Article World of Tanks Implements Ray Traced Shadows for ANY DX11 GPU

[Stuka87]

Meant to post this yesterday, but I forgot. The video link below goes through how the implemented it, and its an interesting watch. For anybody wondering why its only shadows, its because each tank has in excess of 50k polygons, and you can have up to 50 tanks on screen at any given time. Currently no hardware exists that can handle that kind of load with RT reflections/GI. Wargaming has always tried to make their engine run on even low end hardware, and not require special hardware.

However, with how they implemented the RT Shadows, with help from Intel, it can run on any GPU. I play the game a lot, so I downloaded their test client, with the game settings at Ultra, and RT Shadows at HIGH, my system with a 4.5GHz 4690K and an RX480 4GB ran it really quite well at 1080P. It did drop below 60fps a few times. As a note, this does not use any specialized RT hardware, even if the GPU has it. This uses a combination of CPU and GPU to handle the calculations, and they use only 1 ray per pixel, with a custom dithering filter that they came up with which is shown in the video.

Details: https://worldoftanks.com/en/news/general-news/ray-tracing/

Developer Video:

 

[Stuka87]

I stand by my statements, its you guys who don't understand how percentages work during benchmarking or when comparing fps .. It's STILL is 70% difference! Go read some books before you babble about trivia.

Here is how people do compare fps numbers:

SMH

Ok, let me work this out for you. Our baseline is 203. 203 is 100%. If performance dropped by 50%, we would end up with 101.5fps. This is our baseline cut in half. The 119fps with RT is GREATER THAN 101.5. This means 119 cannot possibly be 70% slower than 203. A 70% drop from 203 is 60.9.

To get the correct number we take 119 divided by 203. This gives us 58.62. We then take 100 and subtract 58.62, which gives us 41.38. Which is how we arrive at how much our frame rate decreased.

 

[DXDiag]

The difference between TR/Metro compared to WoT is the complexity of the objects. In WoT you can have up to 60 tanks on screen at one time, each having 50K+ polygons. Thats significantly more than either of the other games have on screen.

GI and AO are much more complex than mere vehicle shadows. Metro scenes are filled to the prim with tessellated complex objects and characters. Tomb Raider have fully detailed characters and environments with multiple lights at any given moment, this gives each scene dozens of RT shadows to work with. Control has Shadows + GI + Reflections all at the same time.

In contrast we have only one light source in World Of Tanks, environment and other objects are much simpler, and lighting is less complex. Also the benchmark has only a handful of tanks on screen at any given time, yet despite this the drop in fps is substantial. You lose close to a 85fps @1080p for less than a dozen tank. The more tanks you have on screen the higher the drop will be. Expect 30fps or less if you have more than 20 tanks on screen. Also the implementation only works on a max of 30 Tanks.

In essence this puts everything into context, software RT is never going to be equal to hardware RT. That's why consoles had to have hardware RT.

 

[maddie]

SMH

Ok, let me work this out for you. Our baseline is 203. 203 is 100%. If performance dropped by 50%, we would end up with 101.5fps. This is our baseline cut in half. The 119fps with RT is GREATER THAN 101.5. This means 119 cannot possibly be 70% slower than 203. A 70% drop from 203 is 60.9.

To get the correct number we take 119 divided by 203. This gives us 58.62. We then take 100 and subtract 58.62, which gives us 41.38. Which is how we arrive at how much our frame rate decreased.

Or we could just say (203-119)/203.

I've always been pretty good at mathematics and have sometimes had a hard time when teaching, to understand why something so obvious could be so difficult to grasp. I now have a slightly better appreciation, based on those exchanges.

 

[DisarmedDespot]

I'm disinclined to draw any sort of performance metrics from World Of Tanks. Unless they've change it, their engine is ancient and they spent years trying to implement Havok physics and never actually managed it.

 

[Paul98]

I stand by my statements, its you guys who don't understand how percentages work during benchmarking or when comparing fps .. It's STILL is 70% difference! Go read some books before you babble about trivia.

Rather than being defensive, how about go back and understand what you said incorrectly so you won't do it again. Just looking at that your statement and the numbers presented it should be obvious that it's incorrect.

Saying something is 70% faster is not the same as saying something dropped by 70%. Like a 50% drop is the same as 100% faster, like looking at something half the speed vs double speed.

Just look at the numbers 119 and 203, clearly not a 50% drop, as if it was 50% it would be 101.5( or just a little above 100 )

So taking 119/203 and getting 59%, that's not telling you the percentage for the drop ether, look at something like 1/203 would be a .5% drop in FPS, and a 203/203 would be a 100% drop in fps( clearly nonsense ). it actually is 41%, as it's 1-(that 59%) or in the examples I gave 1-.005 or 1-1

Words do matter not just the number, otherwise it is incorrect and gives a totally different picture of the situation

 

[Stuka87]

I'm disinclined to draw any sort of performance metrics from World Of Tanks. Unless they've change it, their engine is ancient and they spent years trying to implement Havok physics and never actually managed it.

The engine was completely replaced last year. It wasn't just an upgrade or something, its literally an entirely different engine.

 

[sontin]

In essence this puts everything into context, software RT is never going to be equal to hardware RT. That's why consoles had to have hardware RT.

As long as Raytracing needs certain workloads "hardware RT" will always be faster than using compute units alone. We have been there 20 years ago. T&L/Vertex shader has moved geometry processing from the CPU to the GPU.

World of Tanks is just a marketing stunt from Intel. The quality is mediocore and the performance lost is on par with Raytracing in Tomb Raider with Ultra settings...

 

[BFG10K]

T&L/Vertex shader has moved geometry processing from the CPU to the GPU.

True, but then fixed hardwired T&L was offloaded onto more generalized programmable vertex shaders. Then dedicated vertex shaders became generalized into unified shaders. So if you run basic T&L today, it's actually emulated on the GPU.