Support NeoGAF

[64gigabyteram]

Spitballing an idea in a low effort thread. not like anybody from that company will see this, but why not. Worst case scenario i'd get some people more educated than me in the replies telling me why it's a dumb idea

Even in the modern age with all the progress made, Ray tracing and especially path tracing isn't panning out so well for most modern games in terms of performance.
So I've wondered why Nvidia doesn't release some kind of card dedicated to pure raytracing/path tracing capabilities. no traditional GPU compute, just pure RT number crunching. It'd probably be eons faster than even the best 5090 when it comes to RT/PT processing and people with worse off GPUs can add this in and get infinitely faster RT/PT performance. You wouldn't even need DLSS to run something like HL2 RTX at full speed.

It's an Idea i came up with after talking with my friend about how he bought a Nvidia Tesla card alongside his 9070 XT for PhysX performance.

 

[salva]

"Fortunately we have a product for people who aren't able to do path tracing, it's called a 5090"

 

[Crayon]

You forgot to mention how much your 5090-killing pt card will cost me. Then I can tell you how good of an idea it is.

 

[Magic Carpet]

How many 8pin power plugs will this need?

 

[Bitstream]

Man, you made a thread to ask why nvidia hasn't released a more powerful card than the 5090 yet?

 

[64gigabyteram]

You forgot to mention how much your 5090-killing pt card will cost me. Then I can tell you how good of an idea it is.

if it were a not greedy company who had the GPU monopoly realistically it'd be 300-400

unfortunately this is Nvidia we are talking about

 

[RodriGames40]

It's feasible, but at the same time you'd probably make the engineers wonder why don't they just drop the rasterization cores in the actual cards instead... They did drop Physx at the hardware level recently.

 

[crazyriddler]

Why would they do that? One of the advantages of their cards is that they offer better RT/PT performance than the competition. This allows them to sell their graphics cards at a higher price than similar ones from AMD. Releasing a dedicated card for RT would only benefit the competition.

 

[fersnake]

so back to Ageia days? lol

 

[Hoddi]

Getting data from a traditional GPU to a secondary RT card would need go over PCIe. I'd imagine that would be a major bottleneck.

It would make more sense in this case to have a dedicated accelerator chip on the same physical card as the primary GPU as they could then share a unified memory bus. But that just leaves you with a more expensive 5090 with two chips instead of one and higher latency between them.

 

[Danny Dudekisser]

I, too, think about why I'm not giving Nvidia more money for their overpriced shit.

 

[KuraiShidosha]

I'm torn. On the one hand, I love ray tracing and am ready to worship our path tracing overlords. (I already own a custom water cooled 5090 and previously had a 4090 FE that got stolen by FedEx when RMAing it.) On the other hand, I'm a stickler for backwards compatibility and old game preservation. The idea of a 5090 class card that totally drops rasterization die space in favor of path tracing accelerators is exciting but there just isn't enough content to justify it yet. You can only get so far playing Cyberpunk, Quake 2, and Half Life 2.

 

[pacman4000]

Getting data from a traditional GPU to a secondary RT card would need go over PCIe. I'd imagine that would be a major bottleneck.

It would make more sense in this case to have a dedicated accelerator chip on the same physical card as the primary GPU as they could then share a unified memory bus. But that just leaves you with a more expensive 5090 with two chips instead of one and higher latency between them.

We are done here.

Not all ideas are great.

 

[Gaiff]

Dedicated physx cards didn't work all that well. Transferring data between two PCIe buses has its fair share of problems.

 

[Pimpbaa]

I'd sell about as well as those physics acceleration cards.

 

[diffusionx]

The issue comes down to two things, one is doing the math necessary to do the path tracing and then to transport that to the other GPU that could use it. The first is definitely doable, the second, I don't think it is feasible over the current design of the PC.

There is a reason why Nvidia GPUs are enormous with tons of transistors, it's because it does all this on the chip to do it as fast as possible.

 

[Buggy Loop]

Would perform like shit

Path tracing is basically glued to shader pipeline and share low level cache memories and the shader pipeline is actually doing a bit part of the path tracing, lighting up pixels of the shaders after a BVH. The BVH part is done on dedicated RT cores but the pipeline is not idle in the meantime,

The "compute" part of BVH is already so fast because of the ASIC nature of RT blocks on Nvidia side, the trigger of a light hit to pipeline and order that color/brightness of a pixel is still very much typical rendering pipeline.

And now more than ever they are going with.m neural solutions for path tracing so on top of that you need ML tensor cores nearby

Imagine having another PCB with dedicated hardware that has to go through some slow ass communication bus (comparatively to silicon cache, it'll always be slower) for tasks that require the smallest latencies and idling possible in all of modern graphic rendering because it is already costly to enable?

Yea it would go to shit. Imagine the worse of SLI and multiply tenfold the problems.

 

[Radical_3d]

You lost me at "nVidia".

 

[Imtjnotu]

$2499 GPU
$2498.99 Path tracing card

 

[Soodanim]

Spitballing an idea in a low effort thread. not like anybody from that company will see this, but why not. Worst case scenario i'd get some people more educated than me in the replies telling me why it's a dumb idea

Even in the modern age with all the progress made, Ray tracing and especially path tracing isn't panning out so well for most modern games in terms of performance.
So I've wondered why Nvidia doesn't release some kind of card dedicated to pure raytracing/path tracing capabilities. no traditional GPU compute, just pure RT number crunching. It'd probably be eons faster than even the best 5090 when it comes to RT/PT processing and people with worse off GPUs can add this in and get infinitely faster RT/PT performance. You wouldn't even need DLSS to run something like HL2 RTX at full speed.

It's an Idea i came up with after talking with my friend about how he bought a Nvidia Tesla card alongside his 9070 XT for PhysX performance.

Forget the topic, tell me more about this AMD+nV solution your mate did. Did mixing GPUs not create a black hole?

 

[Ar¢tos]

Isn't nvidia much more interested in making cards for AI purposes than gaming nowadays?

 

[Hudo]

so back to Ageia days? lol

This is what I was immediately reminded off.