AMD: RDNA 3 Speculation, Rumours and Discussion

Not even a HotChips keynote.
AMD are primordial shitlords uhuhuhuhuh.

 

And now the same leaker says there's going to be two multi die... dies?

Fuck it, I severely doubt this leaker now. There's little reason to tape out a separate die a third smaller than your big die, when you can just have a GPU with one of those modular dies instead of two. Heck you go a bit further into binning and the multi die setup is the relatively low power, low clocked bin while the single die bins are the ones you can just amp the clock up a ton. Considering salvage gets you another 2 separate configs at least, making another MCD setup with an entire other die makes severely little sense. There's no hole in your lineup to fill.

If any leak is true it's from the reliable Red Tech gaming. Two 6900s on the same package, probably with better raytracing. Doesn't require any unbelievable leap in power efficiency, doesn't require the practical implementation of SRAM to shrink dramatically between generations for good die sizes and yields, and it even tracks with the relative compute power between CDNA and RDNA. CDNA was more power efficient for compute than RDNA, having RDNA both catch all the way up and go multi die at the same time, without CDNA making the same relative advance, doesn't ring true.

 

Yes.

You really really need to understand what the fuck are they doing with the packaging there.

MCDs are now what you think they are.

It's an entirely different uArch that bears not much resemblance to RDNA1/2.

Chief MI100 to MI200 is 15.5-ish to >42TF DPFP per 500W OAM.

 

I've been wondering this several pages back.
Apparently it's to be "faster", not cost less.
So we all should prepare for some records, if only in pricing I feel.

 

AMD is a business, they'll build what sells, not for bragging rights. I think your original notion was right, this doesn't make business sense. Heck "to be faster" doesn't make any sense with two chiplet dies (that's better). They already have a "faster" one, it's already the fastest, the one that's big there's no point.

Hell if you wanted to cut a third that badly, you could just disable an SE on each one of the big dies and go further salvage. You know, almost exactly what they do with navi 21 and the 6800 non xt.

 

Oh man.
AMD runs an entirely redundant client HEDT lineup for that very purpose.

dGPUs don't make business sense from an AMD POV at all.
They still do them.

 

... Thanks to arguing without any real specification whatsoever. "Yes" that makes no sense. Can we get back to actually spelling out any reasoning whatsoever? Like, I went through and laid out the entire reason the business case is silly. "They're not what you think they are" is your response. Gee thanks.

This forum is getting full of this. What do you "think" MCDs are? Are the somehow heterogenous dies, despite AMD already having homogenous dies for CPUs and that making the most business sense, and even if it is that doesn't address why there isn't a further salvage die. Look, see, detailed response, trying to figure things out instead of arguing for the sake of argument. Instead there's stuff like the ridiculous notion that a Fortune 500 company climbing up in market cap wastes money because it feels like it.

 

Of course it does.
A tapeout is a tapeout!

What even.
From that POV anything besides some value part is a silly business case for AMD.
Like what's the point of N21 besides looking fancy?
It sold jack shit.
It exists to make a statement.

Read the goddamn thread for once.
Those things are small and thus facilitate the reuse that way.

Not anymore!
Zen4 has >2 CCDs and multitude of N6 IODs.
APU tapeouts? I don't even bother counting anymore...

N O T A N Y M O R E.
AMD's a big fish now.
And big fish needs to address discrete markets discretely.

Of course there will be chopped GCD configs for N31 and 32 both.

 

So what is the expected performance gain for the GPU tier at the 4-500$ pricepoint? More realistic 30-50%?

 

Stop turning every thread into a Sales or Market Share Topic.

 

Yea but your VRAM gets cut.

 

Less than the 16/8 they already have in that range?

 

But 450 buck range is 12GB now.
Gets down to 8 next year unless somehow lucky and JEDEC updates G6 spec and all.

 

That’s right, i was misremembering NV GPUs as the only ones with 12. But ya 8Gb in a 2022 4-500$ GPU is not good

 

I would also count the initial Threadripper in the "Let's do this, it's fun" territory. And honestly, I think it is nice to see (sometimes unchecked) engineering spirit at work ... not the bean counters or propaganda ministers.

 

There was not much engineering to be done on those, though, to thoroughly kick Intels behind.

 

Original Threadripper was supposedly engineers pet project they did on their free time

 

400 for a 25tf GPU? If this happens, my worries about an expensive high end only future would not be justified. Looks like a very good offer, even if 8GB requires to tone down settings.
But i assume it will be much more expensive in the end.
I hope for future RDNA2 / Ampere refreshes to solve this...

 

It can be 12GB too; we've qualified 24Gb DDR5/LP5 parts

Not happening really.

 

That really lets your enthusiast heart relate and bond, doesn't it?

For gamers 8GByte on a new card would not be desireable for 2022 and onward, given console have larger memories and it tends to get utilized there to good effect. Unless of course someone's aiming at the high-fps e-sports crowd exclusively. Which might not be the worst idea for brand-building.