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

My guess is within $100 of the 3080 if these numbers hold up. Probably a cheaper option to compete with the 3070.

 

I dunno, like others have said, exact settings matter but they are definitely within striking distance.

How? 96CUs at 2500mhz barely breaks 30TF.

 

they are aiming for the 3080 in their comparisons, quite clearly, this is going to be interesting. I want to know more about RT performance but this is looking good

 

Infinity FP32 , of course.

 

That statement is (a) false and (b) makes no sense. The 3070 has already been announced. Oct 29 is when it's available for purchase. Pushing it to the day *after* the RX reveal provides no competitive benefit to Nvidia -- it gives people a chance to observe the AMD response and make an informed choice.

 

It gives them the ability to make a last minute change to the price if Navi 21 is too close to it.

 

I hope so as well, and given the leaks it certainly might be the second tier card. But even otherwise, it seems within striking distance of the 3080 and that's still a significant step up for AMD, likely at significantly lower power consumption. They still have a month more for additional tweaking/driver maturity as well.

If we go by the increased pricing for Zen 3, AMD is not going to price it cheap. They will price it according to the performance.

The 3070 price has already been announced, doubt they can change one day ahead of retail.

Yea the 80CU 22TF number seems more accurate. But I suggest we don't delve further into the TF discussion as we all know how that ended last time.

 

6800XT targets 3080?

 

It would be cool if they would have a similar solution to DLSS, but we haven't heard about anything similar to tensor cores from AMD.

 

You can do quality upscaling without tensor cores I guess ?

 

I have a feeling they do, and not just me.

https://www.engadget.com/amd-radeon-6000-tease-165623408.html

https://www.pcworld.com/article/3585090/amd-radeon-rx-6000-big-navi-performance-tease-rtx-3080.html

https://hexus.net/tech/news/graphics/146008-amd-outs-radeon-rx-6000-performance-rtx-3080/

 

You can but then it's performance that is "stolen" from normal rendering while tensor cores are "free"?

 

You can? Could you give an example of shader based upsampling with DLSS IQ results?
I wouldn't be so sure about this one in the complete absence of such examples.

 

You can do everything you want in software in theory by using whatever cores are available to you. I said you can, I didn't say performance would be good or IQ the best.. why does it have to be shader based? Are CUDA cores being used solely for shader code? Obviously not as they are used for non graphics applications as well.

 

the numbers pulled from chiphell sure do suggest that this is on par with 3080
https://www.chiphell.com/forum.php?mod=viewthread&tid=2264233&mobile=1
5 days old leak btw

 

Shader based means algorithm based here. Compute shaders are shaders too. This is different from ML / neural network approach which doesn't have a clearly defined algorithm and "dream up" the results from training.
And while you're right that you can run anything on general SIMDs - even DLSS can be run there - an upscaling solution without performance benefits would be completely pointless.

 

Yeah, I see the point. I'm a broken record on this, but I wonder why an "improved" checkerboard system like ps4 pro can't be bring on PC...

 

It's a serialized step in the post-process chain, so I don't see why it would be "stolen". There's just a lot more theoretical TOPs with the amount of tensor cores packed on-chip.

 

Checkerboard is an MSAA based "hack" which is very h/w specific and won't work well across different GPU architectures and with all rendering approaches. It's a suitable option for a fixed h/w console platform but it won't work very well on PC. And many engines are in fact incompatible with it, requiring a lot of changes in both the code and art assets - which is why not all games are using CB on PS4Pro even.

 

Thx for the answer, it's very clear now.

And, have some of you experience a high quality video upscaler ? If so, can a "pure video" solution have a decent quality for video games ? I know it would be only working on the signal, but it would be free for devs... It's maybe super dumb, idk :/