CYBERPUNK 2077 [PC Specific Patches and Settings]

Evaluating SER in Cyberpunk 2077 Path Tracing.

With SER, runtime for DispatchRays calls decreased by 24%, so SER is a large overall win for performance. NVIDIA’s efforts should be applauded here, as SER is able to increase the number of active lanes in a wavefront by 46%.

 
This brings the point home that raytracing and pathtracing workloads are not worse for the GPU than rasterizing + additional software stack. In Cyberpunk i can get 500W with Overdrive, in Layers of Fear my 4090 is stuck at around 370W. So bigger L2 cache, better RT Cores and SER helps alot to better utilize the GPU.
 
This brings the point home that raytracing and pathtracing workloads are not worse for the GPU than rasterizing + additional software stack. In Cyberpunk i can get 500W with Overdrive, in Layers of Fear my 4090 is stuck at around 370W. So bigger L2 cache, better RT Cores and SER helps alot to better utilize the GPU.

Comparing usage in different games is exactly the way to come up with a false conclusion.

Regards,
SB
 
PCGH tested the path tracing in Cyberpunk using Arc GPUs, they tested them at upscaled 1080p. Arc seems behind Turing in this extreme form of ray tracing, as the RTX 2060 is 30% faster than Arc 770. In fact, Turing is even faster than all RDNA2/RDNA3 GPUs, the 2080Ti is 4% faster than 7900XTX and 60% faster than 6900XT. The picture remains the same when tested with upscaled 1440p.

 
This brings the point home that raytracing and pathtracing workloads are not worse for the GPU than rasterizing + additional software stack. In Cyberpunk i can get 500W with Overdrive, in Layers of Fear my 4090 is stuck at around 370W. So bigger L2 cache, better RT Cores and SER helps alot to better utilize the GPU.
I think it does mean it worse, but we can build the hardware to help resolve those problems. We have 30 years of rasterization evolution in hardware and software, people were expecting too much from RT in just its infancy. We are definitely on our way to a pure RT only future
 
I think Raytracing performs much better than expected. You can play Battlefield 5 in 4K with 60FPS+ (more like 120FPS...) on a 4080/4090 card. How long has it take to play Crysis in 4K with 60FPS?
 
I think Raytracing performs much better than expected. You can play Battlefield 5 in 4K with 60FPS+ (more like 120FPS...) on a 4080/4090 card. How long has it take to play Crysis in 4K with 60FPS?
I think that’s Crysis fault right? They thought GPUs would trend one way and it went another ?
 
I think Raytracing performs much better than expected. You can play Battlefield 5 in 4K with 60FPS+ (more like 120FPS...) on a 4080/4090 card. How long has it take to play Crysis in 4K with 60FPS?
You still can't play 2007 Crysis at a locked 60fps all the way through on any PC.

But from a GPU point of view Crysis hasn't been GPU limited for 10+ years.
 
I think that’s Crysis fault right? They thought GPUs would trend one way and it went another ?
The expectation back then was CPU's would evolve to have one or two 8Ghz+ cores/threads and not a pure multi-core path we ultimately ended up going/having.

So it's not really the games fault, it's just the way the industry changed its mind and the game suffered because of it.
 
Last edited:
I think that’s Crysis fault right? They thought GPUs would trend one way and it went another ?
We can use Cyberpunk as another example:
A 2080TI got ~60FPS with 1440p and rasterizing: https://www.computerbase.de/2020-12/cyberpunk-2077-benchmark-test/2/

With a 4090 you get between 30FPS and 60FPS (outside more, inside less) with the Override mode. I think this a huge archivement for only four years of GPU development. So i dont think that Raytracing or Pathtracing workload isnt suited for GPUs. I even think it is much better suited than rasterizing combined with software tricks to get the same level of quality. My 4090 runs at 500W in Cyberpunk Override, so there is no stalling of the GPU.
 
Last edited:
We can use Cyberpunk as another example:
A 2080TI got ~60FPS with 1440p and rasterizing: https://www.computerbase.de/2020-12/cyberpunk-2077-benchmark-test/2/

With a 4090 you get between 30FPS and 60FPS (outside more, inside less) with the Override mode. I think this a huge archivement for only four years of GPU development. So i dont think that Raytracing or Pathtracing workload isnt suited for GPUs. I even think it is much better suited than rasterizing combined with software tricks to get the same level of quality. My 4090 runs at 500W in Cyberpunk Override, so there is no stalling of the GPU.
The software hasn’t even really had a chance to catch up or mature yet. There is more performance gains to be had on the existing architectures
 
So i dont think that Raytracing or Pathtracing workload isnt suited for GPUs. I even think it is much better suited than rasterizing combined with software tricks to get the same level of quality.

I think that’s true but the bar for acceptable quality is currently determined by raster and not RT. It would be interesting to see VSMs go head to head with RT shadows in a complex scene with lots of geometry and a few shadow casting lights. My money is on RT being faster in any non trivial scene.

For basic rendering GPUs are definitely better suited to raster given the amount of fixed function hardware available and less branchy vertex shading vs RT. Higher power consumption doesn’t mean your card is working more efficiently. It just means it’s working harder.
 
Higher power consumption doesn’t mean your card is working more efficiently. It just means it’s working harder.
Then this begs the question, why can't the raster path make GPUs work harder? It seems the raster path has been plagued by one too many limitations, from CPU code limitations to GPU code limitations .. etc. It's not really extracting performance well enough out of the hardware.
 
Then this begs the question, why can't the raster path make GPUs work harder? It seems the raster path has been plagued by one too many limitations, from CPU code limitations to GPU code limitations .. etc. It's not really extracting performance well enough out of the hardware.

It’s simple really. Raster typically uses more fixed function hardware than RT. It’s almost impossible to keep all of the fixed function and compute hardware firing on all cylinders at all times. Nanite is also rasterization but will likely push the hardware very hard because it’s moving a ton of work to the most power hungry parts of the chip and bypassing a lot of fixed function stuff.

I’m not sure what you mean by extracting performance. Power consumption isn’t a measure of performance. If power consumption is lower in a given workload because the fixed function stuff is working efficiently that’s a performance win.

The other obvious reason is that RT workloads are just more taxing due to ray divergence. Shading and memory loads are less efficient and therefore the hardware needs to work harder. This isn’t a good thing.
 
PCGH tested the path tracing in Cyberpunk using Arc GPUs, they tested them at upscaled 1080p. Arc seems behind Turing in this extreme form of ray tracing, as the RTX 2060 is 30% faster than Arc 770. In fact, Turing is even faster than all RDNA2/RDNA3 GPUs, the 2080Ti is 4% faster than 7900XTX and 60% faster than 6900XT. The picture remains the same when tested with upscaled 1440p.


The intel results are a good reference now, they do well in RT workloads outside of this and Portal RTX where AMD also falter badly. I wonder how well intel cards work with the pathtraced updates to Serious Sam/Doom where 6900XT was close to 3070 performance instead of being half of it as here.

In terms of utilization, only Nvidia's current series actually runs smoothly, which can be seen from the filling of the power limit and the clock rates, among other things. While all Radeon GPUs are boosting extremely high, which indicates massive "air bubbles" in the pipeline, Turing and Ampere GPUs are at least being loaded above average.
 
Back
Top