Investigation Shows That NVIDIA GPUs Implement Tile Based Rasterization for Greater Efficiency

[Master Disaster]

An investigation into Nvidia GPUs has shown that during the transition from Kepler to Maxwell Nvidia made some huge changes in the way the card handles Rasterization which is now closer to low power SoCs. 

Quote

As someone who analyzes GPUs for a living, one of the more vexing things in my life has been NVIDIA’s Maxwell architecture. The company’s 28nm refresh offered a huge performance-per-watt increase for only a modest die size increase, essentially allowing NVIDIA to offer a full generation’s performance improvement without a corresponding manufacturing improvement. We’ve had architectural updates on the same node before, but never anything quite like Maxwell.

 

The vexing aspect to me has been that while NVIDIA shared some details about how they improved Maxwell’s efficiency over Kepler, they have never disclosed all of the major improvements under the hood. We know, for example, that Maxwell implemented a significantly altered SM structure that was easier to reach peak utilization on, and thanks to its partitioning wasted much less power on interconnects. We also know that NVIDIA significantly increased the L2 cache size and did a number of low-level (transistor level) optimizations to the design. But NVIDIA has also held back information – the technical advantages that are their secret sauce – so I’ve never had a complete picture of how Maxwell compares to Kepler.

 

For a while now, a number of people have suspected that one of the ingredients of that secret sauce was that NVIDIA had applied some mobile power efficiency technologies to Maxwell. It was, after all, their original mobile-first GPU architecture, and now we have some data to back that up. Friend of AnandTech and all around tech guru David Kanter of Real World Tech has gone digging through Maxwell/Pascal, and in an article & video published this morning, he outlines how he has uncovered very convincing evidence that NVIDIA implemented a tile based rendering system with Maxwell.

 

In short, by playing around with some DirectX code specifically designed to look at triangle rasterization, he has come up with some solid evidence that NVIDIA’s handling of tringles has significantly changed since Kepler, and that their current method of triangle handling is consistent with a tile based renderer.

Tile based rasterization is more common in PowerVR & ARM based chips

Quote

Tile based rendering is something we’ve seen for some time in the mobile space, with both Imagination PowerVR and ARM Mali implementing it. The significance of tiling is that by splitting a scene up into tiles, tiles can be rasterized piece by piece by the GPU almost entirely on die, as opposed to the more memory (and power) intensive process of rasterizing the entire frame at once via immediate mode rendering. The trade-off with tiling, and why it’s a bit surprising to see it here, is that the PC legacy is immediate mode rendering, and this is still how most applications expect PC GPUs to work. So to implement tile based rasterization on Maxwell means that NVIDIA has found a practical means to overcome the drawbacks of the method and the potential compatibility issues.

 

In any case, Real Word Tech’s article goes into greater detail about what’s going on, so I won’t spoil it further. But with this information in hand, we now have a more complete picture of how Maxwell (and Pascal) work, and consequently how NVIDIA was able to improve over Kepler by so much. Finally, at this point in time Real World Tech believes that NVIDIA is the only PC GPU manufacturer to use tile based rasterization, which also helps to explain some of NVIDIA’s current advantages over Intel’s and AMD’s GPU architectures, and gives us an idea of what we may see them do in the future.

http://www.anandtech.com/show/10536/nvidia-maxwell-tile-rasterization-analysis

http://www.realworldtech.com/tile-based-rasterization-nvidia-gpus/

 

Interesting stuff, I'm not 100% sure what it all means though. 

 

Thoughts? 

[Hunter259]

And that might explain why Kepler and Fermi based cards got left in the dust so quickly. 

[PlayStation 2]

I don't really see this as a negative thing moreso clever implementation of something that already exists.

[Master Disaster]

3 minutes ago, Dan Castellaneta said:

I don't really see this as a negative thing moreso clever implementation of something that already exists.

From what I gather it's to do with how the GPU renders the scene. Instead of rendering the entire frame it splits it into small tiles then renders each tile individually. Right? 

 

If so then I'd agree it's not a bad thing as long as it is more efficient but I'd question that fact, rendering 36 small frames compared to just 1 big one must equal the same amount of overall work done so is the efficiency gains due to multithreading and simultaneous processing? 

[Dabombinable]

7 minutes ago, Dan Castellaneta said:

I don't really see this as a negative thing moreso clever implementation of something that already exists.

And considering that Nvidia would have had experience with it in the form of their Tegra, it really shouldn't surprise anyone.

[PlayStation 2]

Just now, Master Disaster said:

From what I gather it's to do with how the GPU renders the scene. Instead of rendering the entire frame it splits it into small tiles then renders each tile individually. Right? 

 

If so then I'd agree it's not a bad thing as long as it is more efficient but I'd question that fact, rendering 36 small frames compared to just 1 big one must equal the same amount of overall work done so is the efficiency gains due to multithreading and simultaneous processing? 

It's probably benefiting from multithreading in a way.

[Hunter259]

1 minute ago, Master Disaster said:

From what I gather it's to do with how the GPU renders the scene. Instead of rendering the entire frame it splits it into small tiles then renders each tile individually. Right? 

 

If so then I'd agree it's not a bad thing as long as it is more efficient but I'd question that fact, rendering 36 small frames compared to just 1 big one must equal the same amount of overall work done so is the efficiency gains due to multithreading and simultaneous processing? 

Rasterizing is just a fancy way of saying rendering really. It renders all tiles and then displays them in 1 frame. So yeah it really is just multi-threading in a way.

[DocSwag]

What I want to see is some testing by people with Kepler, gcn 1.0,1.1, no 1.2, and Polaris GPUs do some testing to see how they behave in this test. If anyone here has one of those I'm curious to see your test results. Thanks!

[DocSwag]

Just now, Hunter259 said:

Rasterizing is just a fancy way of saying rendering really. It renders all tiles and then displays them in 1 frame. So yeah it really is just multi-threading in a way.

So that explains why it increases ROP performance... But then I'm wondering now, he stated it decreases memory utilization.... How does that work?

[Master Disaster]

@patrickjp93Care to chime in and explain how tile rasterization is more efficient than normal? I'm struggling to understand because splitting the frame up still equals the same work done plus then it has to stitch everything together afterwards. 

[Xorbot]

I cannot be certain about the specifics of this type of rendering by NVIDIA.  However, tile-based rendering can lead to other optimizations such as hidden surface removal, the broad term to define this.

 

What happens is if a scene is properly rendered with opaque geometry first, then alpha-blended geometry sorted back-to-front, entire objects can be discarded completely.  Fragment processing would not proceed after this as well, as per usual rasterization.

 

So if a tile is detected as completely opaque, then triangle that lies completely behind the Z/depth values in the tile would be discarded and never even processed further.

[jaggysnake57]

errrrrr any one got a dumbed down version

 

[Master Disaster]

Just now, Xorbot said:

I cannot be certain about the specifics of this type of rendering by NVIDIA.  However, tile-based rendering can lead to other optimizations such as hidden surface removal, the broad term to define this.

 

What happens is if a scene is properly rendered with opaque geometry first, then alpha-blended geometry sorted back-to-front, entire objects can be discarded completely.  Fragment processing would not proceed after this as well, as per usual rasterization.

 

So if a tile is detected as completely opaque, then triangle that lies completely behind the Z/depth values in the tile would be discarded and never even processed further.

OK, so it can ignore some stuff that isn't required to reduce overall workload. Interesting and actually makes a tonne of sense once you think about it

[Hunter259]

Just now, Master Disaster said:

@patrickjp93Care to chime in and explain how tile rasterization is more efficient than normal? I'm struggling to understand because splitting the frame up still equals the same work done plus then it has to stitch everything together afterwards. 

Think of how a single threaded application and a well written multi-threaded application behaves. Multi-threaded is much faster right? This is very similar to this.

Just now, DocSwag said:

So that explains why it increases ROP performance... But then I'm wondering now, he stated it decreases memory utilization.... How does that work?

It only is working on one tiny bit of the single frame at that one moment thus only needing a smaller bit of memory to do so. I imagine by doing this it renders that one tile, drops it from memory, and then immediately reads in new memory into its place. Not an expert by I imagine thats how it works.

[PlayStation 2]

1 minute ago, jaggysnake57 said:

errrrrr any one got a dumbed down version

 

The scene renders in smaller pieces rather than in whole.

[DocSwag]

2 minutes ago, Xorbot said:

I cannot be certain about the specifics of this type of rendering by NVIDIA.  However, tile-based rendering can lead to other optimizations such as hidden surface removal, the broad term to define this.

 

What happens is if a scene is properly rendered with opaque geometry first, then alpha-blended geometry sorted back-to-front, entire objects can be discarded completely.  Fragment processing would not proceed after this as well, as per usual rasterization.

 

So if a tile is detected as completely opaque, then triangle that lies completely behind the Z/depth values in the tile would be discarded and never even processed further.

So basically it could allow for discarding triangles that can't be seen anyways, thus improving performance... Cool!

 

Looks to me like there's a lot of advantages of using tile based rasterization!

[Xorbot]

Just now, Master Disaster said:

OK, so it can ignore some stuff that isn't required to reduce overall workload. Interesting and actually makes a tonne of sense once you think about it

That's why the tiles are small like 32x32 on PSVita, and in this NVIDIA test.. not sure what the smallest tile components' size are.  But they are indeed small compared to the main framebuffer.  Yes, I am a game programmer.

[Tomsen]

Great video from David Kanter.

 

Could this potentially have a small effect on the image quality rendered? (doubt it would be visible to a blind eye).

We haven't seen a real image comparison on GPUs for a long time.

 

14 minutes ago, DocSwag said:

So basically it could allow for discarding triangles that can't be seen anyways, thus improving performance... Cool!

 

Looks to me like there's a lot of advantages of using tile based rasterization!

Color me surprised, but I guess it was those same triangles Nvidia was pushing with gameworks.

[Sharkyx1]

28 minutes ago, Hunter259 said:

Think of how a single threaded application and a well written multi-threaded application behaves. Multi-threaded is much faster right? This is very similar to this.

Graphics rendering is already a massively parallel operation, every triangle, every shadow every pixel is a parallel render, thats why GPU performance scales with core count unlike CPU perf

[Hunter259]

Just now, Sharkyx1 said:

Graphics rendering is already a massively parallel operation, every triangle, every shadow every pixel is a parallel render, thats why GPU performance scales with core count unlike CPU perf

Yes I understand that. This is still just making it even more parallel.

[Dabombinable]

16 minutes ago, Tomsen said:

Great video from David Kanter.

 

Could this potentially have a small effect on the image quality rendered? (doubt it would be visible to a blind eye).

We haven't seen a real image comparison on GPUs for a long time.

 

Color me surprised, but I guess it was those same triangles Nvidia was pushing with gameworks.

And probably why AMD doesn't use it....

[patrickjp93]

43 minutes ago, Master Disaster said:

@patrickjp93Care to chime in and explain how tile rasterization is more efficient than normal? I'm struggling to understand because splitting the frame up still equals the same work done plus then it has to stitch everything together afterwards. 

Deduplication of lighting work. Stamp the same lighting signature across the similar tiles and then tweak as needed. It's the same with tiled texturing.

[Prysin]

19 minutes ago, patrickjp93 said:

Deduplication of lighting work. Stamp the same lighting signature across the similar tiles and then tweak as needed. It's the same with tiled texturing.

THat would explain why Nvidia GPUs get different lightning signatures in Ashes of the Singularity then AMD based cards. Their "tweaking" simply doesnt work and or doesnt fit with the way the game is built.

[Prysin]

25 minutes ago, Dabombinable said:

And probably why AMD doesn't use it....

Ashes of the singularity showed some peculiarities. At first people blamed it on Async compute not working properly. But if it is based on tweaking Lightning, as @patrickjp93 claims, then that would explain why Pascal GPUs consistently make a different lightning signature then Radeon GPUs in that game.

[Prysin]

1 hour ago, DocSwag said:

So basically it could allow for discarding triangles that can't be seen anyways, thus improving performance... Cool!

 

Looks to me like there's a lot of advantages of using tile based rasterization!

AMD introduced a similar system with Polaris. It is one of the Rasterization optimizations that came with GCN4