Ehh, what? So far the only thing AMD acknowledged is the name, and practically all the rumors are WTFtech-level, or worse.
- Status
Ehh, what? So far the only thing AMD acknowledged is the name, and practically all the rumors are WTFtech-level, or worse.
DavidGraham
Veteran
David Wang of AMD says they won't offer DXR support until it's available on all products, which rules out DXR support for Navi.
https://wccftech.com/amds-david-wan...-dxr-until-its-offered-in-all-product-ranges/
https://wccftech.com/amds-david-wan...-dxr-until-its-offered-in-all-product-ranges/
AMD has said it does feature nexgen memory (again typo intended) and focuses on scalability
Actually no he didn't. WCCFTech probably used machine translation or something. Here's a translation by a native japanese guy who tweets about video cards and stuff too:
https://www.overclock3d.net/news/gp..._respond_to_directx_raytracing_-_david_wang/1
DmitryKo
Veteran
Next-gen 'Arcturus' would be a high-end HBM3 part, like Fiji and Vega were, while Navi would be a mid-range GDDR part, sitting well below.
Possibly, possibly not. The focus on scalability would on the contrary say that Navi will be top to bottom
Megadrive1988
Veteran
PS3 was released around the same time Nvidia released the G80 / Geforce 8800 GTX. Both came out Fall 2006.
DavidGraham
Veteran
Actually yes, he denied it by omission, he was specifically asked when are you planning to support DXR, he didn't answer Navi, he just dogged the question and then provided a justification for the dodge. So no, if Navi was to support DXR he would specifically say so.
It also means Vega won't probably support DXR, even for it's fallback layer.
Last edited:
Frenetic Pony
Veteran
It's not terribly hard to support DXR.
Now just how much specific hardware support Navi might have is another question. There's the requirements for thread spawning and etc. And then there's what Turing does, which is to have dedicated ray intersection hardware and cutting warps/whatever to 16 apiece to support more divergence. That takes a more dramatic redesign, and increased costs and time, to implement. I'd not be surprised at minimal HW support for DXR in Navi, but moving beyond that seems less likely.
That being said, so far most games that do more than shadows/ambient occlusion with it are reported to take a dramatic hit in performance for enabling their one feature that uses raytracing. That is of course based off the seemingly NVIDIA specified (since all the targets are the same, mysteriously) target of 1080p 60fps, on a 2080. Considering a 2080 could hit 4k 60 (or near to it) on many of these titles otherwise, well that's a fourth of the performance for one feature; even if the target is a 2070 that's still halving or more the peformance.
All meaning AMD might not even care that much about adding dedicated ray intersection HW at the moment. The smartest thing to use raytacing for is shadows anyway. Nice and coherent, no extra shading, gets rid of a ton of headaches and bad shadow mapping artifacts, lets artists create shadow casting lights with very large radii without killing performance, and can even be backwards compatible if you're doing SDF/voxel tracing. Much of the same goes for ambient occlusion, as while for a better effect you'd need to grab color at ray intersection, you still only need to shade the origin pixel with your data instead of shading your hit pixel, still nice and relatively coherent. Since both can be very fast (faster than shadow maps for shadows) without dedicated RT hardware, and neither relies on divergent shading, AMD might conclude that'd be the way most developers would rather use raytracing for introducing it as a mainstream feature. Thus no need for splitting their wavefronts into smaller units, at least at the moment.
Now just how much specific hardware support Navi might have is another question. There's the requirements for thread spawning and etc. And then there's what Turing does, which is to have dedicated ray intersection hardware and cutting warps/whatever to 16 apiece to support more divergence. That takes a more dramatic redesign, and increased costs and time, to implement. I'd not be surprised at minimal HW support for DXR in Navi, but moving beyond that seems less likely.
That being said, so far most games that do more than shadows/ambient occlusion with it are reported to take a dramatic hit in performance for enabling their one feature that uses raytracing. That is of course based off the seemingly NVIDIA specified (since all the targets are the same, mysteriously) target of 1080p 60fps, on a 2080. Considering a 2080 could hit 4k 60 (or near to it) on many of these titles otherwise, well that's a fourth of the performance for one feature; even if the target is a 2070 that's still halving or more the peformance.
All meaning AMD might not even care that much about adding dedicated ray intersection HW at the moment. The smartest thing to use raytacing for is shadows anyway. Nice and coherent, no extra shading, gets rid of a ton of headaches and bad shadow mapping artifacts, lets artists create shadow casting lights with very large radii without killing performance, and can even be backwards compatible if you're doing SDF/voxel tracing. Much of the same goes for ambient occlusion, as while for a better effect you'd need to grab color at ray intersection, you still only need to shade the origin pixel with your data instead of shading your hit pixel, still nice and relatively coherent. Since both can be very fast (faster than shadow maps for shadows) without dedicated RT hardware, and neither relies on divergent shading, AMD might conclude that'd be the way most developers would rather use raytracing for introducing it as a mainstream feature. Thus no need for splitting their wavefronts into smaller units, at least at the moment.
Err, no. They haven't commented specifics even Zen 2 / Rome which should be closer to actually hitting the markets, not mentioning Navi will or will not support it doesn't say anything.
You are right, I got it wrong - 2007 was 8800 Ultra. This makes the console vs dgpu architecture being the same argument even weaker.
DavidGraham
Veteran
They already mentioned a lot of specifics about Rome and it's main supporting features, but hid others for competitive reasons. DXR is not really a military secret at this point, It's an industry standard, you either show up to support it or you don't, there is no need or advantage of being "coy" about it. Unless you are really not supporting it, which is what the statement already implies.
For example, NVIDIA announced support for DXR in Volta and later architectures, before we knew about Turing and RT cores.
Anyway, I expect we will get further clarification on this soon enough.
Last edited:
DavidGraham
Veteran
The capability of DXR to run on DX12 hardware as a fallback layer is there, whether a manufacturer will produce a working driver for it or not, is up in the air.
It's like DX12, which was supported on almost all DX11 hardware. NVIDIA chose to support it all the way back to Fermi, but AMD chose to leave out VLIW GPUs (HD5000 and HD6000) out of the DX12 loop.
I would bet we'll see it end of 2020 with their 7nm+ Gen. When Nvidia launches their 7nm line, it will be down to the x60 class gpus between 200-300$. In 2020 there's a high risk for amd, that Intel will also launch with raytracing hardware, as they did a lot of research in that space. Whatever amd launches at that timeframe in 7nm+ will probably hold 2 years before we get the next shrink. It's not really important whether they really want to support DXR, the whole risk management in your company will push the gpu hardware designers to include it. As a company you can't take such a massive risk to not support such a big feature for a long time.
Only against Nvidia having console design wins this might work. But intel will also push in the market with game collaborations and you can't risk having 5,6 DXR games per year by Nv and 5,6 by Intel and all tell you it's not possible with amd for 2 years. Your sales will fall of a cliff.
DmitryKo
Veteran
To clarify, compute-based D3D12 Raytracing Fallback Layer is a DirectX SDK sample - however it is deprecated and the code will not be maintained any further, so new revisions of the DirectX Rarytracing API, which may bring new useful features like geometry LOD, would require explicit support in the video driver.
So far only Nvidia Volta has a working "software" DXR implementation in the driver.
Last edited:
DavidGraham
Veteran
Now that explains a lot in the context of David Wang's statement.
But not necessarily in special fixed function hardware. It could probably be implemented as some kind of FP16 Compute Shader, and would at least allow to test some stuff.
The API expects the driver to expose requested features, not how they are handled by the device.
DmitryKo
Veteran
I see little point in a software implementation when even expensive hardware-accelerated cards are struggling to come with acceptable performance levels. Nvidia did it as a stop-gap solution for developers so they could play with the API before GeForce RTX cards were available.
For the time being, AMD will probably stick with their Open-CL based GPUOpen libraries like FireRays/Radeon Rays intersection search and ProRender/Baikal path/ray-tracing renderer; these were designed for professional (CAD/CAM CGI) applications though.
For the time being, AMD will probably stick with their Open-CL based GPUOpen libraries like FireRays/Radeon Rays intersection search and ProRender/Baikal path/ray-tracing renderer; these were designed for professional (CAD/CAM CGI) applications though.
Last edited:
- Status