions used in the Vulkan RT implementation for Quake 2 RTX - well I doubt that nVidia will be in any hurry to let that comparison happen unless they are 100% certain it plays well for them.

 

[kurosaki]

Interesting and important news, especially for running RT games on Linux or on Windows through Vulkan!

The vendor neutral RT extension for Vulkan just went gold in the latest revision of the Vulkan standard.

Link here.

We've been expecting this to come on the tails of AMD's first RT accelerating architecture, so it's good it's finally out there.

As to whether it will replace the nVidia specific RT extensions used in the Vulkan RT implementation for Quake 2 RTX - well I doubt that nVidia will be in any hurry to let that comparison happen unless they are 100% certain it plays well for them.

Do we have any good comparison between the different RT-archs? Nvidia Vs AMD. It all seems so unclear who will have a good lead in 6 months. On paper, AMD has more RT cores than Nvidia in some comparisons, but still fall short in performance. In very specific unbiased benchmarks. Godfall and Dirt 5 is "Only" RT shadows, so not that impressed. Yet.

 

[GodisanAtheist]

Are we going through another generation where ray tracing is still a gimmick feature?

I am getting the impression that Ray Tracing is only useful to add nice lighting to games that did not have it to begin with.

- I feel like adding in RT effects to the slew of remasters coming out is a no brainier. Mass Effect Legacy, for example, where the original games relied so heavily on lighting that a good RTRT implementation would not only make the games look better but give scenes a completely different feel.

Older games seem like the perfect candidates thanks to the hige FPS baseline to work with as well.

So long as studios don't completely butcher it like the Crisis remaster, there be gold in them hills.

 

[moinmoin]

Are we going through another generation where ray tracing is still a gimmick feature?

I am getting the impression that Ray Tracing is only useful to add nice lighting to games that did not have it to begin with.

I expect it to be used more and more widely this gen, it's a differentiation feature on the next gen consoles after all. But the "it" won't be full out RT with all bells and whistles, instead different accelerated parts like path tracing will be used to improve conventional lighting models, as already happens with RT shadows.

 

[Stuka87]

Do we have any good comparison between the different RT-archs? Nvidia Vs AMD. It all seems so unclear who will have a good lead in 6 months. On paper, AMD has more RT cores than Nvidia in some comparisons, but still fall short in performance. In very specific unbiased benchmarks. Godfall and Dirt 5 is "Only" RT shadows, so not that impressed. Yet.

Existing games (outside godfall and Dirt 5) were all coded for nVidia and do not contain any AMD optimizations. We wont have better idea on performance until existing games get updated, and more new games come out.

Maybe one day Anandtech will release a deep dive or, well, anything at this point.

 

[Mopetar]

This is a tiny bit simplistic, I think? Both companies have one huge compute focused chip.

The biggest difference is that NV have put a lot of machine learning acceleration into nearly their entire line up, from A100 down. There's a huge market for that these days.

As to whether you think that's worth it for their gaming chips? You have to decide if you're going to 'allow' DLSS or not. If not, then no. If you do then its hugely worthwhile in terms of the extra performance.

It is simplistic, but that doesn't make it any less true. DLSS is in some ways a product of Nvidia justifying the inclusion of tensor cores in their mainstream GPUs so that they don't sit idle until some game figures out an actual use for them. That it helps cover for lackluster RT capabilities is a bit of extra icing on the cake.

Sure there's a big market for ML, which is why Nvidia can make a mint on each GA-100 they sell, but that market doesn't intersect with traditional gaming without a contrived solution to a different problem that should have been addressed head on.

Calling it all worthwhile is like the battered wife saying that at least her husband brings her flowers after bouncing her off the walls. If we're going to mix our metaphors everyone would just be better off without something so performance hitting.

Give us a pure gaming focused architecture that doesn't need DLSS and I'll be far more happy. Hopefully whatever open standard AMD comes up with kills off DLSS so Nvidia can focus more on pure RT performance and not pushing another vendor proprietary solution that I don't want as a fix to performance problems in the first place.

I have a feeling that Nvidia will head in that direction anyway. With AMD actually competing they can't afford to skimp by with less than their absolute best. I

 

[Mopetar]

Do we have any good comparison between the different RT-archs? Nvidia Vs AMD. It all seems so unclear who will have a good lead in 6 months. On paper, AMD has more RT cores than Nvidia in some comparisons, but still fall short in performance. In very specific unbiased benchmarks. Godfall and Dirt 5 is "Only" RT shadows, so not that impressed. Yet.

The only RT comparison we have are for titles basically optimized for one architecture or the other. They run however well on the cards they're optimized for and like garbage on anything else.

Current results are only useful to anyone wanting to use the technology right now and not for any kind of future performance

- I feel like adding in RT effects to the slew of remasters coming out is a no brainier. Mass Effect Legacy, for example, where the original games relied so heavily on lighting that a good RTRT implementation would not only make the games look better but give scenes a completely different feel.

Older games seem like the perfect candidates thanks to the hige FPS baseline to work with as well.

So long as studios don't completely butcher it like the Crisis remaster, there be gold in them hills.

The RT performance is probably as much of a performance hit on old titles as it is new. Maybe the rest of the graphics being ancient means you can just run it at 1080p instead of trying to upscale it for no real benefit unless you've rally drunk the marketing koolaid.

 

[beginner99]

Give us a pure gaming focused architecture that doesn't need DLSS and I'll be far more happy. Hopefully whatever open standard AMD comes up with kills off DLSS so Nvidia can focus more on pure RT performance and not pushing another vendor proprietary solution that I don't want as a fix to performance problems in the first place.

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

 

[Kenmitch]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

As far as I'm concerned....you left off dung!

 

[moinmoin]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

That elephant is on a sunny island both AMD and Nvidia can't reach anyway though.

 

[Leeea]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

At a $900 + walled garden so Apple can fork you over the coals on every software purchase, there is nothing low end on how much that GPU costs. The mini is firmly in discrete GPU territory price wise, but has a pathetic APU.

It is over 3x the cost of its in tier AMD/Intel APU competitors, and does not even include the monitor, keyboard, and mouse. When you throw in closed standards, lack of long term support, and lack of expandability the value prospect goes right out the window.


The mini sits in an odd spot, it is to expensive to be low end, but it does not have the performance or specs to match its price bracket. $900 only gets you 16 GB of RAM, 512 GB SSD, and a iGPU.

It is a fashion statement today, and the day Apple drops support, an e-waste brick.

 

[kurosaki]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

Tile based renderer though. Differs quite a lot from our immidiate mode tile based deferred renderers at NV and AMD. Think pros is efficiency and low mem bandwidth demand, cons is roughly about everything else.

 

[Stuka87]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

You can't really compare a custom built SoC to an APU that has to work on existing motherboards. The M1 doesn't have to fit into any existing platforms. So lots of things can be done to improve performance that cannot be done on ATX motherboards.

If you want to compare the M1 to something, compare it to the SoC in the Nintendo switch which offers great performance for its form factor, or even the SoC used in the big consoles. Yes the big consoles use way more power, but they also use a fully custom SoC.

 

[Mopetar]

Not RT related but the elephant in the room is also Apples M1 iGPU which blows out any AMD iGPU or low end from both vendors especially also in terms of performance/watt.

Yeah, but can it run Crysis?

That is sort of a serious question in a sense because it really doesn't matter how efficient it is compared to AMD/Nvidia if Apple never scales it up to a level where it competes with AMD/Nvidia. Apple caters to a more casual gaming market that isn't interested in running games at 4K ultra and would only want to pay $$$$ for a GPU if it had certified professional grade drivers for professional work.

If I were AMD or Nvidia I wouldn't worry about Apple going after any of the gaming market. Instead I'd worry if they decide to target the professional market that buys Quadros for CAD work. That sounds a lot more like the kind of market that Apple might want to go after, especially since it lets them get the kind of high-end margins that they want to target.

 

[Hitman928]

Yeah, but can it run Crysis?

That is sort of a serious question in a sense because it really doesn't matter how efficient it is compared to AMD/Nvidia if Apple never scales it up to a level where it competes with AMD/Nvidia. Apple caters to a more casual gaming market that isn't interested in running games at 4K ultra and would only want to pay $$$$ for a GPU if it had certified professional grade drivers for professional work.

If I were AMD or Nvidia I wouldn't worry about Apple going after any of the gaming market. Instead I'd worry if they decide to target the professional market that buys Quadros for CAD work. That sounds a lot more like the kind of market that Apple might want to go after, especially since it lets them get the kind of high-end margins that they want to target.

I'm no graphics architecture expert, but from my understanding, the M1 GPU being a tile based rendering architecture benefits greatly (especially compared to Intel and AMD APUs) from having the large on die system cache. I'm sure the tile size on the M1 was chosen specifically around the on die cache and the on die cache doesn't have to be incredibly large because the GPU is using tile based rendering. This allows the GPU to be efficient in both terms of power and especially memory bandwidth. However, as it is asked to render scenes with more complex geometry, it will have to spend more and more time sorting.

Additionally, as resolution increases, you now have a lot more tiles to handle and I think with modern techniques for avoiding culled geometry, IMR based GPUs will scale a lot better with really high resolutions. So for lower resolutions and simpler geometry games, the M1 based GPUs should "punch above their weight", but when you get games with a lot of complexity at high resolutions, it won't be able to keep up and even a scaled up version will under perform. Others who have a deeper knowledge can correct me if I am mistaken.

 

[Qwertilot]

Calling it all worthwhile is like the battered wife saying that at least her husband brings her flowers after bouncing her off the walls. If we're going to mix our metaphors everyone would just beive us a pure gaming focused architecture that doesn't need DLSS and I'll be far more happy. Hopefully whatever open standard AMD comes up with kills off DLSS so Nvidia can focus more on pure RT performance and not pushing another vendor proprietary solution that I don't want as a fix to performance problems in the first place.

I have a feeling that Nvidia will head in that direction anyway. With AMD actually competing they can't afford to skimp by with less than their absolute best. I

You can't have your cake and eat it here.

Even if/(when) you get a vendor neutral solution to neural network upscaling it won't give you anything like the same speed up without dedicated machine learning hardware.

NV would probably say this is their best, especially for ray tracing - the two things do go quite closely together. It's the only real hope for plausible RT performance right now.

 

[Leeea]

Even if/(when) you get a vendor neutral solution to neural network upscaling it won't give you anything like the same speed up without dedicated machine learning hardware.

Why would it need to be neural network? DLSS comes with an expensive developer cost and limited support. Add in neural network artifacting in games where developers did not spend time hand tuning it, and it is less then ideal. It could be claimed some results are inferior to no upscaling at all*.


Something like 2xSal would be a big step up from bicubic, and allow upscaling for all games.
Another approach would be to using AA style techniques to upscale, something Amd has already hinted at.


I would be completely shocked if AMDs upscaling solution uses neural networking at all.

* https://bit.ly/3pTLafJ
( I detest url shorteners, but I detest awful crap forum software that insists on mangling my post with automatic reddit unfurl )

 

[Mopetar]

You can't have your cake and eat it here.

Even if/(when) you get a vendor neutral solution to neural network upscaling it won't give you anything like the same speed up without dedicated machine learning hardware.

NV would probably say this is their best, especially for ray tracing - the two things do go quite closely together. It's the only real hope for plausible RT performance right now.

Well they would say it is their best because who wants to buy something even the seller will full admit is substandard? I don't even think an upscaling solution needs to be an ML-based implementation to succeed either. That's really just artificially limiting it to some arbitrary decision by Nvidia for no good reason. It's just saying that upscaling needs to exist to solve the problem of sub-standard ray tracing performance and that's just an excuse to cover up for poor performance. I'd rather have something that exists to scale older games to native display resolutions.

The real solution to the current problem is that developers should reduce the amount of RT being added to their games so that it doesn't absolutely tank the performance. At least that gets them to start familiarizing themselves with implementing it and it allows for the hardware to evolve to a point where it is good enough to allow for more RT to be used. It's pretty obvious that RT is still in its infancy and what we see a decade from now will probably be radically different from the present architectures. If you look back at traditional rasterized graphics implementations, cards used to have separate pixel and vertex shaders.

I expect we'll eventually get some watershed moments with RT assuming it isn't a fad and it may well not be since we're really reaching the point of diminishing returns with traditional graphics and resolutions so we may as well devote processing power to a better technique.

 

[kurosaki]

Well they would say it is their best because who wants to buy something even the seller will full admit is substandard? I don't even think an upscaling solution needs to be an ML-based implementation to succeed either. That's really just artificially limiting it to some arbitrary decision by Nvidia for no good reason. It's just saying that upscaling needs to exist to solve the problem of sub-standard ray tracing performance and that's just an excuse to cover up for poor performance. I'd rather have something that exists to scale older games to native display resolutions.

The real solution to the current problem is that developers should reduce the amount of RT being added to their games so that it doesn't absolutely tank the performance. At least that gets them to start familiarizing themselves with implementing it and it allows for the hardware to evolve to a point where it is good enough to allow for more RT to be used. It's pretty obvious that RT is still in its infancy and what we see a decade from now will probably be radically different from the present architectures. If you look back at traditional rasterized graphics implementations, cards used to have separate pixel and vertex shaders.

I expect we'll eventually get some watershed moments with RT assuming it isn't a fad and it may well not be since we're really reaching the point of diminishing returns with traditional graphics and resolutions so we may as well devote processing power to a better technique.

And that tech surely isn't DLSS...

 

[Saylick]

Well they would say it is their best because who wants to buy something even the seller will full admit is substandard? I don't even think an upscaling solution needs to be an ML-based implementation to succeed either. That's really just artificially limiting it to some arbitrary decision by Nvidia for no good reason. It's just saying that upscaling needs to exist to solve the problem of sub-standard ray tracing performance and that's just an excuse to cover up for poor performance. I'd rather have something that exists to scale older games to native display resolutions.

The real solution to the current problem is that developers should reduce the amount of RT being added to their games so that it doesn't absolutely tank the performance. At least that gets them to start familiarizing themselves with implementing it and it allows for the hardware to evolve to a point where it is good enough to allow for more RT to be used. It's pretty obvious that RT is still in its infancy and what we see a decade from now will probably be radically different from the present architectures. If you look back at traditional rasterized graphics implementations, cards used to have separate pixel and vertex shaders.

I expect we'll eventually get some watershed moments with RT assuming it isn't a fad and it may well not be since we're really reaching the point of diminishing returns with traditional graphics and resolutions so we may as well devote processing power to a better technique.

I have a feeling that you're right. If history has taught us anything about GPU architectures, we've been moving away from fixed function units on any part of the graphics pipeline that requires computation; RT is no exception. If I'm not mistaken, at some point in the future, portions of the traditional rasterization pipeline be completely tossed out and replaced with a RT pipeline if the RT hardware is fast enough. If this is true, then as we start approaching the other end of the traditional-RT spectrum, it seems that it makes more sense to have flexible execution units that can accommodate both pipelines rather than having dedicated units that only help with one or the other. Your example of pixel and vertex shaders is a perfect example of how those units were completely replaced with a unified shader model, and while I was too young to own any major graphics card back in the day with pixel and vertex shaders, I do know that various games would run faster on certain GPUs simply because the ratio of pixel:vertex shaders were more favorable for that game. Once we moved towards a unified shader architecture, performance became more consistent from game to game. I can see something similar happening for GPUs moving forward whereby all of the execution units can be used for a game that only uses traditional rasterization, all of them can be used for a game that relies on purely RT, and a blend if the game uses a hybrid approach.

 

[GodisanAtheist]

The RT performance is probably as much of a performance hit on old titles as it is new. Maybe the rest of the graphics being ancient means you can just run it at 1080p instead of trying to upscale it for no real benefit unless you've rally drunk the marketing koolaid.

- A 50% performance hit on 300FPS is preferable to a 50% hit on 60 FPS.

Remasters should be the RTRT playground.

Cheap and easy way to make something old new again...

 

[Saylick]

I thought this was an interesting video comparing RDNA 2 and Ampere. He, too, thinks that RDNA 2 is the more elegant and efficient architecture, and believes that AMD's RTRT implementation just needs more optimization to fully leverage the Infinity Cache and the higher clocked traversal units. He also has a slide of Time Spy Extreme scores relative to TFLOPS and it's apparent that Ampere just can't leverage all of it's raw TFLOP count (no surprise here). Since the Infinity Cache is used for RT, I can't wait to see what RDNA 3 has in store, assuming it's when AMD moves to a chiplet architecture. I would totally expect to see more Infinity Cache per Shader Engine with maybe even twice as many TA units.

 

[GoodRevrnd]

While I could give two ploops about RT shadows in Dirt 5, it's interesting RT shadows are pretty impactful in this game, moreso in terrain. Actually I don't care about CoD either, but it's at least a more meaningful tech demo. Wonder if the faster RT on XSX is due to more shaders or better / more mature API. Would like to see benches on PC!

 

[soresu]

Do we have any good comparison between the different RT-archs? Nvidia Vs AMD. It all seems so unclear who will have a good lead in 6 months. On paper, AMD has more RT cores than Nvidia in some comparisons, but still fall short in performance. In very specific unbiased benchmarks. Godfall and Dirt 5 is "Only" RT shadows, so not that impressed. Yet.

It seems like AMD's RT acceleration uArch plays better to the new DXR 1.1 inline raytracing feature from what I've heard, though I may have gotten things muddled there.

So basically a lot of the older 'RTX' games using DXR 1.0 probably won't play so well, but newer games might tip the balance some.

Of course that is if nVidia doesn't pull a DX10.1 and somehow discourage most game devs from using those new features, which will be much harder this time given AMD's absolute console RT hardware dominance for the time being until Nintendo do a Switch 2 with RT.

 

[jrdls]

If history has taught us anything about GPU architectures, we've been moving away from fixed function units on any part of the graphics pipeline that requires computation; RT is no exception

Pardon the interruption but, if the talk by Bill Dally, chief scientist at Nvidia, is anything to go by, I don't think Nvidia will move away from fixed function units to handle raytracing anytime soon, nor AMD BTW.

He makes a compelling case for them as he argues that that type of unit is the way to continue to improve performance per watt as Dennard scaling is no more. I think AMD is aware of this and they will introduce more of those types of units on RDNA3 as future nodes would allow them to dedicate more transistors to them and further improve on this metric.

Not that it matters, but I believe AMD did the right thing when they decided to go the IC route so as to keep the data on chip and not go to memory as much. In the future they'll have to introduce specialised hardware judiciously if they intend to keep up with the trend they currently have.