News Balking AMD's Claims, Nvidia Says Smartphones Aren't Ready for Ray Tracing

[-Fran-]

People most often make generalization when it comes to VRAM situation between AMD and nvidia but in reality it is a bit more complex. But the short version is nvidia to certain extend can get away with less amount of VRAM than AMD.

Ah, there's a myth like that? I kid, I kid.

The data compression for "on the fly" transmission is good, but it doesn't make it so they can get away with having half the ram or those werabaouts. It may make nVidia have a 20% edge? That's on the very best case scenario of the compression IIRC. Still, both AMD and nVidia are still bandwidth starved anyway, so having a bigger buffer (something AMD realized and implemented) helps a ton; specially on the EOL part of the card. Look at the RX480's and 1060's with half the VRAM and how they fare in games today vs their 8/6 GB siblings. While the 1060 was a borderline scam as it was a different GPU, AMD's clocked their 4GB variants slightly slower as well to make them a bit more different. Point is: 1060 3GB is dead in the water now even with the compression advantage you mention over the RX480 with 4GB. I mean, I'm still using my RX480 for VR and it's capped all the time with >3.8GB usage. I'm just projecting down the line here, even with the advantage you mention, it is just not enough to make up for more than 50% extra VRAM on AMD's side; that's just, more or less, a fact.

Also, please don't become Stans of any company. Not AMD's, not nVidia's, not Intel. None of them want to babysit you and make you happy: they all want your hard earned dollars, so make sure you make them work for them. You all sound like beaten up wives saying "BUT HE LOVES ME!". Yeah, ok... If you're happy, I guess?

Regards.

 

[watzupken]

I feel Nvidia jumped in to make the comments because they don't have the opportunity to bring RT to mobile so far.

Having said that, I don't think we will see RT lighting, shadows, etc, implemented on mobile devices anytime soon. Considering a PC that requires hundred of watts of power struggles in RT enabled games, it is impossible to implement (with existing technology) RT on hardware that ideally should utilise sub 10W kind of power. However, it is factual that AMD's RDNA 2 contains a Ray Accelerator tightly coupled per CU, which means the hardware is there. So for them to announce that they are bringing RT to mobile is not incorrect. Whether its being used or not is another story. I am sure the RA will end up on the RDNA2 iGPUs, but of course, we know that is not going to give you any playable experience if you enable RT.

 

[renz496]

as i said to certain extend. just look what happen with horizon zero dawn for 4GB card between nvidia and AMD:

View: https://www.youtube.com/watch?v=693NBRwk8z4


in TPU test when things being push to the limit (4k) AMD 4GB card unable to load the game at all.

another game that we saw how lack of VRAM can impact AMD cards:

4k result:

look at RX5600XT vs RX580 or RX590.

 

[watzupken]

as i said to certain extend. just look what happen with horizon zero dawn for 4GB card between nvidia and AMD:

View: https://www.youtube.com/watch?v=693NBRwk8z4


in TPU test when things being push to the limit (4k) AMD 4GB card unable to load the game at all.

another game that we saw how lack of VRAM can impact AMD cards:

4k result:

look at RX5600XT vs RX580 or RX590.

And your point being? At 4K, none of the lower end cards can make it. While AMD may crash to the bottom of the chart, the comparable cards are all displaying a slide show, not quite a game.

 

[hotaru.hino]

I feel Nvidia jumped in to make the comments because they don't have the opportunity to bring RT to mobile so far.

Having said that, I don't think we will see RT lighting, shadows, etc, implemented on mobile devices anytime soon. Considering a PC that requires hundred of watts of power struggles in RT enabled games, it is impossible to implement (with existing technology) RT on hardware that ideally should utilise sub 10W kind of power. However, it is factual that AMD's RDNA 2 contains a Ray Accelerator tightly coupled per CU, which means the hardware is there. So for them to announce that they are bringing RT to mobile is not incorrect. Whether its being used or not is another story. I am sure the RA will end up on the RDNA2 iGPUs, but of course, we know that is not going to give you any playable experience if you enable RT.

It's not impossible for real-time ray tracing to happen on a mobile power envelope. Imagination showed us this 5 years ago.

 

[renz496]

And your point being? At 4K, none of the lower end cards can make it. While AMD may crash to the bottom of the chart, the comparable cards are all displaying a slide show, not quite a game.

my point is on how lack of VRAM affecting AMD card. and the original video from youtube was tested on 1080p. alsoyou can read the article done by TPU. in watch dog legion RX5600 only perform as it should at 1080p. even at 1440p it was beaten by RX580/590 because the game use around 6500mb of VRAM at that res. such issue did not happen on nvidia 6GB card.

 

[hotaru.hino]

my point is on how lack of VRAM affecting AMD card. and the original video from youtube was tested on 1080p. alsoyou can read the article done by TPU. in watch dog legion RX5600 only perform as it should at 1080p. even at 1440p it was beaten by RX580/590 because the game use around 6500mb of VRAM at that res. such issue did not happen on nvidia 6GB card.

The problem with anything reporting VRAM is it's not indicative of what's actually "needed." All of the reporting metrics at best can tell us is how much VRAM was allocated to the application. It doesn't really tell us how much VRAM is actually in use.

 

[-Fran-]

as i said to certain extend. just look what happen with horizon zero dawn for 4GB card between nvidia and AMD:

View: https://www.youtube.com/watch?v=693NBRwk8z4


in TPU test when things being push to the limit (4k) AMD 4GB card unable to load the game at all.

another game that we saw how lack of VRAM can impact AMD cards:

4k result:

look at RX5600XT vs RX580 or RX590.

Those graphs are useful and all, but if a game fails to load, it means there's a problem in the game engine more than the card. Most "out of memory" errors are just bad programming practices when you have fallback mechanisms to rely on. Think RAM vs virtual/swap.

And as hotaru.hinu pointed out, reported VRAM doesn't always mean "used" VRAM. It's a semantic distinction, but an important one to make. You will only "see" when a game struggles with VRAM when 2 cards using the same GPU and different capacities do not perform the exact same. The video is comparing different generations of GPUs at 4GB, but the comparison should be with the same generation using different capacities. So, the interesting ones are: 1060 3GB vs 6GB and RX570 4GB vs rx580 8GB; these two pairs are close enough to make the VRAM capacity difference the real distinction point on how they perform.

Unfortunately, after all, this doesn't work as evidence to your point, although I do understand what you meant. I'm just not keen on accepting this as justification to give nVidia a pass on cheaping out on VRAM capacities this generation.

Regards.

 

[renz496]

Those graphs are useful and all, but if a game fails to load, it means there's a problem in the game engine more than the card. Most "out of memory" errors are just bad programming practices when you have fallback mechanisms to rely on. Think RAM vs virtual/swap.

And as hotaru.hinu pointed out, reported VRAM doesn't always mean "used" VRAM. It's a semantic distinction, but an important one to make. You will only "see" when a game struggles with VRAM when 2 cards using the same GPU and different capacities do not perform the exact same. The video is comparing different generations of GPUs at 4GB, but the comparison should be with the same generation using different capacities. So, the interesting ones are: 1060 3GB vs 6GB and RX570 4GB vs rx580 8GB; these two pairs are close enough to make the VRAM capacity difference the real distinction point on how they perform.

Unfortunately, after all, this doesn't work as evidence to your point, although I do understand what you meant. I'm just not keen on accepting this as justification to give nVidia a pass on cheaping out on VRAM capacities this generation.

Regards.

i'm not giving nvidia the free pass on the VRAM part. but they are very well known to give you just enough. and when they dictate for example 3080 should be enough with 10GB only at 4k it might be true for their card. but most often people make simple assumption that VRAM situation are the same between AMD and nvidia. and the video that make the comparison are done on 4GB cards only because back then it is the issue for AMD 4GB card that did not exist on 8GB version of the card. then they just compare it to nvidia 4GB cards. and in case of watch dogs legion while the performance end up where it should be at 1080p at 1440p and 4k it was beaten by RX580 and RX590 that has more VRAM but should be slower on raw performance. so you still want to blame game engine issue rather than the card amount of VRAM?

this is what TPU said in their conclusion for watch dog legion:

Because of the high-res texture pack, which is an optional download, we can finally make use of all the VRAM paired with modern GPUs. Even at 1080p, Watch Dog Legion will use more than 6 GB, which is a good thing as textures are much improved. NVIDIA cards with 3 GB and 4 GB VRAM do much better than their AMD counterparts in this memory constrained situation. For example, the 4 GB RX 5500 XT drops to only 12 FPS, whereas the GTX 1030 3 GB runs twice as fast at 22 FPS. This just confirms what we've been seeing in such scenarios previously—NVIDIA handles memory pressure much better than AMD.

https://www.techpowerup.com/review/watch-dogs-legion-benchmark-test-performance-analysis/7.html

 

[-Fran-]

i'm not giving nvidia the free pass on the VRAM part. but they are very well known to give you just enough. and when they dictate for example 3080 should be enough with 10GB only at 4k it might be true for their card. but most often people make simple assumption that VRAM situation are the same between AMD and nvidia. and the video that make the comparison are done on 4GB cards only because back then it is the issue for AMD 4GB card that did not exist on 8GB version of the card. then they just compare it to nvidia 4GB cards. and in case of watch dogs legion while the performance end up where it should be at 1080p at 1440p and 4k it was beaten by RX580 and RX590 that has more VRAM but should be slower on raw performance. so you still want to blame game engine issue rather than the card amount of VRAM?

this is what TPU said in their conclusion for watch dog legion:

Because of the high-res texture pack, which is an optional download, we can finally make use of all the VRAM paired with modern GPUs. Even at 1080p, Watch Dog Legion will use more than 6 GB, which is a good thing as textures are much improved. NVIDIA cards with 3 GB and 4 GB VRAM do much better than their AMD counterparts in this memory constrained situation. For example, the 4 GB RX 5500 XT drops to only 12 FPS, whereas the GTX 1030 3 GB runs twice as fast at 22 FPS. This just confirms what we've been seeing in such scenarios previously—NVIDIA handles memory pressure much better than AMD.

https://www.techpowerup.com/review/watch-dogs-legion-benchmark-test-performance-analysis/7.html

Yes, that's what I said when I replied to you originally: nVidia handles memory better due to their on-the-fly compression, so it makes sense that it would handle memory constrained scenarios better; no surprises* there. How they allocate assets in memory is a completely different topic and it will depend on how each game implements the engine, so it's not apples to apples necessarily and hence why I said the comparisons are better made per-GPU pairs with different capacities.

I guess you could say, this gen is the equivalent of someone getting a 1060 3GB instead of a RX580 8GB? That is the bulk of my argument, really. It is not more complicated than that.

Regards.

 

[hotaru.hino]

Yes, that's what I said when I replied to you originally: nVidia handles memory better due to their on-the-fly compression, so it makes sense that it would handle memory constrained scenarios better; no surprises* there. How they allocate assets in memory is a completely different topic and it will depend on how each game implements the engine, so it's not apples to apples necessarily and hence why I said the comparisons are better made per-GPU pairs with different capacities.

I guess you could say, this gen is the equivalent of someone getting a 1060 3GB instead of a RX580 8GB? That is the bulk of my argument, really. It is not more complicated than that.

Regards.

AMD had compression from Polaris onward (https://www.amd.com/system/files/documents/polaris-whitepaper.pdf). Or at least AMD started using it aggressively.

I'd argue that NVIDIA employing tiled rasterization since Maxwell was a bigger factor. AMD started using it in Vega. Tiled rasterization significantly cuts down on memory bandwidth needs.

 

[-Fran-]

AMD had compression from Polaris onward (https://www.amd.com/system/files/documents/polaris-whitepaper.pdf). Or at least AMD started using it aggressively.

I'd argue that NVIDIA employing tiled rasterization since Maxwell was a bigger factor. AMD started using it in Vega. Tiled rasterization significantly cuts down on memory bandwidth needs.

Ugh, ok. AMD has been using compression techniques for a good while as well, but they've always been less spectacular than nVidia. It's not like they haven't had any at all. GCN came out with a form of compression, but it was like the most basic things you could do with colours or something? nVidia works their compression on more areas I can't remember, but it is quite a bit better; again, i can't remember what made it so much better than AMD's. Still, Polaris did improve greatly, yes, but not enough to catch up to nVidia's way of moving data from and to the GPU. I don't want to read through all the technical sheets released for each gen, but I'm sure it won't be too hard to find articles that compared the two at some point?

And I can't remember what Tiled Raster did; I'll have to read on it.

Regards.

 

[hotaru.hino]

And I can't remember what Tiled Raster did; I'll have to read on it.

If you want the tl;dr, it breaks up the frame into tiled chunks (usually 16x16 or 32x32) and rendering is performed on those chunks. The immediate benefit is these tiled chunks are smaller to pass around. The secondary benefit is that polygon sorting usually happens here to figure out which polygons are seen in the tiles. This can result in less overdraw.

Explanation in the OG hardware: https://www.anandtech.com/show/558/4

 

[-Fran-]

If you want the tl;dr, it breaks up the frame into tiled chunks (usually 16x16 or 32x32) and rendering is performed on those chunks. The immediate benefit is these tiled chunks are smaller to pass around. The secondary benefit is that polygon sorting usually happens here to figure out which polygons are seen in the tiles. This can result in less overdraw.

Explanation in the OG hardware: https://www.anandtech.com/show/558/4

Ah, I remember now. Yes.

Isn't this used by AMD and nVidia now, though?

Regards.

 

[hotaru.hino]

Ah, I remember now. Yes.

Isn't this used by AMD and nVidia now, though?

Regards.

Yes, that's what I mentioned. NVIDIA's used this since Maxwell 2, AMD since Vega.

 

[-Fran-]

Yes, that's what I mentioned. NVIDIA's used this since Maxwell 2, AMD since Vega.

Ah, you did, yes. Apologies as I've missed that.

Regards.