Next Generation Hardware Speculation with a Technical Spin [post E3 2019, pre GDC 2020] [XBSX, PS5]

[iroboto]

We can safely postulate the level of their RT support to not exceed an RTX 2070 Super

I wouldn't postulate that. 20xx series are their first crack at RT. I suspect their second attempt will be significantly improved now that they see what developers are doing with it.
That being said, everyone arriving later should be able to
a) see how it performs on production game code (and what developers actually do with it)
b) make modifications where necessary to improve performance (at least wrt) beating the performance of nvidia's first generation of RT acceleration.

 

[iroboto]

RT only solves visibility, you still have to do the shading. While it might reduce some of the compute you need to calculate some stuff, like shadow cascades for one, it also adds its a lot of compute load on its own.

Well i'm not necessarily sure that's true at all.
If we separate the triangle intersection and the updating of the acceleration structure which is the bulk of the speed drop off and looking purely at the shaders ; one could probably make a case that the shaders operating on the triangles directly are probably simpler and less work than standard T&L ones as you remove out all the _sorting_, exception processing, weird cases, weird hacks etc.

But on the topic of cascaded shadows vs RT shadows: the difference is pretty staggering in terms of load drop off.

An older article but still likely applicable today:
https://www.imgtec.com/blog/ray-traced-shadows-vs-cascaded-shadow-maps/

Over cascaded shadow maps:
50% bandwidth reduction
50% bandwidth footprint
rendering/compute time 50% less

 

[RDGoodla]

According to the report yesterday, does the haptic feedback in PS5 controller has anything different from HD Rumble in switch’s joycon?

 

[Miniature Kaiju]

Well i'm not necessarily sure that's true at all.
If we separate the triangle intersection and the updating of the acceleration structure which is the bulk of the speed drop off and looking purely at the shaders ; one could probably make a case that the shaders operating on the triangles directly are probably simpler and less work than standard T&L ones as you remove out all the exception processing, weird cases, weird hacks etc.

These machines aren't doing RT at a level that can get it around traditional raster rendering pipeline complexity. We're still pretty far from that dreamland.

 

[iroboto]

These machines aren't doing RT at a level that can get it around traditional raster rendering pipeline complexity. We're still pretty far from that dreamland.

I updated my post above, that's an older Power VR article there which has been a proponent of hybrid RT/rasterization forever. That's a significant load reduction there. I'm sure we can find other instances of things like that.

 

[Shifty Geezer]

It would be an absolute marketing disaster if nothing else.

The only thing that sells consoles is the games and what's on the screen. No-one cares how many flops a machine can do if it looks better than anything else. Hell, what if it's a super-processor with one single unified instruction "RENDERSCREEN" that can generate photorealistic graphic in a single floating point operation? OMG it only runs at 60 FLOPS!!!!

RT only solves visibility, you still have to do the shading. While it might reduce some of the compute you need to calculate some stuff, like shadow cascades for one...

Shadow cascades? Shadows are one of the things raytracing solves, with no need for cascades, and none of their inaccuracies.

Some people are being really myopic about these numbers and interpreting them purely in relation to how current graphics are produced. We're talking about a new hardware generation with a potential paradigm shift and ground-up hybrid renderers using tech far more optimally. The very way geometry is constructed, lit and shaded could be different. We have zero knowledge on Sony's RT solution and enough reasons to doubt it matches RTX that we shouldn't be drawing any parallels. Numbers only tell a small part of the story. Plenty of contributors have been long enough on this board that they should know that by now and not be jumping to conclusions, although I'm aware some people prefer to form conclusions for everything instead of working with open uncertainties and not collapsing them to a final conclusion when there's enough data to create that robust conclusion.

As tot the technical debate, how much does RT hardware add to RTX? Isn't it something like 10%? What could be achieved on a large GPU with 50% of it dedicated to raytracing? Someone compare the rumoured die-sizes (300 mm²) with other GPUs and determine how much of that die would likely be 7TFs of compute and give us a die size for the RT side of things. That'll give some impression of what might be going on under the hood of this hypothetical machine.

Personally, I find the possibility quite exciting. If it is only 7TFs and a considerable amount of RT hardware, it means Sony for one are going whole hog and that might mean a true generational advance in games, not just in rendering which would be prettier, but also development if devs don't need to keep refining hacks. Though cross-platform development will be a challenge.

 

[Rootax]

I wonder how much they can go "custom RT", in a sense that they have to be compatible with what they're doing on PC and futur Xbox.

 

[iroboto]

I wonder how much they can go "custom RT", in a sense that they have to be compatible with what they're doing on PC and futur Xbox.

as custom as they'd like to. It's a wide open field here. The pipeline for RT is very simple, so the customization will go towards improving the performance at each step.

If customization means changing the pipeline, then yes, I suppose there could be some issues.

 

[iroboto]

The very way geometry is constructed, lit and shaded could be different.

Agreed.
with Ray Tracing, sorting triangles is removed, and in essence I suspect culling is more straightforward as well (at least more straightforward)
Currently the more visually capitivating the game, the less onscreen actors there are. Culling and sorting become a larger issue. WIth RT I wonder if it can push the upper limits here and have some very complex scenes where there are just layers and layers of NPCs etc in a dense scene and RT by nature will always return the closest triangle.

 

[DieH@rd]

According to the report yesterday, does the haptic feedback in PS5 controller has anything different from HD Rumble in switch’s joycon?

Joycons have a precise rumble motors.

DS5 seemingly have force feedback for triggers [programmable force that pushes back as you press the trigger] , and precise vibration on the analogue sticks that enables it to create wide range of haptic effects. We don't know if analogue sticks also have force feedback, but Peter Rubin did mention that some effects felt "heavy" [mud] or "slick" [ice].

 

[chris1515]

https://twitter.com/x/status/1181903581353566208

 

[Jay]

The only people who care about TF is a small corner of the net.
Even then that changes, remember before RDNA cards came out, people kept saying marketing nightmare if console is a flop less than 12TF due to 1X being 6TF. Don't really hear that anymore do we.

If they feel like marketing it with a scale, then they'll just move from TF to RadeonRay's or something to show how much more better it is in comparison to old console.
Same way bits was no longer used.

 

[iroboto]

people kept saying marketing nightmare if console is a flop less than 12TF due to 1X being 6TF. Don't really hear that anymore do we.

It's a marketing nightmare which is avoided by not talking about TF
But technically speaking it's not a nightmare and I think that is what the current discussion is about.
Most people don't understand the transformation of game engines. This generation it was the move toward compute based engines. PBR etc.
Next gen it will be a move towards HRT. Mesh shaders etc.

 

[ultragpu]

The only thing that sells consoles is the games and what's on the screen. No-one cares how many flops a machine can do if it looks better than anything else. Hell, what if it's a super-processor with one single unified instruction "RENDERSCREEN" that can generate photorealistic graphic in a single floating point operation? OMG it only runs at 60 FLOPS!!!!

Shadow cascades? Shadows are one of the things raytracing solves, with no need for cascades, and none of their inaccuracies.

Some people are being really myopic about these numbers and interpreting them purely in relation to how current graphics are produced. We're talking about a new hardware generation with a potential paradigm shift and ground-up hybrid renderers using tech far more optimally. The very way geometry is constructed, lit and shaded could be different. We have zero knowledge on Sony's RT solution and enough reasons to doubt it matches RTX that we shouldn't be drawing any parallels. Numbers only tell a small part of the story. Plenty of contributors have been long enough on this board that they should know that by now and not be jumping to conclusions, although I'm aware some people prefer to form conclusions for everything instead of working with open uncertainties and not collapsing them to a final conclusion when there's enough data to create that robust conclusion.

As tot the technical debate, how much does RT hardware add to RTX? Isn't it something like 10%? What could be achieved on a large GPU with 50% of it dedicated to raytracing? Someone compare the rumoured die-sizes (300 mm²) with other GPUs and determine how much of that die would likely be 7TFs of compute and give us a die size for the RT side of things. That'll give some impression of what might be going on under the hood of this hypothetical machine.

Personally, I find the possibility quite exciting. If it is only 7TFs and a considerable amount of RT hardware, it means Sony for one are going whole hog and that might mean a true generational advance in games, not just in rendering which would be prettier, but also development if devs don't need to keep refining hacks. Though cross-platform development will be a challenge.

Well this new renderer better deliver like you said and at a respectable resolution for us 4k adopters otherwise we riot, it just ain't right if we're going backwards to 1080p or even 1440p especially in late 2020, you need a good res to appreciate nice graphics after all. I am really hoping the Pro rumor is true now.

 

[Jay]

It's a marketing nightmare which is avoided by not talking about TF
But technically speaking it's not a nightmare and I think that is what the current discussion is about.
Most people don't understand the transformation of game engines. This generation it was the move toward compute based engines. PBR etc.
Next gen it will be a move towards HRT. Mesh shaders etc.

We'll talk about TF as it is a reasonable point of comparison, just not in isolation.
I don't think it's a nightmare in marketing or technically speaking until we know more.

But I've seen it mentioned here in regards to marketing to the public, and a couple posts about it in regards to technical implications as you say.

My main point is, that people say marketing nightmare, but that changes as more details are known. Once released the figures will mean very little if the overall package is appealing.

 

[chris1515]

https://www.eurogamer.net/articles/digitalfoundry-2019-playstation-5-new-details-analysis

 

[iroboto]

We'll talk about TF as it is a reasonable point of comparison, just not in isolation.
I don't think it's a nightmare in marketing or technically speaking until we know more.

But I've seen it mentioned here in regards to marketing to the public, and a couple posts about it in regards to technical implications as you say.

Technically speaking it's never been a reasonable comparison at all. That's why we got the whole nvidia flops vs AMD flops. TF is a single number performing 1 major function of what the GPU must do. We make assumptions that the rest of the pipeline and bandwidth is beefed up equally to support the TF, but we know that's actually not the case. But we use the term anyway. Sadly.

 

[Deleted member 13524]

The only people who care about TF is a small corner of the net.

And you know, people who actually develop games and care for performance indicators the game engines that have increasingly been leaning on compute throughput for pixel shaders, geometry culling, post-processing, reconstruction techniques, and many others.

BTW, using Turing's TFLOPs as comparable to RDNA's isn't great. Turing will probably always deliver more performance per-rated-TFLOPs than RDNA because Turing does concurrent FLOPs and IOPs due to dedicated INT32 units, whereas RDNA repurposes its FP32 ALUs for integer calculations, and games use around 2 to 4 INT32 operations for every 10 FP32 operation.

So unless RDNA2 is bringing dedicated INT32 ALUs like Volta/Turing, you'll always need more TFLOPs on AMD's architecture to achieve the same FP+INT throughput.
As it stands, a 7 TFLOPs RDNA GPU will not get the same practical perfromance as a 7 TFLOPs Turing. Which is why you need a ~9TFLOP 5700XT to be competitive against a 7.5 TFLOPs RTX 2070.

 

[iroboto]

BTW, using Turing's TFLOPs as comparable to RDNA's isn't great. Turing will probably always deliver more performance per-rated-TFLOPs than RDNA because Turing does concurrent FLOPs and IOPs due to dedicated INT32 units, whereas RDNA repurposes its FP32 ALUs for integer calculations, and games use around 2 to 4 INT32 operations for every 10 FP32 operation.

TIL.
But i stand by that AMD FLOPS are Nvidia FLOPS. AMD doesn't have concurrent IOPS is the story here.

 

[iroboto]

So a user who has been vetted for inside information has revealed that PS5 was due to launch in 2019 (around this time) but they were held up by BC issues.
According to the poster, "It is my belief that this is one of (but not the only) reasons Sony delayed PlayStation 5 to holiday 2020."
https://www.resetera.com/threads/ne...ly-fan-noise-and-hot-air.129193/post-25291605

IIRC I think @DSoup made this exact prediction. 2019 delayed to 2020 for BC. I suspect being 2017, ray tracing must have been a point as well, DXR demos were revealed by end of 2018. And we know that there is no RT until RDNA2 which is releasing next year.

Depending on how accurate all of this is, this does bring into light why Sony brought in more engineers etc as we discussed.