AMD RDNA3 Specifications Discussion Thread

[Digidi]

90% of people are stupid. They listen more to people which are good in wording, than to people which know the technical things but can not articulate.

If you ask me what is the major problem of this world, it is exactly this. We give people with superficialities (good looking, good wording) more trust then people which are genius in there field but can not good talk.

 

[LordEC911]

Agree with your assessment and at least to me, they way he was presenting and answering questions is much more preferred to more marketing oriented people.

PS. That disc ... I can claim all the letters, depending on the day

Well that is one of the things that is always stressed to us, everyone has all those "abilities/traits" in them and depends on the exact scenario, but over the course of ~100 questions, you generally show your base 1or2 traits.
Like for example, you were a lead on a project and had to be more of a "D" but then on a different project you had to be more adaptable so you were a "S".

 

[techuse]

Sam doesn't look a good presenter, in fact he looks a bit bored ? withdrawn ?

He probably is a bit bored having to give extremely simplified answers non-engineers can understand while also having to tread carefully not to reveal anything he shouldn't. I much prefer him over marketing execs who don't actually know anything about the hardware.

 

[LordEC911]

What also shines through is his passion for his work.
Plus, we realize what a rare opportunity this is, to meet/talk with someone that was a major driving force inside AMD.

We have had some talks/interviews with Keller over the years, which were pretty interesting.
We also got to hear about Dave Baumann pushing RV770 after being there for ~3months.
We got to know what Carrell Killebrew did with the small-die strategy and his fight for RV870, and making Eyefinity essentially in secret.
We know Rick Bergman was there heading the group and making compromises throughout this time.
Then we got all those interviews with Raja Koduri after he came back in 2013.
I know there are probably dozens of others that we have heard from time to time that I can't think of...

We also know that ATI/AMD in the mid/late 2000's was a totally different company than 2015/2016 AMD.

Sam Naffziger goes down in history with those legends.

I've always wondered about who got AMD to go down the Bulldozer path.
That person saw a specific vision of where the industry(software) was headed, got AMD to follow his vision, and ended up wrong.

 

[trinibwoy]

…not good talk.

I think you mean “not speak well”

 

[Rootax]

Good ITW but, the main question imo is still not asked : what about compute chiplet.

 

[LordEC911]

Good ITW but, the main question imo is still not asked : what about compute chiplet.

??? You mean multiple GCDs? Or their CDNA products?
If you mean GCDs, it was touched on throughout the entire interview... and was partially answered in the first minute of the video then again at ~9:30 and ~10:50.

 

[del42sa]

Good ITW but, the main question imo is still not asked : what about compute chiplet.

I think is´t still a long way to go ...

 

[Qesa]

They said in the interview there were too many wires between shader engines for it to be feasible.

Though this was with InFO as the baseline I believe, a silicon interposer would allow for more bandwidth but at the higher cost which could also make it not worthwhile

 

[yuri]

I've always wondered about who got AMD to go down the Bulldozer path.
That person saw a specific vision of where the industry(software) was headed, got AMD to follow his vision, and ended up wrong.

A full Bulldozer post mortem would be a dream - covering the inception from Andy Glew's times (2004), through Charles R. Moore CBMT times with targeting 45nm, the release on 32nm under Mike Butler, and finally the highperf roadmap cancellation with focus on stop-gaps until Zen.

There is a 2011 paper called Bulldozer: An Approach to Multithreaded Compute Performance which describes the motivation/vision for BD. There are multiple authors (one of them is a key figure of Zen 3 development) but no real back story.

 

[Bondrewd]

a silicon interposer would allow for more bandwidth

Nope, that InFO_OS already runs class leading bump/wire pitch.
The only way up is SoIC.

 

[Rootax]

??? You mean multiple GCDs? Or their CDNA products?
If you mean GCDs, it was touched on throughout the entire interview... and was partially answered in the first minute of the video then again at ~9:30 and ~10:50.

They said in the interview there were too many wires between shader engines for it to be feasible.

Though this was with InFO as the baseline I believe, a silicon interposer would allow for more bandwidth but at the higher cost which could also make it not worthwhile

Well yeah but then find a solution, or another type of interposer, etc. I mean, few months ago, the big deal in the rumors was "oh, multiples compute dies, let's go !" In the end, well, I see the advantage of what they've done but, they can't really do more, it's doesn't scale that much after rdna3, it's not comparable to zen.

So, I guess the "solution" is multiple small gpus, and we're back at square one when the software thing seems hell for gaming. But this "chiplet" hype for rdna3 is a nothing burger now imo.

 

[techuse]

It’s a win for AMD margins but there is no benefit for consumers. I don’t see why they are using it as a talking point for marketing. They clearly aren’t passing on any of the price savings to us.

 

[Granath]

I wonder why they didn't go 500 mm sq. GCD. That would be a real justification for chiplets.

 

[Putas]

I wonder why they didn't go 500 mm sq. GCD. That would be a real justification for chiplets.

Minimizing risks might be one reason. The clocks of the current thing impressed no one and the other vendor tried to innovate power delivery just to feed its big chip.

 

[Kaotik]

It’s a win for AMD margins but there is no benefit for consumers. I don’t see why they are using it as a talking point for marketing. They clearly aren’t passing on any of the price savings to us.

Pretty much sure they did price it lower than competition, had they gone monolithic who's to say they could have priced it where they did?

 

[eastmen]

It’s a win for AMD margins but there is no benefit for consumers. I don’t see why they are using it as a talking point for marketing. They clearly aren’t passing on any of the price savings to us.

It's evocative of AMD's turn around in the processor market . Also lets remember at the $900 / $1000 price point AMD is offering 20/24 gigs of graphics ram. It's competitor was only going to offer 12 gigs at the $900 price point and 16 gigs at the $1200 price point

I wonder why they didn't go 500 mm sq. GCD. That would be a real justification for chiplets.

Could still be coming for all we know. AMD could announce a larger chip early in the new year to fill that huge gap between the $1k price and $1500 price

 

[Inuhanyou]

Pretty much sure they did price it lower than competition, had they gone monolithic who's to say they could have priced it where they did?

More ram and cheaper price seems pretty good to me if that's a benefit of chiplet. And it also is easier to produce and manufacture? Seems like a net win

 

[Bondrewd]

Could still be coming for all we know

Nope.

AMD could announce a larger chip early in the new yea

It's very very dead.
You're probably getting something else instead.

 

[PSman1700]

Why would AMD be announcing a even larger chip anyways? Raster performance isnt something they have to improve on so much, maybe clock speeds are enough there.