AMD RDNA3 Specifications Discussion Thread

Have you seen the frequencies of the 7900XT?
Basis: 1,500MHz (vs. 1,900MHz)
Game: 2,000MHz (vs. 2,300MHz)
Boost: 2,300MHz (vs. 2,400MHz)

Its a huge downgrade from the 7900XTX. And using the game frequency compute performance would have been only 8% higher than the 4080 12gb (2620Mhz boost/game frequency).
 
I wonder if they're limited by the front end being clocked faster?

The CU's might have 20% higher clock headroom in them but if the front end can't also go +20% then you can't increase the the CU clock as you'll get stability issues due to the front end.
 
I wonder if they're limited by the front end being clocked faster?

The CU's might have 20% higher clock headroom in them but if the front end can't also go +20% then you can't increase the the CU clock as you'll get stability issues due to the front end.
The front end clocks 200MHz higher. The compute units are the limitation.
 
The front end clocks 200MHz higher. The compute units are the limitation.

The front end will always be the limit as you're limited by the maximum stable clock of the fastest part.

If the CU's can do 2.9Ghz the front end will have to do 3.1Ghz which is a big ask.

If the front end maxes out at 2.8Ghz then the CU's will be limited to 2.6Ghz even if they have room to go higher.

On old Nvidia GPU's that ran the core and shaders at different clock rates there was an option to decouple the clock speeds and clock them separately, hopefully there's a similar thing for these new AMD cards.
 
The shader engines are clocked lower. So they are the limiting factor. Otherwise AMD would have gone for a 1:1 clock rate. And when you see that the 7900XT has a 300MHz lower "game frequency" it shows that the compute units are not as efficient as they should be to be clocked much higher.
 
The shader engines are clocked lower. So they are the limiting factor. Otherwise AMD would have gone for a 1:1 clock rate. And when you see that the 7900XT has a 300MHz lower "game frequency" it shows that the compute units are not as efficient as they should be to be clocked much higher.

AMD claim clocking them lower saved them 25% power draw, so that's likely the main reason for their low clock speed and not because they can't clock higher.
 
Introducing curved surfaces is going to contribute to noise due to the divergent directions of the rays ...
Not necessary. Smooth curved surfaces would reflect rays evenly in similar directions (not random), so you will get stretched reflection and a bit of aliasing on geometry edges without TAA and with just 1 rpp or below inside of the reflection, but you won't get noise, at least if you sample textures with mips in hit points.

Coherent rays ? Curved surfaces ? Spatio-temporal stable image ?
Curved surfaces can provide very smooth mirror reflections (if these are processed per triangles faces), you can spot them in many games and demos with RT, check reflections in CP2077 on glass surfaces for example.

I don't believe many real-time applications with realistically be able to use 1000 mirrors
Yes, of course, and that's the whole point, as I said in the very beginning of our discussion, even a single quarter res planar reflection can drop performance by a factor of 2x in Cyberpunk 2077, so adding more planar surfaces certainly won't be feasible.


The tradeoff between full blown RT reflections and planar reflections in terms of quality/elegance isn't one-sided as you would implicate them to be ...
Sorry, but I am just too lazy to argue even further. It's quite one-sided and that's why people move to RT reflections, and not otherwise.
 
AMD claim clocking them lower saved them 25% power draw, so that's likely the main reason for their low clock speed and not because they can't clock higher.
Yes, they have to down clock the shader engines because these compute units are not as efficient as they should be. AMD is power limited and they wont go above 375W. So they lose around 5% performance for 25% energy savings.
 
Yes, they have to down clock the shader engines because these compute units are not as efficient as they should be.
The chiplet design and moving data off die surely doesn't help with the efficiency either. The interconnect also costs them a non-insignificant area. If their current chiplet design is an overall win in terms of PPA, it doesn't seem like a big one, but I suppose it's only one step forward towards a greater goal. If they only did more than the bare minimum for RT performance, I'd have been satisfied with this.
 
Not necessary. Smooth curved surfaces would reflect rays evenly in similar directions (not random), so you will get stretched reflection and a bit of aliasing on geometry edges without TAA and with just 1 rpp or below inside of the reflection, but you won't get noise, at least if you sample textures with mips in hit points.
@Bold only if the points in the reflective surface share a direction similar to their normals which is basically near planar surfaces ...

Arbitrary normals in surface = Arbitrary ray directions
Curved surfaces can provide very smooth mirror reflections (if these are processed per triangles faces), you can spot them in many games and demos with RT, check reflections in CP2077 on glass surfaces for example.
Doesn't really help your case that the image uses Nvidia Ansel which is capable of cheating by rendering ingame scenes with higher sample counts ...

Yes, of course, and that's the whole point, as I said in the very beginning of our discussion, even a single quarter res planar reflection can drop performance by a factor of 2x in Cyberpunk 2077, so adding more planar surfaces certainly won't be feasible.

Sorry, but I am just too lazy to argue even further. It's quite one-sided and that's why people move to RT reflections, and not otherwise.
It really isn't one-sided, cheating in an acceleration structure really isn't superior in all aspects to rendering multiple geometry passes especially with the hardware we're talking about in this thread where it may very well be the superior option in terms of both quality and performance ...
 
Btw, 6nm for the infinitycache is still a good process. On paper, something like 12nm wouldn't be possible ? For a cheaper product ?
 
So 7970XTX have 6144 SP, each of them manage dual FMA/clock?

It actually seems to be working pretty well. Without co-issue you’re looking at 30 tflops which is just 30% more than the 6900xt. In order to hit AMDs 50-70% numbers there must be quite a bit of coissue firing. Or is there some other trickery where multiple wavefronts can execute in the same clock?
 
Back
Top