AMD's information on DX12 and new rendering technique for multi-gpu

[miagisan]

Seems good. I only have 1 question.

What is the TDP of the cards.! Really want to see if they will run hotter than the 290/Xs or cooler.

troll much?

[mr moose]

APEX physics, the ones run on the CPU crippling AMD ecosystems.

Im all against anti-competative measures, but I've been waiting for AMD for years to have an alternative PhysX (APEX type) (TressFX is not the same obviously) and while its not good to buy based off a single feature, the one game I wanna play out of them all,..Project Cars, uses it and I want full performance.

 

More or less, this is how the industry moves forward, with competition.  When a company can't compete they lose sales, this is their incentive to produce better products. Any one who expects a consumer to wait and buy an inferior product simply to support a company that can't keep up doesn't understand how the system works.

 

And besides all that, producing a better product is not anti competitive.  Producing a product that actually (in reality) stops another product from working might be, but we aren't seeing that here.

[ivan134]

More or less, this is how the industry moves forward, with competition.  When a company can't compete they lose sales, this is their incentive to produce better products. Any one who expects a consumer to wait and buy an inferior product simply to support a company that can't keep up doesn't understand how the system works.

 

And besides all that, producing a better product is not anti competitive.  Producing a product that actually (in reality) stops another product from working might be, but we aren't seeing that here.

Witcher 3 had 64x tessellation by default. AMD owners are actually in a better place because Kepler owners don't have the option to reduce it.

[mr moose]

Witcher 3 had 64x tessellation by default. AMD owners are actually in a better place because Kepler owners don't have the option to reduce it.

 

And the said competition wheel keeps turning forcing each company to fix their mistakes and make better products.   

[othertomperson]

"Frames no longer need to be queued" they do if it's remotely useful to be looking at a single image by the end

 

Looks like hsync is going to be thing soon

[SkilledRebuilds]

Witcher 3 had 64x tessellation by default. AMD owners are actually in a better place because Kepler owners don't have the option to reduce it.

https://www.youtube.com/watch?v=XnWkSFqo5A4

 

Fix for Kepler GPU's to get even more performance (up to 20%) with little to none visual degradation seen.

Video states NO visual losses, yet extra performance.

[sgloux3470]

Seems as if the new Split-frames method will cause screen tearing no?

Not anymore than current multigpu sets do. Civ beyond earth uses it for mantle crossfire with no issues.

Didn't realize multigpu was native now. I wonder if this will translate into 3-4 GPU's being a viable choice.

[That Norwegian Guy]

Some old dual GPU cards are going to become powerhouses again. GTX 590, GTX 690 and GTX 760 MARS, all kneecapped by their tiny to small frame buffers, will be able to do good work again with all their memory combined.

[ivan134]

https://www.youtube.com/watch?v=XnWkSFqo5A4

 

Fix for Kepler GPU's to get even more performance (up to 20%) with little to none visual degradation seen.

Video states NO visual losses, yet extra performance.

This is exactly what is meant when people say Nvidia sabotages performance. Tessellation that high offers no tangible visual improvements, but it lets them craft this false narrative that their competition is lacking behind, or their new cards are way better than their old ones. I'm glad someone found a fix for them. I'm not sure why people think it's okay for the game to run fine on only a grand total of 6 graphics cards.

[Misanthrope]

It sounds like this will depend on bandwidth a lot so it might get an advantage down the line with HBM and more over it might be key for the current 4gb limitation

[Megahurt]

SFR might make me consider a multi gpu system, I'm much concerned about frame latency than fps, seem like SFR would make both better.

[SkilledRebuilds]

This is exactly what is meant when people say Nvidia sabotages performance. Tessellation that high offers no tangible visual improvements, but it lets them craft this false narrative that their competition is lacking behind, or their new cards are way better than their old ones. I'm glad someone found a fix for them. I'm not sure why people think it's okay for the game to run fine on only a grand total of 6 graphics cards.

Yeah, Kepler performs as much as Maxwell due to the above change in the video....a simple dll file, causing so much grief, but when deleted, back to top tier performance rivaling the 970.

 

There have also been other game performance reviews I've seen where it looks like Kepler got ZERO optimizations while Maxwell gets the golden egg, and it shows in performance.. 780Ti losing to the GTX970 by more than 15-20fps, Pretty sure the new memory compression techniques are not that good... nor the Tessellation being improved upon that much. But I cannot say for sure, Im no engineer, but the 780Ti>GTX970 performance gap seems WAY WAY too wide.

[Xorbot]

The split frame rendering should not cause tearing if it works the way I think it does.

 

The GPU's work together with the same data, yes? Then the two halves should not be rendering two different viewpoints that would cause tearing, nor should it be pushing out two halves of a frame independently causing weird tearing on the monitor side of things. The GPU's should be rendering the two halves using the same data and combining it into one frame that it pushes out to the monitor essentially only combining the GPU power while sharing the same memory.

 

The only way I see this causing tearing is if the GPU's are not synced with each other and each is pulling data independent of each other, which should not happen if it is using pooled memory.

 

@Arreat

 

This is pretty much correct.

 

Mobile GPUs or SoCs have been doing this for a while with much smaller tiles.  My work on the PSVita shows that it uses 32x32 tiles.  This is why the vertical resolution is 544 pixels, not 540 which is half of 1080.  544 is evenly divisible by 32.

 

The benefit of this is that if there are quads that have the highest values in the z-buffer, and those quads were completely opaque (e.g. a wall), then geometry behind this that was meant to be rendered in a scene such as a tree can be discarded.  The game engine can still determine that the tree needs to be drawn because it is still visible in the final output image, in a specific 32x32 area of the screen, the house may be completely blocking the tree.

 

So breaking up the workloads into smaller tiles can have huge benefits.  I am no expert on DirectX 12 yet, but there could be these types of optimizations which go far beyond just saying "GPU #1: take the left side of the screen and GPU #2 take the right side."

 

So indeed, tearing should not exist because all rendering resource are contributing to the same frame in a tile-based rendering technique (even if it is just 2 tiles).

 

However, one apparently obvious downside is that this means a frame can be rendered only as fast as the slowest GPU.  If the left side is a Titan X and the right-side is a 650 Ti, then that left side will be done way faster than the right side.  Although I am no expert on DX 12 yet, perhaps if they could break this down into more smaller tiles, the Titan X can complete 85% of the frame's tiles while your old useless 650 Ti still has a tiny bit of use remaining and render the final 15% of tiles.  Sure, there may be a bit more overhead with breaking up a render it an increased amount of tiles, but at a smaller size, though it may be better for "very lop-sided" multi-gpu configurations.

 

Time will tell.

[SteveGrabowski0]

Please tell me the Hawaii rebrand isn't shipping with that horrible reference cooler shown in those ads.

[Doobeedoo]

This looks great!

I really want to see more in-depth stuff about multi-GPU in real world usage with new DirectX. It will be interesting to see how they behave, and how frames are delivered, also frame timing.

[Toddwjp]

AMD is really pushing the dx12 advertising, guess they want people to associate it with them even though it also gives Nvidia cards better performance as well.

 

although it will be nice to see them have cards that are equal to each other.

[Warren Pnigel]

This has probably been asked an answered before, but will DX12 allow vram stacking for nvidia cards?

[zMeul]

the problem with AMD's claim 4+4=8Gb is that for every video card powered by GCN above 1.0 GPU, AMD has renounced to use CFX bridges and the communications is done exclusively via PCIe

1. AMD's chipsets are not PCIe 3.0, but PCIe 2.x
2. the more cards you add, the less PCIe lanes each card has access to - CFX works with a low as 4X PCIe 2.x

[osbios]

Mobile GPUs or SoCs have been doing this for a while with much smaller tiles.  My work on the PSVita shows that it uses 32x32 tiles.  This is why the vertical resolution is 544 pixels, not 540 which is half of 1080.  544 is evenly divisible by 32.

Tilled rendering is widely used in mobile and desktop GPUs. But not because of multi GPU support. I just turns out that cache locality is one of the most critical parts with todays GPU speeds and memory latency. Also its really easy to scale tilled rendering by just throwing more computation cores at it.

 

The benefit of this is that if there are quads that have the highest values in the z-buffer, and those quads were completely opaque (e.g. a wall), then geometry behind this that was meant to be rendered in a scene such as a tree can be discarded.  The game engine can still determine that the tree needs to be drawn because it is still visible in the final output image, in a specific 32x32 area of the screen, the house may be completely blocking the tree.

That is something different, called Hierarchical Z-buffer and early Z rejection. They technically don't need tilling, but of course they work perfectly together and I'm sure there functionality to be unified in modern GPUs.

 

So breaking up the workloads into smaller tiles can have huge benefits.  I am no expert on DirectX 12 yet, but there could be these types of optimizations which go far beyond just saying "GPU #1: take the left side of the screen and GPU #2 take the right side."

 

So indeed, tearing should not exist because all rendering resource are contributing to the same frame in a tile-based rendering technique (even if it is just 2 tiles).

This technics have nothing to do with tearing! Everything uses double buffering. What you notice as tearing is the switch from one fully rendered image to the next fully rendered image without waiting for the monitor scanline to be at the start of the screen again.

[Agent181]

possibly, though it will be much easier on each gpu as the effective resolution is halfed

wouldn't this be a lot more noticeable and then if you have a free sync monitor it should help in theory but wouldn't that cause some problems since its more than 1 gpu for a single free sync module?

[PlayModeOn]

Does this mean we can use Intel's integrated graphics with an AMD/Nvidia discrete GPU?

[Megahurt]

Split frame rendering is probably how "vram stacking" will work, each gpu only renders half the frame, so it needs less vram to do so. Sharing memory through the 16 GB/s pci-e connection (and that's if both cards are running pci-e 3.0 @ 16X) would make no sense at all.

[Xorbot]

This technics have nothing to do with tearing! Everything uses double buffering. What you notice as tearing is the switch from one fully rendered image to the next fully rendered image without waiting for the monitor scanline to be at the start of the screen again.

 

Indeed.  I wasn't referring to the tearing associated with v-sync, I was speaking more about DX 12 showing the front buffer before both "tiles" have completed.  V-sync tearing would still exist if it was disabled and the framerate was not the same as the monitor's refresh rate.

[Opcode]

Does this mean we can use Intel's integrated graphics with an AMD/Nvidia discrete GPU?

For post-processing I would imagine so as we have a news thread on here about that. Although for something like asymmetric rendering it could be a little more complex.

 

Split frame rendering is probably how "vram stacking" will work, each gpu only renders half the frame, so it needs less vram to do so. Sharing memory through the 16 GB/s pci-e connection (and that's if both cards are running pci-e 3.0 @ 16X) would make no sense at all.

It makes perfect sense as traditionally in a multiple GPU configuration an entire frame gets piped to both cards for rendering. With DirectX 12 only half the frame needs to be sent to each card not only effectively reducing the dependency on VRAM density but also bandwidth requirements on the PCIe interface.

[trag1c]

Split frame rendering is probably how "vram stacking" will work, each gpu only renders half the frame, so it needs less vram to do so. Sharing memory through the 16 GB/s pci-e connection (and that's if both cards are running pci-e 3.0 @ 16X) would make no sense at all.

I don't think the frames will actually be "split" since there would be no way to actually determine which resources each card needs to have loaded to actually split the rendering. The frames would probably be done through work queues where each card has a set of resources and is given a job when a draw call uses one of those resources, after all jobs have been completed it would probably be sent to the master card for final image computation.