Support NeoGAF

[In Cold Blood]

The link you have shared demonstrates nothing which contradicts my original post or information, it's just a deep dive on how Primitive Shaders function and is based on information which is now outdated.

Both Primitive and Mesh Shaders are set up in the GPU's geometry engine and command processors, the hardware changes made to these components was done in AMD's Vega architecture and up until and including RDNA 2 there have been absolutely zero changes made to either these components according to AMD documentation.

Here's a link to AMD's Vega white paper and on page 6 it highlights the "NGG fast path" (Next Generation Geometry). Both RDNA 1 and RDNA 2 documentation show there was zero changes made to the hardware to accommodate Mesh Shaders.

In fact all AMD RDNA 2 GPU's are converting Mesh Shaders into Primitive Shaders in code, this includes the Series X/S as well as the discrete GPU's.

It has nothing to do with what I "believe" lol it's literally on paper if you know where to look.

LeviathanGamer2 has done an excellent job on covering this topic.

Not quite.
At the definition level of the problem they seek to solve, we could say that they seek to solve the same problem, but these are two different solutions.

First of all, the Primitive Shader retains the use of the teselation unit for the amplification of geometry.

While the 3D pipeline with the Mesh Shader completely eliminates that part.

However, one might think that if Primitive Shader= Mesh Shader then Amplification Shader = Surface Shader. Right? The AMD patent of the Primitive Shaders is clear about this.

The Shader Surface is active when the teselation is. When the teselation is active, the shaer surface implements the functionality of the vertex and hull shader stages. The teselation stage is still implemented in fixed-function hardware. The Shader Surface is disable when the teselation is deactivated. The Shader Surface is implemented both both partial and totally as a shader program.
When the teselation is active, the Primitive Shader implements the functionality of the Domain Shaer and Geometry Shader Stage, but if the latter is active when the teselation is active, the Primitive Shader implements the functions of the Vertex Shader stage.

On the other hand, if we look at Microsoft documentation on the Shaders Ampliplification, we are told:

While the Mesh Shaders are a rather flexible tool, they do not allow all teselation scenarios and are not always the most efficient way to implement selective selection per instance. To do so, we have the Shader Ampliplification. What they do is simple: send groups of subprocesses from Mesh Shaders. Each Mesh Shader has access to the data from the main Shader Amplification and does not return anything. The Shader Amplification is optional and also has access to group shared memory, making it a powerful tool to allow Mesh Shader to replace any current channelling scenario.

In short, the Shader Amplification allows to indicate the volume of geometry with which we want the Mesh Shaders to work. It is therefore thanks to them that the teselation unit is not necessary in this approach, being this the main change between the Primitive Shaders' approach and the Mesh Shaders.

So in short Mesh Shaders and Primitive shaders are doing the same job, just that the Mesh Shader pipeline is less than the Primitive pipeline.

 

[supernova8]

Doesn’t state it’s PlayStation or Xbox related, it could be for anything, maybe the next Nintendo console since that would be coming much sooner.

Sshhh they need the clicks

 

[Soosa]

I know that Nvidia doesnt make custom SoCs, but I wanted them to make these for consoles so freakind bad

DLSS and way better RT performance are so important

Why people hope for this madness? While NVIDIA makes fast PC gpus, it doesnt mean that they could make FAST + CHEAP + power efficent console SoC, and they dont have x86 license anyway.

This is as weird as looking at some million dollar super car and thinking that because it is so fast, why wont they design an engine for 15 000€ car?

AMD is clearly the best manufacturer for x86 SoC, consoles cant have some 800$ SoC that eats 1000 W of power. And NVIDIA isnt really known for cheap parts, which is essential for consoles.
Also they screwed up PS3 with their GPU so they are bad choise if anything.

DLSS is just one technique, consoles have their own ways to upscale stuff and RT is what it is, even with imaginary NVIDIA SoC we would probably not have much better RT for this gen, and by next gen AMD will also have better RT implementations.

Nintendo switch has a custom tegra soc (according to Nintendo official website)

At least the original switch launched basically with Stock tegra X1, but with lower clock speeds. Nintendo fanboys were sure before the launch, that they will get super custom SoC, but nope. Maybe they call it custom because some tiny changes, it could be basically "custom" if they cut out some features.

Nvidia makes Tegra SOC and has done so for many years. Used in the switch, shield etc.

They talked about CUSTOM SoC, not about stock stuff. Custom usually means stuff that is tailored to certain system and have features that other devices dont have.

Could NVIDIA make custom tegra with much better features and performance than X1? yes, but whom is willing to pay for it is another question. Nintendo surely didnt want to pay, as they just took the X1 and used it basically as stock chip.


Dreams are dreams, but AMD is currently the best and basically only manufacturer that can sell x86 SoC that have good performance/price factor.

Intel could do x86 + gpu, but it would probably cost too much, Nvidia could make ARM based SoC but it would also probably cost too much + it would make backwards compability worse, and probably performance would not be that much better than similarry priced AMD SoC, it could also be worse at same price point.

People act like AMD would be some bad choise, when it is basically the second best GPU manufacturer on the market (with small margin) and best or shared 1st place as CPU manufacturer

 

[TheRedRiders]

Not quite.
At the definition level of the problem they seek to solve, we could say that they seek to solve the same problem, but these are two different solutions.

First of all, the Primitive Shader retains the use of the teselation unit for the amplification of geometry.

While the 3D pipeline with the Mesh Shader completely eliminates that part.

However, one might think that if Primitive Shader= Mesh Shader then Amplification Shader = Surface Shader. Right? The AMD patent of the Primitive Shaders is clear about this.


On the other hand, if we look at Microsoft documentation on the Shaders Ampliplification, we are told:


In short, the Shader Amplification allows to indicate the volume of geometry with which we want the Mesh Shaders to work. It is therefore thanks to them that the teselation unit is not necessary in this approach, being this the main change between the Primitive Shaders' approach and the Mesh Shaders.

So in short Mesh Shaders and Primitive shaders are doing the same job, just that the Mesh Shader pipeline is less than the Primitive pipeline.

Again none of this proves your argument that the hardware for Mesh Shaders and Primitive Shaders is different.

 

[ToTTenTranz]

DLSS and way better RT performance are so important

Perhaps the fact that neither Sony nor Microsoft demanded AMD to spend more die area on those is a clear indication that they aren't as important as other features at this stage.
The difference between DLSS and shader-based temporal upsampling is almost negligible in motion on a 4K TV at normal viewing distances, and better RT acceleration would have taken millions/billions of transistors that were otherwise spent on more ALUs, caches and wider buses.

People seem to forget that a RTX3060 (just the GPU) has 13.25 billion transistors and consumes 170W, while the Series X SoC (GPU + 8-core CPU + infinity fabric + I/O + sound acceleration) has 15.3 billion transistors and the entire console with SoC+SSD+WiFi+BT+disc+etc. pulls 170W at the wall (210W at ~80% efficiency PSU). Yet the Series X probably performs better than a PC with the RTX3060 in most games.
DLSS and better RT wouldn't have come for free had Nvidia designed the console GPUs. Everything is a trade-off.

 

[Hugare]

Perhaps the fact that neither Sony nor Microsoft demanded AMD to spend more die area on those is a clear indication that they aren't as important as other features at this stage.

Nonsense

DLSS is so good, that AMD is trying really hard to reach it as fast as possible with FSR (with no success). They are trying, but they just dont know how.
It's by far better than any other temporal upscaling sollution that I've ever seen, while more than doubling performance in most cases. We would be playing console games at least at 1800p right now instead of 1440p, most of the time.

And about RT, let's just agree to disagree.
'Cause I preffer playing Cyberpunk with everything maxed on my laptop equiped with a rtx 2060 at 1080/30 instead of playing it on my PS5 at "dynamic 4K"/60.
For some games, RT makes more impact than performance or resolution.
Specially when it has RT GI

 

[In Cold Blood]

Again none of this proves your argument that the hardware for Mesh Shaders and Primitive Shaders is different.

Um..the fact that mesh shaders have done away with the tessellation unit and Primitive Shaders still have it absolutely shows that there is a hardware difference.
So Primitive shaders have a tessellation unit.
Mesh shaders don't have a tessellation unit.
If you are arguing that there is no difference in the hardware at this point there must be a reason that has nothing to do with actual facts.

 

[TheRedRiders]

Um..the fact that mesh shaders have done away with the tessellation unit and Primitive Shaders still have it absolutely shows that there is a hardware difference.
So Primitive shaders have a tessellation unit.
Mesh shaders don't have a tessellation unit.
If you are arguing that there is no difference in the hardware at this point there must be a reason that has nothing to do with actual facts.

Incorrect, as I've pointed out already, AMD have highlighted zero changes to their GPU architecture to accommodate Mesh Shaders, AMD's ISA on RDNA 2 makes no mention of Mesh Shaders either.

The whole point of Mesh/Primitive Shaders is to send compute based workloads to the GPU without the bottlenecks of a compute shader. All the Primitive Shader stages are fully programmable according to AMD documentation and patents.

As I mentioned before, LeviathanGamer2 (who specialises in gaming hardware architecture at a degree level) has covered this already.

 

[Tams]

I know that Nvidia doesnt make custom SoCs, but I wanted them to make these for consoles so freakind bad

DLSS and way better RT performance are so important

You tell them not to be such utter cunts to work with, and you might have a shot.