Rumored specs of RDNA3 architecture has changed a little - up to 12288 Stream Processors

Published by

Click here to post a comment for Rumored specs of RDNA3 architecture has changed a little - up to 12288 Stream Processors on our message forum
#6014122
https://forums.guru3d.com/data/avatars/m/230/230258.jpg
One thing i know for certain... GPU makers got the taste of the hind part of the miners very well. So, hashrate is what the next-gen gpus gonna excel ... that is what know for sure
#6014125
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
AuerX:

What would a AMD GPU do with Tensor Hardware?
There is 0 indication of AMD planning to add anY matrix arrays of any kind from AMD or leakers , they may do it in future generations they might do it now .... We do not know , from the specs i see that for now they are probably going to brute force it I guess.
#6014140
https://forums.guru3d.com/data/avatars/m/248/248994.jpg
TimmyP:

It is an given that hardware based tensor mathematics acceleration in graphics is absolutely unavoidable in the future. AMD would be foolish not to introduce this for their third gen straight; even Intel has hardware XeSS.
Yeah, Intel has hardware XeSS. Nobody else has it because Intel refuses to release the graphics cards to the market. I'm sure Raja Koduri likes to play with it in the lab.
#6014151
https://forums.guru3d.com/data/avatars/m/260/260048.jpg
Too early to judge. With MCM design, they can come up with practically anything in the following months.
#6014158
https://forums.guru3d.com/data/avatars/m/248/248291.jpg
At this point, these leakers are just throwing radom guesses, until something is right.
#6014160
https://forums.guru3d.com/data/avatars/m/282/282473.jpg
it sounds reasonable, full navi21 die is +300w alone on 7nm,with 5nm they can reduce the power draw a fair bit,but still fitting three of those on a single card would require a miracle.Not to mention the die size.20% cu reduction on every die might have been required to stay under 500w and 800mm2,especially when cache sizes need to double and they occupy a tremendous amount of the die I'm still excited to see this card,even with 3x 4096 it will be crazy fast
#6014168
https://forums.guru3d.com/data/avatars/m/273/273838.jpg
Are we absolutely sure AMD is not going back to the "multiple SKUs per tier" they used to do? I mean, back in the old days of the HD series, there were HD 4850, HD 4870 etc on the HD 48xx series, and also HD 4750, HD 4770 etc on the HD 47xx series
#6014171
https://forums.guru3d.com/data/avatars/m/282/282473.jpg
Mpampis:

Are we absolutely sure AMD is not going back to the "multiple SKUs per tier" they used to do? I mean, back in the old days of the HD series, there were HD 4850, HD 4870 etc on the HD 48xx series, and also HD 4750, HD 4770 etc on the HD 47xx series
xx50 and xx70 are still there,they're xt and non-xt now
#6014182
https://forums.guru3d.com/data/avatars/m/260/260048.jpg
Mpampis:

Are we absolutely sure AMD is not going back to the "multiple SKUs per tier" they used to do? I mean, back in the old days of the HD series, there were HD 4850, HD 4870 etc on the HD 48xx series, and also HD 4750, HD 4770 etc on the HD 47xx series
MCM gives them the flexibility to come up with a practically indefinite amount of variants. I think, even with miners in mind, we won't see so many issues with supply as we did with past generations as yields should be higher.
#6014183
data/avatar/default/avatar08.webp
AuerX:

Maybe they just keep winning in rasterization, after all it has gotten them to were they are on the GPU front now.
You mean in second place in sales behind NVIDIA?
#6014191
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
Crazy Joe:

You mean in second place in sales behind NVIDIA?
Being faster can boost the sales for sure maybe even overtake nvidia.
#6014197
https://forums.guru3d.com/data/avatars/m/229/229509.jpg
Hopefully it'll have a reasonable TDP!
#6014217
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
Undying:

Being faster can boost the sales for sure maybe even overtake nvidia.
Boost sales sure they can ,overtake though ..... I don't know Nvidia outsold amd on the 2xx era where AMD had clearly the superior cards and cheaper on the same price category ! They sure sold more than other gens though a lot more. Meanwhile Nvidia has a vastly stronger mindshare today ..,. So while AMD can sure get from 18-20% all the way up to 40% I seriously doubt Nvidia will go bellow 60% maybe if Nvidia loose clearly for 2-3 gens then AMD could overtake em.
#6014224
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
Venix:

Boost sales sure they can ,overtake though ..... I don't know Nvidia outsold amd on the 2xx era where AMD had clearly the superior cards and cheaper on the same price category ! They sure sold more than other gens though a lot more. Meanwhile Nvidia has a vastly stronger mindshare today ..,. So while AMD can sure get from 18-20% all the way up to 40% I seriously doubt Nvidia will go bellow 60% maybe if Nvidia loose clearly for 2-3 gens then AMD could overtake em.
I agree, something like this needs to happen. Rdna1 was a start, rdna2 was amd comeback to high end, rdna3 fastest consumer gpu and mcm debut so rdna4 can finally overtake nvidia. In the meantime nvidia to not have a successful and power hungry 40 series and being one step behind with 50 series becouse we know it will be their first mcm gpu while amd will have second generation by then. Something like 6900xt only being fast as 2080ti when using rt. Interesting times ahead.
#6014232
data/avatar/default/avatar20.webp
Remember Tesselation? A couple of generations later both AMD and Nvidia were equal and now nobody cared anymore. Same thing is gonna happen with RT. If you're not lying to yourself, even on nvidia 3080 enabling RT wrecks FPS to unplayable levels at 4K. But you have DLSS to compensate. On AMD hardware enabling RT wrecks FPS even more but you don't have DLSS-like option to get some FPS back. But in games where RT is used to just add finishing touches, both can handle it at 1440p somewhat OK. At 4K even 3090 can't maintain playable FPS without DLSS. I usually have RT on OFF or low and have great FPS. --- Regarding leaks. They are fun to follow but in the end, specifics about shader configurations don't tell you anything about performance. Who would have guessed 6900/6800XT with only 256 bit bus + infinity cache would be so fast in raster?
#6014237
https://forums.guru3d.com/data/avatars/m/108/108389.jpg
ah rumor rumor
f6c127657c2ecf62696f499e7fea3546d7b8cec5418075fbe305af125b039e50.gif
#6014258
data/avatar/default/avatar19.webp
Venix:

There is 0 indication of AMD planning to add anY matrix arrays of any kind from AMD or leakers , they may do it in future generations they might do it now .... We do not know , from the specs i see that for now they are probably going to brute force it I guess.
Lol did you actually just try and google tensor math and come here thinking you know even a fraction of a percent of what is going on? Jesus. Brute force tensor math? Haha. Oh my god.
#6014261
https://forums.guru3d.com/data/avatars/m/275/275921.jpg
Crazy Joe:

You mean in second place in sales behind NVIDIA?
Of course, what else would I mean?
#6014269
data/avatar/default/avatar06.webp
alpha007org:

Remember Tesselation? A couple of generations later both AMD and Nvidia were equal and now nobody cared anymore. Same thing is gonna happen with RT. If you're not lying to yourself, even on nvidia 3080 enabling RT wrecks FPS to unplayable levels at 4K. But you have DLSS to compensate. On AMD hardware enabling RT wrecks FPS even more but you don't have DLSS-like option to get some FPS back. But in games where RT is used to just add finishing touches, both can handle it at 1440p somewhat OK. At 4K even 3090 can't maintain playable FPS without DLSS. I usually have RT on OFF or low and have great FPS. --- Regarding leaks. They are fun to follow but in the end, specifics about shader configurations don't tell you anything about performance. Who would have guessed 6900/6800XT with only 256 bit bus + infinity cache would be so fast in raster?
AMD not including tensor hardware would be more akin to when ATI explicitly elected not to support shader model 3 in the x800 series, and drove away 50+ percent of their consumer base. (*I was one of them) *Apparently they DO have the hardware already on CDNA, they just have to bring it to RDNA. They are called "Matrix Core Engines."
#6014282
https://forums.guru3d.com/data/avatars/m/282/282473.jpg
TimmyP:

Lol did you actually just try and google tensor math and come here thinking you know even a fraction of a percent of what is going on? Jesus. Brute force tensor math? Haha. Oh my god.
you can run ML reconstruction in shahders,in fact afaik the early dlss version of control did
Of course, this isn't the first DLSS implementation we've seen in Control. The game shipped with a decent enough rendition of the technology that didn't actually use the machine learning Tensor core component of the Nvidia Turing architecture, relying on the standard CUDA cores instead. It still provided a huge performance boost, and generally looked better the higher up the resolution chain you went, but the new 2.0 revision offers a profound improvement.
https://www.eurogamer.net/digitalfoundry-2020-control-dlss-2-dot-zero-analysis

Previous Page

Next Page