AMD RDNA graphics architecture whitepaper reveals the Radeon RX 5700's

https://www.amd.com/en/technologies/rdna

Here is the whitepaper --> PDF

I’ve been intrigued by the new architecture since we saw it ina slide a couple years ago as the “scalability” gpu. Now it makes sense to me. AMD was playing the long game for sure with this, and with them able to exceed 64CU’s with RDNA there will be some serious competition on the horizon.

My thoughts on scalability was always on how AMD could go beyond 64CU, but RDNA being able to go as low as smart phones and 5W power to scalable systems for cloud gaming. I’m interested in seeing the “Big” Navi now. (72CU’s ??? maybe 80CU’s???)

2 Likes

Wow I didn’t realize how tiny the chip is
2080 Ti 775 mm^2
2080 545 mm^2
Vega 64 495 mm^2
1080 Ti 471 mm^2
2060 445 mm^2
1080 314 mm^2
5700 251 mm^2
1060 200 mm^2

3 Likes

510mm^2 for a 60 or 64CU chip… Theoretically would be interesting to see how it all comes together ?

1 Like

Well, small scalable chips worked out for CPU’s all right, as per recent news. I wouldn’t be surprised if this was the case for GPU’s too

1 Like

Don’t get me wrong here, I love AMD as much as the next guy (maybe even a little more) but navi is still using a decent amount of power and making a decent amount of heat at the size it is now. Scale that up amd those go up too. Its is cool they can do it but nVidia definitely still are way out in front in certain ways.

Edit: Also the topic title makes no sense as it is incomplete. We all ready have the RX 5700, not a reveal of it in any way, unless you are just coming out of a coma.

Edit 2: seriously can some.with the privelages finish the title properly?

Would be slightly different to have say 2 gpu’s on a die and have the OS recognize it as one and function as such. Let’s see what comes of this because I have the sneaky suspicion that the scalability will come from CU’s instead of multiple dies.

True.
However, if you were to lower clocks (and therefor voltage a bit), TDP would reduce dramaticaly. AMD tends to be overly aggressive on voltage and clock, on the other hand, less performance than Nvidia would mean bad press.
Then you just do the ZEN-style paralelization (massive CU count) and done is a scaleable GPU for data center use.

I’mnot so convinced that Navi will go to the data center… If anything a new implementation of GCN will come out specifically for that.

We had a small thread talking about something

1 Like

ah, interesting !

Regarding Navi’s compute /data center ambitions.
There are some interesting details pertaining to deep learning / neural networks / inference.

https://www.amd.com/system/files/documents/rdna-whitepaper.pdf

“Additional Multi-Precision Ops for Some “Navi” Variants”

AMD added a new mixed precision FMA unit to the lanes including various 8bit /4bit dot operation support
To my understanding these features are not present or enabled on current Navi 10 (5700 / XT) GPU’s

3 Likes

@catsay Do you think that AMD is laying the groundwork for a more heterogeneous architecture ( probably for their next gen APU’s ) ? They have attempted a form of HSA some time ago, but with their platform change, this could be a good time to strike while the iron is hot.

new trancendental execution units to accelerate more complicated math operations that are used in both graphics and general computing…