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

 

[eek2121]

Well, it is vague, as RDNA 4 comes before 2024 - but there is a stupid arrow the as per AMD SOP. I posted this hoping someone with more up to date knowledge might chime in. If such information is out there. Perhaps we will need to wait till this year's summer FAD for a bit more info.

Marketing slide. RDNA4 is 2H 2024.

So will the 7950 xtx and xt bug fix re-spins or just oced?

Just OCed, possibly with faster memory. See: 6950xt vs 6900xt.

We will get x50 variants starting later this year (along with new NVIDIA offerings like the 4090ti), and RDNA4 late next year (along with NVIDIA 5xxx).

No silicon respin or anything. The “defective“ rumors remain exactly that. Rumors.

 

[linkgoron]

Marketing slide. RDNA4 is 2H 2024.

Just OCed, possibly with faster memory. See: 6950xt vs 6900xt.

We will get x50 variants starting later this year (along with new NVIDIA offerings like the 4090ti), and RDNA4 late next year (along with NVIDIA 5xxx).

No silicon respin or anything. The “defective“ rumors remain exactly that. Rumors.

After the successful RDNA1 and RDNA2 launches (and especially Zen) people prefer to believe that there are issues with the RDNA3 that can be fixed using a respin, instead of believing that the arch is just fundamentally bad (which is obviously worse).

 

[Mopetar]

6700 XT

4090

As others have already pointed out there are multiple bins for both Navi 22 and AD-102.

Navi 22 had six different parts across desktop and mobile and yields must have been pretty good on it since they didn't cut it below 90%.

AD-102 only has 2 bins being sold right now, and the 4090 is the cut down part that isn't being sold as a data center card. It's highly likely we see a 4090 Ti at some point in the future that's also based on this die and potentially an even more cut down part in the form of a 4080 Ti as well.

Since the 4080 is already so close to full due AD-103 that there's not a lot of room for improvement. That and using AD-102 dies with a partially disabled memory bus would let them sell a card with an extra 4 GB of VRAM. NVidia is probably stockpiling these.

 

[jpiniero]

As others have already pointed out there are multiple bins for both Navi 22 and AD-102.

But it shows that even if they do end up going ahead with N32, they don't have to do more than 1 desktop part.

Same could be said for N33 for that matter.

 

[Aapje]

But it shows that even if they do end up going ahead with N32, they don't have to do more than 1 desktop part.

Same could be said for N33 for that matter.

True, but it makes little sense for them to not have a 7600 and 7600 XT, and a 12 GB and 16 GB version of N32, on the desktop.

 

[Heartbreaker]

But it shows that even if they do end up going ahead with N32, they don't have to do more than 1 desktop part.

All previous history seems to indicate, is that they don't have to start with more than one part. Is there any example of a chip than NEVER had more than one desktp part?

Note: It's premature to use RTX 4000 series because they aren't done yet.

 

[PJVol]

Just curious, where does all this talk about the single n32 sku come from?
If it originates from the "pull request leak", then even assuming it was accurate the list itself doesn't indicate it in any way, rather suggest the set of to-be-supported GPUs towards the certain software release date.

 

[jpiniero]

True, but it makes little sense for them to not have a 7600 and 7600 XT

The two SKUs would just end up competing with each other.

I am still of the opinion that N32 is completely cancelled.

 

[Aapje]

The two SKUs would just end up competing with each other.

I am still of the opinion that N32 is completely cancelled.

Could be, but then there will definitely be multiple N33 cuts.

 

[Kepler_L2]

The two SKUs would just end up competing with each other.

I am still of the opinion that N32 is completely cancelled.

It isn't.

 

[jpiniero]

It isn't.

They would be crazy to release it, given the current market. Unless they are expecting Mining to return this year which is a stretch.

 

[eek2121]

After the successful RDNA1 and RDNA2 launches (and especially Zen) people prefer to believe that there are issues with the RDNA3 that can be fixed using a respin, instead of believing that the arch is just fundamentally bad (which is obviously worse).

It may very well be that the actual 7900xtx was destined to be a 2-chiplet card that was supposed to compete with NVIDIA’s best, but when the cut-off came and they couldn’t make it work, they changed things up.

AMD should think about a new model number scheme. Far too easy to want to compare 9s, 8s, and 7s. They also really need to stop NVIDIA from setting price points. Just set a point with a healthy margin and go with it. I bet AMD would have made a lot of folks happy if:

Rename 7900xtx -> 7800xt -> $899
Rename 7900xt -> 7700xt -> $599

That would not have hurt AMD at all, and would have made their competitor look bad, and I suspect that slightly lower margins wouldn’t have mattered since volume would be up.

Also, AMD is in dire need of a “cream-of-crop” first party designs. While their recent cards look great, they have had QC issues and don’t offer anything unique over competitors.

NVIDIA’s FE cards push the envelope in terms of size (how compact/small it is), noise, looks, and performance. No NVIDIA partner has matched the FE cards in all these metrics combined. AMD should do the same.

 

[Aapje]

Also, AMD is in dire need of a “cream-of-crop” first party designs. While their recent cards look great, they have had QC issues and don’t offer anything unique over competitors.

EVGA?

We've seen both Nvidia and AMD lock down the freedom of AIBs more and more, but AMD could decide to do the opposite. However, these companies now seem to want to prevent customers getting 'excessive' value for too cheap a price.

 

[Heartbreaker]

They would be crazy to release it, given the current market.

Why? It seems like if they have already done the expensive up front work, and having the masking layers, they'd be crazy not to release it.

Though, if they haven't done that work, I could see skipping it.

 

[jpiniero]

Why? It seems like if they have already done the expensive up front work, and having the masking layers, they'd be crazy not to release it.

Though, if they haven't done that work, I could see skipping it.

I doubt it would sell in the current environment at the price needed to make it work.

 

[maddie]

I doubt it would sell in the current environment at the price needed to make it work.

That's right. Voluntarily lower your TAM for all the R&D costs already spent, and simultaneously increase the cost of the models your produce (lower card #s for sharing said R&D).

This is how you compete.
Seriously now, lower margins is the future and I expect us to baby step our way down to new levels. Warren Buffett’s long-time business partner Charlie Munger, said, “Get used to making less”.

Concerning a 16GB card, either the 4060 or 7600. This equates to 4 additional 2GB memory chips and as we have seen recently,they're spot selling for <$4 per. Is $20 ($5 per) in memory bulk-buy and say <$10 additional PCB cost outrageous? Add a total margin of (66% summed, for all the distribution stages) and yes, $50 is very achievable. The real troubling issue, for Nvidia & AMD, is what will this do to customer perception of your higher end models and their desired MSRPs.

 

[jpiniero]

This is how you compete.
Seriously now, lower margins is the future and I expect us to baby step our way down to new levels.

Like how AMD's client business is currently unprofitable?

 

[maddie]

Like how AMD's client business is currently unprofitable?

As simply put as possible, for this debate. Profit = sales - costs. Costs will have to fall, with all the disruptions implied in that statement.

I never said it would be pretty. I expect AI tools and services are going to "employed" in workforce reductions on a massive scale. Keep your core top people. Much pain incoming.

Regarding your return to free money printing as a solution, I would advise reading more about the growing worldwide trend into non $US transactions between rapidly growing nations that are equivalent to the G7 even in nominal GDP, much less PPP levels. This is not our parents world. What do you think will happen in such a situation with unlimited money printing?

 

[Aapje]

With Moore's law slowing down, they can provide less value for customers and so one way or another, profits will go down. Probably because upgrade cycles will get longer.

Nvidia is actually being quite smart here by introducing more and more software features like DLSS 1/2/3. Conveniently, these all required upgrading to newer cards. Still, it's hard to imagine that this will make up for slower gains in transistors per $.

 

[Heartbreaker]

I doubt it would sell in the current environment at the price needed to make it work.

Shouldn't Navi 32 be cheaper to produce than Navi 21, to replace the 6800 XT.

Though it might be a bit problematic replacing the 6700 Xt.

 

[jpiniero]

Shouldn't Navi 32 be cheaper to produce than Navi 21, to replace the 6800 XT.

Though it might be a bit problematic replacing the 6700 Xt.

I assume N32 is decently more than N22 but less than N21. I would agree that N22's original MSRPs would be unrealistic for N32 given the current state of the market.

 

[PJVol]

that N22's original MSRPs would be unrealistic for N32 given the current state of the market.

Idk, IMO 7700xt and 7800xt (or 7800 and 7800XT) sku's fit nicely 6700xt's $480 and 6750xt's $550 msrp respectively.
The performance of both ~ in a 6800 - 6900xt range, and launch with the navi21 inventory exhaust.

 

[MrTeal]

It may very well be that the actual 7900xtx was destined to be a 2-chiplet card that was supposed to compete with NVIDIA’s best, but when the cut-off came and they couldn’t make it work, they changed things up.

AMD should think about a new model number scheme. Far too easy to want to compare 9s, 8s, and 7s. They also really need to stop NVIDIA from setting price points. Just set a point with a healthy margin and go with it. I bet AMD would have made a lot of folks happy if:

Rename 7900xtx -> 7800xt -> $899
Rename 7900xt -> 7700xt -> $599

That would not have hurt AMD at all, and would have made their competitor look bad, and I suspect that slightly lower margins wouldn’t have mattered since volume would be up.

A renamed 7900XT for $600 would have made everyone happy considering it's 20GB and 30-40% faster at 4k than the 4070 at the same price, but that's pretty unrealistic. Dropping the price 1/3rd from $900 to $600 is a little more than slightly lower margins.

 

[jpiniero]

AMD Radeon RX 7600 8GB graphics card spotted in Asian store - VideoCardz.com

Imminent launch of Radeon RX 7600 confirmed by retailer We received an anonymous tip with Radeon RX 7600 already on shelves. According to the photograph and the information that was provided, an undisclosed store in Asia has AMD Radeon RX 7600 graphics cards in stock. This does not mean that...

videocardz.com

Ooh, Videocardz seems to have gotten a picture of the Sapphire Pulse 7600. I'd say... maybe legit? Of note it would be the full design and not the cut. I guess if this is legit then the 7500 would be the cut model and the 7700 would be the 16 GB clamshell... if they go ahead with either of those.

 

[maddie]

Full die too, and purportedly $250,