AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
64 ROPs seems like an... interesting design decision if that's what it ends up as.
Higher clocks might help a bit compared to Navi10, but it'll be interesting to see how it stacks up at super high resolutions and frame rates.

For example, games at 8k resolution or even 4k with a 120fps target, or anything VR which tends to be very light on compute but heavy on ROPs and bandwidth. Doubly so for those that use forward renderers and MSAA.
8K resolution is in all terms completely irrelevant even for highest end. As long as you can do 4k it's enough
 
4K requires as much fillrate as 1080p 4xAA. Add all those fancy buffers and shadow maps that games create every frame and more fillrate may actually be useful.
Yeah but have ROPs been a bottleneck at all in the last 10 or so years? And as already stated, clock speeds have skyrocketed recently.
 
4K requires as much fillrate as 1080p 4xAA. Add all those fancy buffers and shadow maps that games create every frame and more fillrate may actually be useful.

And forward rendering and MSAA is definitely not dead for VR - in fact, it's the recommended state-of-the-art approach, although clustered forward rendering is probably a better compromise. An awful lot of effort has been spent making good forward renderers specifically because in VR you want 4x or even 8x MSAA.

Combine that with a 4k or higher resolution frame buffer per eye and then rendering out at 90fps in most cases, and even 120 or 144hz with the most modern headsets and it's pretty easy to see how you can be fillrate bound in a hurry.

https://docs.unrealengine.com/en-US/Engine/Performance/ForwardRenderer/index.html

It's one of the reasons why in a lot of VR titles the 1080Ti punches way above its usual weight class against Turing. Clock speeds barely budged with Turing and ROP counts didn't change tier to tier. For a non-VR game you'd expect the 2080 and especially the 2080 Super to beat the 1080Ti in every single scenario. In VR that's not the case, as the +50% fillrate advantage actually matters and they're generally quite light on compute.

Here's hoping big Navi is a beast in that regard!
 
Yeah but have ROPs been a bottleneck at all in the last 10 or so years? And as already stated, clock speeds have skyrocketed recently.

We still need ROPs because they have a blending unit included in them. Even with Nvidia's Marble RTX demo in all of it's glory it lacked ray traced caustics on the marbles themselves ... (their shadowing felt unnaturally uniform for a transparent object)

Alpha blending/transparency is a part of the wider phenomena of light scattering. Light rays don't just reflect but they can scatter inside the medium as well like we see with volumetric objects such as clouds which can either brighten or darken with respect to their thickness.

ROPs were traditionally used as a rough approximation for some of these wider phenomena and we'll continue to use them because I highly doubt that the near future will involve ray traced volumetric rendering or transparency/caustics on a widespread basis. It'll be a problem to attempt for the next generation after this coming generation ...
 
And forward rendering and MSAA is definitely not dead for VR - in fact, it's the recommended state-of-the-art approach, although clustered forward rendering is probably a better compromise. An awful lot of effort has been spent making good forward renderers specifically because in VR you want 4x or even 8x MSAA.

Combine that with a 4k or higher resolution frame buffer per eye and then rendering out at 90fps in most cases, and even 120 or 144hz with the most modern headsets and it's pretty easy to see how you can be fillrate bound in a hurry.

https://docs.unrealengine.com/en-US/Engine/Performance/ForwardRenderer/index.html

It's one of the reasons why in a lot of VR titles the 1080Ti punches way above its usual weight class against Turing. Clock speeds barely budged with Turing and ROP counts didn't change tier to tier. For a non-VR game you'd expect the 2080 and especially the 2080 Super to beat the 1080Ti in every single scenario. In VR that's not the case, as the +50% fillrate advantage actually matters and they're generally quite light on compute.

Here's hoping big Navi is a beast in that regard!
VR is an incredibly niche market though. Its not showing much growth either.
 
We still need ROPs because they have a blending unit included in them. Even with Nvidia's Marble RTX demo in all of it's glory it lacked ray traced caustics on the marbles themselves ... (their shadowing felt unnaturally uniform for a transparent object)

Alpha blending/transparency is a part of the wider phenomena of light scattering. Light rays don't just reflect but they can scatter inside the medium as well like we see with volumetric objects such as clouds which can either brighten or darken with respect to their thickness.

ROPs were traditionally used as a rough approximation for some of these wider phenomena and we'll continue to use them because I highly doubt that the near future will involve ray traced volumetric rendering or transparency/caustics on a widespread basis. It'll be a problem to attempt for the next generation after this coming generation ...
Actually I think marbles rtx did have caustics on the marble, the Thing is that the denoiser in motion erases them. Still shots from the Demo have caustics on the marble for example.
 
Fun experiment - take a glass in your kitchen and fill it with water. Put it on the table. Barring sun shining through a window and falling on th glass, you won’t see any at all. Next, take it out into the sun. If you put it down in the sun you will see caustics, but question #1 - could you have accurately predicted the pattern before you saw it? Question #2 - could you have accurately predicted it’s intensity given degree of haze/clouds and the surface the glass stands on?
For me, even having done these experiments and explicitly looking for the results, the answer to both of these are still NO!
The phenomenon is both complex, and belongs to that plethora of visual data that the brain wisely treats as noise and discards.
And the conlusion from that is that you can either fake the effect (quite crudely is perfectly fine), or disregard it entirely.
Your brain won’t mind, it doesn’t care and has no idea what would be physically correct anyway.
 
Fun experiment - take a glass in your kitchen and fill it with water. Put it on the table. Barring sun shining through a window and falling on th glass, you won’t see any at all. Next, take it out into the sun. If you put it down in the sun you will see caustics, but question #1 - could you have accurately predicted the pattern before you saw it? Question #2 - could you have accurately predicted it’s intensity given degree of haze/clouds and the surface the glass stands on?
For me, even having done these experiments and explicitly looking for the results, the answer to both of these are still NO!
The phenomenon is both complex, and belongs to that plethora of visual data that the brain wisely treats as noise and discards.
And the conlusion from that is that you can either fake the effect (quite crudely is perfectly fine), or disregard it entirely.
Your brain won’t mind, it doesn’t care and has no idea what would be physically correct anyway.

Maybe for glasses of water but my brain certainly doesn't discard all caustics. Take a swim in the caribbean. Guarantee you will notice and appreciate them!
 
Maybe for glasses of water but my brain certainly doesn't discard all caustics. Take a swim in the caribbean. Guarantee you will notice and appreciate them!
Of course. It’s not that there aren’t conditions where they exist, it’s that outside academia it’s pointless to have an accurate rendition of them.
 
It's also pointless to render accurate caustics, as long we are unable to do proper fluid simulation in realtime at all. Cant' hide the fake :D
 
So it's 64 ROPs again, for a ~20TFLOPs GPU?

Even if the core is clocking at 2.2GHz or so, the fillrate/compute ratio will be completely different to Navi 10 and Vega.
 
So it's 64 ROPs again, for a ~20TFLOPs GPU?

Even if the core is clocking at 2.2GHz or so, the fillrate/compute ratio will be completely different to Navi 10 and Vega.
Does it matter, though? It's ought to be more than enough for 4K
 
So it's 64 ROPs again, for a ~20TFLOPs GPU?

Even if the core is clocking at 2.2GHz or so, the fillrate/compute ratio will be completely different to Navi 10 and Vega.
The FLOPS to fillrate ratio is very close to Polaris / RX 480 (2304 SPs / 32 ROPs) and not far from Tahiti / HD 7970 in 2012 (2048 SPs / 32 ROPs) :)

/corrected
 
Last edited:
Status
Not open for further replies.
Back
Top