AMD: RDNA 3 Speculation, Rumours and Discussion

I do think it makes sense to reuse Navi 22 and Navi 23 at 6nm, if all Navi 3x are on their way to becoming such expensive beasts.
Even if there's no performance upgrade, at least they'll have something to put inside laptops with console-like performance and featureset, using a hopefully cheap node.

 

Yes, performance-wise it would make sense, but i.e. if you consider N33 and N22, we would have an higher performing part with a smaller memory bus and possibly (if N33 follows the same trend as N23) a lower number of PCI-E lines. And I don't know how much cheaper N6 is compared to N7. Yes, it has some less passes (but it also uses EUV AFAIK) and it saves 10% of area, but after that we are looking at chips bigger than 300mm^2 and 200 mm^2, respectively, and without any other cost saving features (smaller buses, lower power consumption) compared to current line. So it may be, but it makes a lot of confusion if true.

 

6nm process is EUV, 7nm isn't. It shoudn't be possible to create 6nm Navi 2x by using just a simple optical shrink. Does it make sense to use RDNA 2 architecture to make new desing once you have RDNA 3 building blocks? Is it really so much cheaper that the architectural advantages of RDNA 3 (perf./watt, bandwidth efficiency etc.) could be outweighed?

 

Irrelevant; they're fully DR compatible and the migration consists of pressing 3.5 EDA buttons.

Yes it is.
Literally a mask swap.

Yes but AMD works in a different way anyways.

 

I wonder wether is possible to implement Decision Feedback Equaliser in memory controller to provide better quality of eye diagram ?

 

But logic and SRAM etc. scales differently. So it has to be arranged so how.

 

AMD has previously used shrinks smaller properties as a way to use dark silicon as thermal barriers in Zen vs Zen+
https://www.anandtech.com/show/12625/amd-second-generation-ryzen-7-2700x-2700-ryzen-5-2600x-2600/2

(white square in 12LP image is just 14LPP size for reference)

 

https://www.hardwaretimes.com/nvidi...0-super-series-expected-by-end-of-2021-rumor/

 

Last I checked that's not RDNA2, except for the part of the article that's regurgitated existing rumours.

 

You know the drill.

 

DTCs are supposed to bring an improvement equal to a whole process node uplift, really only Apple has been all-in with it till now, wonder when we'll see it from AMD in higher power products.

 

Seems RDNA3 has quite a bunch of chips in the family
1030->1035
https://github.com/RadeonOpenCompute/ROCm-Device-Libs/tree/amd-stg-open/oclc/src

 

https://lists.freedesktop.org/archives/amd-gfx/attachments/20210909/a989ed35/attachment.obj
https://lists.freedesktop.org/archives/amd-gfx/2021-August/068375.html
New MPIO block added to amdgpu, sounds similar to the one in the Zen4 manual made public by some leaks
This may not be related to RDNA though, could be the MI300 stuff for dealing with CXL, IF3.x+

 

Those aren't RDNA3 parts, GFX1030 is Navi21 so they're probably all RDNA2 parts.

 

Indeed, I just had a brain fart

 

Could be upcoming RMB and the other APUs with RDNA2

 

Yes!

 

I thought RDNA3 will be released in Q1/2022 not Q4/2022?

 

I don't know who greymon is, but either him or @Bondrewd are wrong, then.

 

It's always been late 2022 since I first heard about it.