Raster/ray traced performance

[ravenhold]

I don't understand why GPU vendors don't treat raster and ray tracing performance the same way. For example, if GPU is bound with frame number at 1080p, the same frame number would have with raytracing ON.
So it would have the exact number of ray traced/shader cores or any kind of cores dedicated to ray tracing as main ones.

My point is with lower end side of GPUs, you would get ray tracing performance at that scale as it is raster.

 

[cvaldes]

GPU manufacturers design chips to function in a variety of usage cases. They don't design them for one person's usage case. It's not your mother in the kitchen cooking you your favorite breakfast.

All of these things are a balance of features, compromises, etc. Remember that it's a finite amount of silicon wafer space and they need to consider how frequently any given type of transistor is going to be used -- raster cores, ray tracing cores, machine learning cores, etc. -- in a wide variety of real world situations, not just one random guy living in his mother's basement running benchmarks.

If running applications that benefits from the presence of ray tracing cores becomes more popular I'm guessing that these GPU manufacturers will include more of them. There's not much incentive for them to be included if most of the time they go unused. That was AMD's philosophy until RDNA2 and unsurprisingly ray tracing performance on the RDNA2 generation GPUs is inferior to NVIDIA's Ampere generation GPUs.

It's not like Jensen or Dr. Su can wave a magic wand and add 5x RT cores to a die for free.

Adding more ray tracing cores to a given GPU die means subtracting other transistors elsewhere. That might mean lesser raster performance in exchange for better ray tracing performance. Is that something you'd be interested in?

 

[ravenhold]

What about new GPU architecture doctrine where raster and ray traced cores are modulary driven, designers could add more raytracing cores without affecting raster performance.

 

[eazen]

That’s way too soon, first we need efficient code to deal with RT in games and software first then all GPUs can transition to full or more RT hardware and you can get what you want. Essentially the long time goal is full RT but we’re still far away from it.

 

[Vayra86]

What about new GPU architecture doctrine where raster and ray traced cores are modulary driven, designers could add more raytracing cores without affecting raster performance.

They're radically different kinds of calculations.

If you look at the papers on Ampere and Turing you will see that INT (integer) functions were added and the new cores carry those functions, because they're built to handle them faster. Efficiency in GPUs is often obtained by reducing the functionality of cores, making them more single-purpose. Example of that is how Nvidia don't offer high precision floating point on Geforce, and did on Titan, but eventually killed that too.

But in the history of GPU development there is always a shift back and forth, as new functionality is added, and refined, and at some point becomes a known quantity. There's always a phase where the resources a GPU gets are not perfectly aligned with what games want. That's why there are differences in performance between engines/GPU families/games.

Pascal to Turing: less performance per shader, lower perf per clock, too, AND lower clocks, but the functionality was expanded. Ampere iterated on that with a further refinement of shader count, and the sacrifice was made in TDP to remain competitive in raster performance.

Either way, yes, I do agree the ideal situation is one where you don't waste die space on cores that are going to be idle at any point in time. We have yet to see if that is feasible.

 

[eazen]

Pascal to Turing: less performance per shader, lower perf per clock, too, AND lower clocks, but the functionality was expanded. Ampere iterated on that with a further refinement of shader count, and the sacrifice was made in TDP to remain competitive in raster performance.

I’m curious how you would come to conclusion that Nvidia made the shaders worse with a new architecture? Turing has higher IPC than Pascal in every way, and the clocks are same more or less. The only GPU with straight lower clocks is the 2080 Ti cause it was a huge GPU (still the biggest gaming GPU ever) and had to make due with a “low” tdp of just 280W compared to its size.

 

[ravenhold]

Is there a future of GPU architecture where we ditch raster alltogether and use only raytracing cores? If not RTX5000 then RTX6000 series maybe?

If that happens, is it required to rewrite complete shader system/Directx/Vulkan?

 

[ratirt]

I’m curious how you would come to conclusion that Nvidia made the shaders worse with a new architecture? Turing has higher IPC than Pascal in every way, and the clocks are same more or less. The only GPU with straight lower clocks is the 2080 Ti cause it was a huge GPU (still the biggest gaming GPU ever) and had to make due with a “low” tdp of just 280W compared to its size.

If you look at the 2080 ti and 3060 ti. The 3060 has higher clocks and more cores and yet is around 15% slower. Obviously the 2080 Ti is bigger since it's been made on a 12nm node and the 3060 ti 8nm so it is smaller. By the cores and clocks it should have been faster but it isn't. Maybe the memory plays a role here but I doubt it. So there are some sort of limitations implemented I suppose.
Same goes for pascal?

 

[Vayra86]

I’m curious how you would come to conclusion that Nvidia made the shaders worse with a new architecture? Turing has higher IPC than Pascal in every way, and the clocks are same more or less. The only GPU with straight lower clocks is the 2080 Ti cause it was a huge GPU (still the biggest gaming GPU ever) and had to make due with a “low” tdp of just 280W compared to its size.

You are correct, I mixed things up. Pascal clocks higher, overall. A good 100 mhz boost advantage is common, often more