computerbaseAshes of the Singularity Beta1 DirectX 12 Benchmarks

[railven]

Interesting info on the Hitman thing. Makes me wonder if it will bring extra Image Quality or just performance gains? Also makes me wonder if it will mirror Ashes - ie, an AMD user not using DX12 is screwed.

Secondly, I wonder if NV's DX11 path performance would be equal to AMD's DX12 path performance.

So many questions. But I look forward to some of these answers. After all the DX12 sales pitch and promising future of (more than 100% GPU utilization? WAAAAAT?) it's about time we get something concrete.

 

[dacostafilipe]

I've only seen max/min/avg FPS comparison of Ashes between DX11 and DX12.

Did somebody see a frametime variation graph somewhere? Thx

 

[guskline]

SilverForce11, I'm at work right now, but when I get home this evening I'll try to run the Ashes of Singularity benchmark on both of my rigs below to see if I can detect the differences you point out.

 

[Spjut]

Don't nvidia support Asynchronous Compute? No. Under DX12, Nvidia do not support Asynchronous Compute.

Could Nvidia support it via NVAPI? Like how Nvidia did support certain DX10.1 features on its Tesla GPUs despite not being fully compatible?

And is the asynchronous compute problem something that could be solved relatively easy in a minor DX12 revision?

 

[Mahigan]

Could Nvidia support it via NVAPI? Like how Nvidia did support certain DX10.1 features on its Tesla GPUs despite not being fully compatible?

And is the asynchronous compute problem something that could be solved relatively easy in a minor DX12 revision?

The Async compute this is likely due to nvidia not supporting some mundane DX12 requirement in hardware.

Nvidia do support Asynchronous Compute when working with CUDA. An example of this is PhysX and DX11. Titles which use PhysX are using CUDA and by extension execute compute and graphics commands concurrently (though not asynchronously).

For full asynchronous compute support, nvidia would need to incorporate ACE-like units in their hardware. This is what nvidia lack right now. I mean their scheduler relies heavilly on their software driver and much of CUDA's HyperQ solution (nvidia's async compute-like feature) is a software implementation.

This leads me to believe that under Vulcan, nvidia may be able to support concurrent execution of graphics and compute workloads.

 

[Spjut]

Sorry if I'm getting on your nerves now, but I guess what I'm asking is, why can't Microsoft and Nvidia make an additional DX12 Asynchronous Compute feature adapted to Kepler/Maxwell?

 

[dogen1]

This leads me to believe that under Vulcan, nvidia may be able to support concurrent execution of graphics and compute workloads.

If you're right, that's great news for amd and nvidia users alike.

It definitely seems likely considering cuda, and how vulkan is designed to work well for all vendors.

 

[Paul98]

Sorry if I'm getting on your nerves now, but I guess what I'm asking is, why can't Microsoft and Nvidia make an additional DX12 Asynchronous Compute feature adapted to Kepler/Maxwell?

It can't if the hardware doesn't support it.

 

[Spjut]

It can't if the hardware doesn't support it.

But he says Nvidia has a similar solution that works when using CUDA

 

[dogen1]

It can't if the hardware doesn't support it.

He's asking why can't they add a workaround for maxwell that'll allow them to do what they already seem to be able to do with cuda.

I have no idea. It's probably more complex than you and I realize anyway.

 

[ThatBuzzkiller]

But he says Nvidia has a similar solution that works when using CUDA

Nvidia has no hardware solution ...

They can only run multiple compute queues but they can't interleave graphics and compute queues ...

Nvidia may support concurrent copy queues but it doesn't go any further than that ...

 

[Mahigan]

Sorry if I'm getting on your nerves now, but I guess what I'm asking is, why can't Microsoft and Nvidia make an additional DX12 Asynchronous Compute feature adapted to Kepler/Maxwell?

I'm sure they could but Microsoft releases standards and expects that vendors will comply. Nvidia probably made a mistake in their implementation. Since DX12 is a standard to follow, nvidia will have to incorporate a fix in future GPU designs.

Graphics manufacturers must comply with API requirements. Not the other way around.

 

[3DVagabond]

Really bad quality images.

Biggest difference is missing Bloom.
Not really sure if there is actually any difference in quality of lighting. (Done by their 'object space rendering' approach.(Which is basically shading objects in texture space or Reyes shading in textures instead of micropolygons.))

Bloom? The biggest difference? Do you mean the lighting? Lots of lighting. There's also a lot less smoke. Possibly done with particles. When you look at scenes where there are multiple fighters moving quickly the nVidia rendering is a stutterfest.

This is night and day difference. nVidia said they were going to implement the DX12 async compute in drivers. Are they? Lots of people here seemed to think it was a simple task with nVidia's driver team, money, and software capabilities.

 

[xthetenth]

I'm sure they could but Microsoft releases standards and expects that vendors will comply. Nvidia probably made a mistake in their implementation. Since DX12 is a standard to follow, nvidia will have to incorporate a fix in future GPU designs.

Graphics manufacturers must comply with API requirements. Not the other way around.

Yep, if you want the spec changed you join the group and lobby for it. If you want the actually codified part of the spec that people have had a chance to implement against changed (or heck, an experimental that's been around long enough that people have implementations), you have fun with that. It's not going to happen.

 

[Mahigan]

Bloom? The biggest difference? Do you mean the lighting? Lots of lighting. There's also a lot less smoke. Possibly done with particles. When you look at scenes where there are multiple fighters moving quickly the nVidia rendering is a stutterfest.

This is night and day difference. nVidia said they were going to implement the DX12 async compute in drivers. Are they? Lots of people here seemed to think it was a simple task with nVidia's driver team, money, and software capabilities.

From information I've received (source) they were working on doing just that but the solution hammered the CPU hard under AotS. Mostly because the CPU is already hammered hard under that title.

Its been several months now since then (I received the info back in October). It looks to me like the idea was abandoned and that instead nvidia will push DX11 + game works instead. Partnering with studios in order to counter what would otherwise be a rather embarrassing scenario.

All that having been said, the GTX 980 Ti is an incredible piece of hardware. It is every card under it that really lacks any sort of punch.

 

[tential]

Sorry if I'm getting on your nerves now, but I guess what I'm asking is, why can't Microsoft and Nvidia make an additional DX12 Asynchronous Compute feature adapted to Kepler/Maxwell?

This is nvidia we're talking about. Why would they do this? It just means less reasons to upgrade to pascal.

 

[Azix]

Is there another side by side comparison that I can check out? Anyone talked about this difference?

https://www.youtube.com/watch?v=o2b6Nncu6zY

 

[Silverforce11]

Bloom? The biggest difference? Do you mean the lighting? Lots of lighting. There's also a lot less smoke. Possibly done with particles. When you look at scenes where there are multiple fighters moving quickly the nVidia rendering is a stutterfest.

This is night and day difference. nVidia said they were going to implement the DX12 async compute in drivers. Are they? Lots of people here seemed to think it was a simple task with nVidia's driver team, money, and software capabilities.

This was the original Ashes unveil very early on, running on Mantle.

https://www.youtube.com/watch?v=t9UACXikdR0&feature=youtu.be&t=101

You notice a lot of dynamic lights on all the projectiles, lots of smoke trails, basically similar to the images we see on the side by side comparison for the 390X.

Notice Oxide specifically mentions "Thanks to Mantle & you can do this on DX12 or Vulkan too... every single shot is casting light." You can clearly see dynamic lighting of the projectiles.

I noticed it was clearly absent in the first alpha benchmark that tech sites did.

On the 970 @ Digital Foundry
http://www.eurogamer.net/articles/digitalfoundry-2015-ashes-of-the-singularity-dx12-benchmark-tested

There is absolutely zero dynamic light casting in that initial release for NV GPUs.

Compared to the 390 @ Digital Foundry

Notice there is some dynamic lights, but much less than the initial Mantle reveal and much less than the recent 390X side by side video.

Then it makes sense when you put that into context with what Oxide have said, they disabled async compute at NV's request. That there is indeed a vendor specific path and it's for NV only.

The Radeons are rendering a scene with many dynamic light sources all that time while the NV GPUs according to these shots from tech sites above are not.

 

[railven]

Is there another side by side comparison that I can check out? Anyone talked about this difference?

https://www.youtube.com/watch?v=o2b6Nncu6zY

Already being discussed in the oriignal thread, no need for a new one.

 

[Azix]

Already being discussed in the oriignal thread, no need for a new one.

was this a massive thread? about which page

nvm searched video

 

[Azix]

The specs showcase DX11 for nvidia cards, the DX12 path is being added by a joint venture between I/O and AMDs Gaming Evolved initiative.

The game would require two separate DX12 paths in order to support nvidia, unless nvidia spend the money to implement their own path it won't get done.

Nvidia cards will not be able to run the AMD path as they don't support Asynchronous Compute.

you said something like this before on OCN and I still wonder how accurate it is. I guess we'll see when the game comes out if dx12 works on nvidia cards.

My thinking is they simply run the compute tasks through the graphics queue for nvidia. It would just be slower.

What's the verdict on the side-by-side video? I don't think its just monitor because even then you would see the effects, just with less brightness etc.

 

[Mahigan]

you said something like this before on OCN and I still wonder how accurate it is. I guess we'll see when the game comes out if dx12 works on nvidia cards.

My thinking is they simply run the compute tasks through the graphics queue for nvidia. It would just be slower.

What's the verdict on the side-by-side video? I don't think its just monitor because even then you would see the effects, just with less brightness etc.

Sure, you can run the compute tasks sequentially in the graphics queue but the DX12 path, for hitman, is built around GCN and Asynchronous compute. The shaders being used will be GCN optimized (AMD Gaming Evolved title). You end up with the same scenario we had with AotS. In this scenario the developer couldn't get even a GTX 980 Ti to run the code at a playable frame rate. Oxide worked with nvidia and implemented a vendor ID specific path without Async Compute and with nvidia optimized shaders. That's what nvidia runs today in AotS.

AotS barely used Async Compute as per Oxide. Hitman will be leveraging Asynchronous compute moreso than any PC title to date. In other words, far more compute work loads than AotS. If you run those sequentially, you'll likely hit a wall on a GTX 980 Ti.

But you're right, only time will tell.

 

[Mahigan]

This was the original Ashes unveil very early on, running on Mantle.

https://www.youtube.com/watch?v=t9UACXikdR0&feature=youtu.be&t=101

You notice a lot of dynamic lights on all the projectiles, lots of smoke trails, basically similar to the images we see on the side by side comparison for the 390X.

Notice Oxide specifically mentions "Thanks to Mantle & you can do this on DX12 or Vulkan too... every single shot is casting light." You can clearly see dynamic lighting of the projectiles.

I noticed it was clearly absent in the first alpha benchmark that tech sites did.

On the 970 @ Digital Foundry
http://www.eurogamer.net/articles/digitalfoundry-2015-ashes-of-the-singularity-dx12-benchmark-tested

There is absolutely zero dynamic light casting in that initial release for NV GPUs.

Compared to the 390 @ Digital Foundry

Notice there is some dynamic lights, but much less than the initial Mantle reveal and much less than the recent 390X side by side video.

Then it makes sense when you put that into context with what Oxide have said, they disabled async compute at NV's request. That there is indeed a vendor specific path and it's for NV only.

The Radeons are rendering a scene with many dynamic light sources all that time while the NV GPUs according to these shots from tech sites above are not.

I never noticed that before...

 

[Silverforce11]

I never noticed that before...

Lol yeah I didn't either until I saw the recent side-by-side... then I went back to the initial press release benchmark and there it was.

NV GPUs are not rendering dynamic lights. Exactly matching Oxide's statement that their mass usage of dynamic lights are possible due to Mantle/DX12/Vulkan features, and that they disabled Async Compute at NV's request, adding NV specific rendering path with optimized shaders for them.

When they posted on OC3D forums and they actually said along the lines of Ashes is actually more optimized for NV. Makes sense, AMD GPUs are working so much more work.

 

[Mahigan]

Lol yeah I didn't either until I saw the recent side-by-side... then I went back to the initial press release benchmark and there it was.

NV GPUs are not rendering dynamic lights. Exactly matching Oxide's statement that their mass usage of dynamic lights are possible due to Mantle/DX12/Vulkan features, and that they disabled Async Compute at NV's request, adding NV specific rendering path with optimized shaders for them.

And we've been comparing the benchmarks as if they're apples to apples when in fact the GCN cards are processing more effects...

Hmm...