Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[In2Photos]

According to Hardware Canucks they asked AMD about competing with the 4090 and AMD's response was that they are not trying to compete with it and can't compete with it right now.

Around the 5:25 mark.

 

[lobz]

-Doesnt even matter if it's the future or not. If NV matches or beats them everywhere then they need to take it seriously because they'll only ever be the second choice for the vast majority of people this way.

If I can get a card for an extra $100 that performs exactly the same in raster and 1.5x as fast in RT, and I'm already spending $1000... then I might as well get the one with more RT whether it's the future or not.

And might as well lose a couple times more of that in the power consumption difference in a year but hey, who cares about power in the middle of an energy crisis inducing actual war in Europe?

 

[igor_kavinski]

I want 4 slot monsters to end with the 4090. No more. Please.

 

[DAPUNISHER]

I want 4 slot monsters to end with the 4090. No more. Please.

You are looking at it all wrong bruh. It is better than 4 way SLI. And works properly in every game.

 

[Asterox]

A short summary of yesterday's or today's shocking events. The W difference should be highlighted more often.

RDNA3, 355W

RTX 4090, 450W

Huh, the graphics card is incomparably smaller compared to the monstrous Nvidia RTX 4090.

 

[Aapje]

Initial thoughts:
$1000 and no DDR6X?

I think that Nvidia has an exclusive contract for GDDR6X, which requires them to buy up all the stock. This is also why they are apparently now making a 3060 with GDDR6X, because they are required to use more of it than they would like to.

I wouldn't be surprised if the 4070 and perhaps even 4060 Ti get it as well, just to get rid of it.

 

[JayMX]

my guess is that they plan to glue two sets of those chiplet GCDs together at some point and create a serious frankengpu in the next 2 years.

They 've done it on MI200 (https://www.amd.com/en/graphics/instinct-server-accelerators) and I think that's their great advantage (scalability).

 

[DAPUNISHER]

Some vidoes from techtubers:

I am going to take it there was some real blowback. That, or more likely, they got the views they were looking for by throwing out the red meat title. Then changed the title to something more professional when the spike subsided. The latest video is much more subdued and positive. Maybe Steve needed the sleep and was crabby before? LOL

 

[DisEnchantment]

Jesus man, MLID blaming viewers and fanboys for misconstruing his leaks on RDNA3
Now he is leaking RDNA4 already.
Good lord, I pity the poor souls relishing those RDNA4 leaks.

 

[biostud]

Initial thoughts:
$1000 and no DDR6X?

Being as they've gone to all that effort with the separate cache chiplets I would have expected more cache? I mean you gotta assume on die would be faster and simpler as it's on the same chip, so if you go off die with all the extra logic and complexity then you've gotta be aiming for a higher quantity but this has less then the 6900XT?

The bandwidth is vastly increased, but maybe the will pull a vcache moment for the refresh model next autumn/winter.

 

[exquisitechar]

The bandwidth is vastly increased, but maybe the will pull a vcache moment for the refresh model next autumn/winter.

Yeah, I don't think V-Cache would do much for N31 with that kind of bandwidth. A larger GCD probably could have achieved >4090 raster performance, but it looks like N31 is as good as it gets.

 

[RnR_au]

I want 4 slot monsters to end with the 4090. No more. Please.

Sorry to be the bearer of bad tidings...

https://twitter.com/x/status/1588449243404042240

In addition, there is this interesting report of RDNA 3 arch being able to reach 3Ghz...

At the 46:26 mark.

Will the AIBs be able to create 450+W sku's that can accelerate a fair way past the reference sku's?

Take salt for now I guess.

 

[Tigerick]

Jesus man, MLID blaming viewers and fanboys for misconstruing his leaks on RDNA3
Now he is leaking RDNA4 already.
Good lord, I pity the poor souls relishing those RDNA4 leaks.

He also looked down at future Geforce 5000 series which is still 2 years away.

By the time, GDDR7 should be standard and power requirement will be lower (GDDR6X is pretty power hungry than GDDR6). With some tweaking, I believe nVidia will maintain the lead unless AMD increases die size of RDNA4 Navi41 to increase performance of RT core.

Anyhow, happy to see how AMD set the prices for current gen. MLID, my ass

 

[DAPUNISHER]

Sorry to be the bearer of bad tidings...

https://twitter.com/x/status/1588449243404042240

In addition, there is this interesting report of RDNA 3 arch being able to reach 3Ghz...

At the 46:26 mark.

Will the AIBs be able to create 450+W sku's that can accelerate a fair way past the reference sku's?

Take salt for now I guess.

This is what some of us were expecting. AMD lets the partners take center stage and justify an upcharge for their version of the cards. The reference version is going to be popular in miniITX builds, me thinks.

 

[DAPUNISHER]

He also looked down at future Geforce 5000 series which is still 2 years away.

By the time, GDDR7 should be standard and power requirement will be lower (GDDR6X is pretty power hungry than GDDR6). With some tweaking, I believe nVidia will maintain the lead unless AMD increases die size of RDNA4 Navi41 to increase performance of RT core.

Anyhow, happy to see how AMD set the prices for current gen. MLID, my ass

Guy is almost like fan fiction.

I share your expectations that Nvidia isn't giving up the lead anytime soon. They have executed well for a long time now. The power adapter issues, will be like the 2080ti failures. It blows up and gets quite a bit of coverage, then is quickly forgotten. Especially after the damage control squad spends countless posts telling us how trivial it was.

 

[Mopetar]

AMD CPU + AMD GPU <<< Show your love for AMD AND Mother Earth. The greener alternative to Intel+Nvidia.

But I thought that NVidia was already the greener alternative to Team Red.

I'm currently building a new rig. Should be finished by next week so what I have in my current sig is outdated

If you're going with a 4090 then waiting on Zen 4D before building a new rig seems like a better move. It's likely to make a difference in some titles and the AM5 platform is going to have a longer lifespan.

 

[Timorous]

Sorry to be the bearer of bad tidings...

https://twitter.com/x/status/1588449243404042240

In addition, there is this interesting report of RDNA 3 arch being able to reach 3Ghz...

At the 46:26 mark.

Will the AIBs be able to create 450+W sku's that can accelerate a fair way past the reference sku's?

Take salt for now I guess.

The AIB 6950XTs were around 5% faster than the reference design and had more OC headroom on top to extend the gap to about 10% when both were OCd.

So there is potential but it will cost a lot of extra power.

So having a look at this, if the 7900XTX comes in at 54% faster than the 6950XT in this chart it gets 131 FPS. The 4080 being 20% faster than the 3090Ti comes in at 109 FPS. That gives the following

4090 cost / frame $11.11
4080 cost / frame $11.01
7900XTX cost / frame $7.63

That gives the 7900XTX a very large cost / frame advantage vs the 4000 series and vs any other current GPU even at current prices.

In RT using 0.46x scaling for the 4000 series (number from that review), 0.31x for the 7900XTX (Same as the 6950XT perf loss) and 0.42x for the 3090Ti we get the following.

4090 RT cost / frame $24.24
4080 RT cost / frame $24.00
7900XTX RT cost / frame $24.37
3090Ti RT cost / frame $28.95

This shows the RT cost / frame is essentially a tie for these products and the 3090Ti is a little more expensive at $1,100 and would need to be around $925 to match these cost / frames in RT.

Now there is no guarantee the 7900XTX will match the 0.31x scaling factor, it may actually do worse but there is also no guarantee the 4080 will match the 4090 scaling factor given how cut down it is.

So as stated already. Pricing is scaling with RT performance by the looks of it but you get a lot more raster performance for your money. Just of note the 4080 12GB @ 899 would also be around $24.30 per RT frame so I wonder what NV will do with that when they launch it as a 4070. Will they cut it a bit further? Will they release the exact same spec but at a lower price point? Interested to find out because it could offer the best RT bang for your buck if priced right.

 

[MrTeal]

Yeah, I don't think V-Cache would do much for N31 with that kind of bandwidth. A larger GCD probably could have achieved >4090 raster performance, but it looks like N31 is as good as it gets.

Perhaps not vcache, but one of the disadvantages of HBM has always been the extra cost associated with the complex packaging needed. With RDNA3 AMD has already taken those costs by moving the memory controllers and cache off the main die. Is there a technical reason why they couldn't put a 2 or 4 high stack of of HBM2e for 4GB VRAM on top of every MCD? It would seem to be a natural progression, and would help lower board costs and smaller PCBs would help cooling.

 

[Geddagod]

They 've done it on MI200 (https://www.amd.com/en/graphics/instinct-server-accelerators) and I think that's their great advantage (scalability).

It's not recognized as one GPU though. They are just two separate GPUs with a really fast interconnect.

 

[Heartbreaker]

It's not recognized as one GPU though. They are just two separate GPUs with a really fast interconnect.

Exactly. This is fine and routine for compute. It isn't for Gaming.

Memory Controller Cache chiplet was a great way for AMD to take advantage of chiplets, while sidestepping the multiple core die issue.

 

[Timorous]

Reading the Anandtech article and this stood out to me.

The Navi 3x GPU (AMD is not confirming the specific GPU name at this time) breaks new ground for AMD not only on the performance front, but in terms of its construction. For the first time from any of the big 3 GPU manufacturers, AMD is utilizing chiplets in the construction of the GPU.

Why would AMD hide the name? Do they have something bigger planned down the road?

 

[Kaluan]

This thread is getting unbearable to follow.

Anyway, V-Cache doesn't seem to be ready yet. Which is not a issue, it's good they're clearly still keeping some aces close to their chest for now.

IDK about you guys, but with these great prices, 7700XT at $450 or even lower sounds plausible next year now. Trolls can cry me a river that AMD is putting huge downward pressure on GPU prices. Oh how terrible.

 

[maddie]

Not much left in the OC tank for aibs apparently...

https://twitter.com/x/status/1588329053530251265

Sorry to be the bearer of bad tidings...

https://twitter.com/x/status/1588449243404042240

In addition, there is this interesting report of RDNA 3 arch being able to reach 3Ghz...

Will the AIBs be able to create 450+W sku's that can accelerate a fair way past the reference sku's?

Take salt for now I guess.

The clock speed miss is the strangest thing to me. Pretty much all claimed higher speeds. A what/if follows.

Coreteks says "only prelim drivers used" for 5% OC & he interprets it to mean the performance is maxed out, but are they power limited?

The 3+ slot coolers for a 5% OC seems like overkill and for companies existing on very tight margins, putting a 50% larger cooler is not smart, unless they truly are selling by mass.

The 4090 class coolers are either needed for maxed OC models or a shortage of either 4090 die or 4090 sales forcing them to use the coolers on the 7900XTX. Either possibility is intriguing.

Guess we'll see soon as info will be coming in volume shortly.

 

[exquisitechar]

IDK about you guys, but with these great prices, 7700XT at $450 or even lower sounds plausible next year now.

For Navi 32?

 

[insertcarehere]

Jesus man, MLID blaming viewers and fanboys for misconstruing his leaks on RDNA3
Now he is leaking RDNA4 already.
Good lord, I pity the poor souls relishing those RDNA4 leaks.

All of these attention-seakers who missed the mark hard on RDNA3 need to be shown with clown profiles for a good while...