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

 

[jpiniero]

This is the only N33 laptop I see on Newegg. It's a third party seller too and it says Shipped from Canada. I wonder what you would actually get if you bought it.

You can't even use the Power Search to find N33 laptops.

https://www.newegg.com/off-black-asus-tuf-gaming-fa617ns-ds71-ca/p/N82E16834236351

 

[gdansk]

I am very skeptical of a significant re-spin.
Energy/money better spent on RDNA4. Market conditions aren't good and it was known even at the end of the last year when such a decision would have been made. Why spend more to fix a product that won't sell much even if it wasn't a dud.

 

[Mopetar]

Unless RDNA4 was planned to be developed and ship not long after RDNA3, AMD is leaving a huge hole in their product line by scrapping N32.

Depending on how much the architectures have in common the fixes they would need for RDNA3 would also be needed for RDNA4 as well.

 

[jpiniero]

Unless RDNA4 was planned to be developed and ship not long after RDNA3, AMD is leaving a huge hole in their product line by scrapping N32.

Depending on how much the architectures have in common the fixes they would need for RDNA3 would also be needed for RDNA4 as well.

They have a product. It's unloading the remaining stock of RDNA 2.

 

[Ajay]

I am very skeptical of a significant re-spin.
Energy/money better spent on RDNA4. Market conditions aren't good and it was known even at the end of the last year when such a decision would have been made. Why spend more to fix a product that won't sell much even if it wasn't a dud.

Unless RDNA4 was planned to be developed and ship not long after RDNA3, AMD is leaving a huge hole in their product line by scrapping N32.

Depending on how much the architectures have in common the fixes they would need for RDNA3 would also be needed for RDNA4 as well.

Well, with the market conditions being poor, especially for AMD GPUs, they can ship or not ship RDNA3 and it may not matter much. 'Wasting' any addition resources on RDNA3 in a down market wouldn't make sense. That said, it's unlike AMD would skip a re-spin if they've already done the engineering for it. They can just keep clearing the channel, sell whatever they can make the most money on, and then ship RDNA4 as planned (FAD presentation Graph hints at end 2023/beginning 2024, but is vague as usual). Last note, the FAD presentation was around 1/3rd RDNA and 2/3rds CDNA - so it's clear that AMD's high performance commercial offerings are really where it is at for them. Not that that is a surprise.

 

[PJVol]

Nothing, not even fake rumors.

Not sure fake or not, but MLID had some today. At least those regarding RDNA3 launch saga seemed noteworthy to me.

 

[linkgoron]

So disappointed with RNDA3 and AMD. After RDNA1 and RDNA2, they just bit more than they could chew with RDNA3. In theory using the chiplet strategy, it should have been simpler to release N32 and N33, but it's clearly that's not the case. Huge Vega vibes. Such a radical redesign + new node might have been too much, and sadly they didn't have a backup plan (RDNA2 doesn't count). N32 and N33 are so late it's a joke.

 

[jpiniero]

So disappointed with RNDA3 and AMD. After RDNA1 and RDNA2, they just bit more than they could chew with RDNA3. In theory using the chiplet strategy, it should have been simpler to release N32 and N33, but it's clearly that's not the case. Huge Vega vibes. Such a radical redesign + new node might have been too much, and sadly they didn't have a backup plan (RDNA2 doesn't count). N32 and N33 are so late it's a joke.

They could have released N32 (and N33 on desktop) by now... but there would have been no point since it would have been redundant with there being apparently plenty of RDNA2 still unsold.

 

[Ajay]

They could have released N32 (and N33 on desktop) by now... but there would have been no point since it would have been redundant with there being apparently plenty of RDNA2 still unsold.

Which is a pretty big screw up in and of itself. Surprised we aren't hearing of more heads rolling.

 

[Mopetar]

N32 and N33 will be out eventually when they're worth releasing. To paraphrase Miyamoto, a bad GPU is bad forever, but a delayed one may eventually be good. Hopefully the lessons learned here pay off with RDNA4.

 

[linkgoron]

They could have released N32 (and N33 on desktop) by now... but there would have been no point since it would have been redundant with there being apparently plenty of RDNA2 still unsold.

That's a joke. AMD are losing market share and mindshare while Nvidia has already released AD102, AD103, AD104 and will supposedly soon release AD106. It's obvious that RDNA3 is just bad, and they had to go back to the drawing board.

 

[biostud]

Is there really so many RDNA2 parts in stock, or are they also freshly made? If AMD and their board partners really had a lot of stock, they probably had a get together and agreed to sell out the stock before AMD could launch N32, otherwise AMD might find it hard to sell GPUs in large quantities in the future.

 

[Mopetar]

If they had to delay a replacement for 6+ months they might continue production on the older parts. Even if there were a lot of inventory, they'd probably still sell through it over a long stretch like that.

By this point the bin quality will have gone up so AMD could always sell at the same price but give the AIB partners higher quality chips that would go into their higher margin parts.

Navi22 at ~$300 is too much of a good deal right now for stock levels not to deplete before Navi32 drops.

 

[jpiniero]

The other issue is whether AMD would be able to get any laptop OEM N32 deals... and whether this theoretical 6 month later refresh would be a big enough improvement to change things if the answer was no.

 

[PJVol]

Is there really so many RDNA2 parts in stock, or are they also freshly made?

I bought mine in july with the chip dated 2214, whatever that means, so probably still many. But the question is if N32 was delayed long before N31 launch event.
Tbh, MLIDs rumoured "artifacts" raises another question, if they admit Rdna4 has the same issue and it was fixed, what the driver team was doing some 2-3 months then, if it obviously is the hardware bug or even the design flaw?

 

[moinmoin]

I bought mine in july with the chip dated 2214, whatever that means

If it's the same scheme as used with CPUs it should be year and week, so 14th week in 2022 which is early April 2022.

 

[PJVol]

If it's the same scheme as used with CPUs it should be year and week, so 14th week in 2022 which is early April 2022.

You know, I've been curious, and now it's an even bigger mystery, where the hell are all these miner's cards? ) Shouldn't there be a huge number of them floating around?

 

[fleshconsumed]

You know, I've been curious, and now it's an even bigger mystery, where the hell are all these miner's cards? ) Shouldn't there be a huge number of them floating around?

They're there, ebay is full of them, I see tons of listings of what are clearly mining cards. However, since apparently tons of people are still buying 40 series despite 50% gen on gen price hike, sellers have no reason to drastically lower RDNA2/30xx prices. Sad to see personally, but I don't set the pricing.

 

[linkgoron]

Tbh, MLIDs rumoured "artifacts" raises another question, if they admit Rdna4 has the same issue and it was fixed, what the driver team was doing some 2-3 months then, if it obviously is the hardware bug or even the design flaw?

Trying to find software mitigations to the hardware bug. It's very common.

 

[PJVol]

Trying to find software mitigations to the hardware bug. It's very common.

I assume you didn't watched his video, where the "source" stated the product was launched with the bug already mitigated, or to be precise the stall was introduced in the pipeline for this purpose.

 

[maddie]

I assume you didn't watched his video, where the "source" stated the product was launched with the bug already mitigated, or to be precise the stall was introduced in the pipeline for this purpose.

Didn't they also say that the performance reduction was greater than expected? If so, then they would be exploring better alternatives

 

[Aapje]

They're there, ebay is full of them, I see tons of listings of what are clearly mining cards. However, since apparently tons of people are still buying 40 series despite 50% gen on gen price hike, sellers have no reason to drastically lower RDNA2/30xx prices. Sad to see personally, but I don't set the pricing.

Your reasoning makes little sense. If people buy 4000 cards instead, sellers would have to lower prices to compete. Lack of demand means lower prices, not higher.

It's much more logical that demand for 2nd hand cards is quite high due to the high prices of the new gen, so sellers don't have to lower prices much to sell.

 

[Aapje]

I assume you didn't watched his video, where the "source" stated the product was launched with the bug already mitigated, or to be precise the stall was introduced in the pipeline for this purpose.

That clearly causes performance degradation, so the bug is still there and they just picked the lesser of two evils to make sure that the card at least functions correctly, if slower.

 

[fleshconsumed]

Your reasoning makes little sense. If people buy 4000 cards instead, sellers would have to lower prices to compete. Lack of demand means lower prices, not higher.

It's much more logical that demand for 2nd hand cards is quite high due to the high prices of the new gen, so sellers don't have to lower prices much to sell.

We're saying same thing. Pricing is determined in part by the next gen pricing, and it's absurdly high so unless the demand craters which it hasn't used card sellers have no reason to lower prices.

 

[jpiniero]

AMD unveils Radeon Pro W7900 48GB and Pro W7800 32GB RDNA3 workstation GPUs - VideoCardz.com

AMD Radeon Pro W7900 & W7800 announced The company confirms specs and pricing of Radeon PRO GPUs with RDNA3 architecture. AMD Radeon Pro W7900 features 96 Compute Units (full Navi 31) combined with 48GB of GDDR6 ECC memory. This is a flagship workstation solution based on the same GPU as the...

videocardz.com

AMD releasing RDNA3 Radeon Pros. The W7900 is basically a 7900 XTX with 48 GB of memory; the W7800 is a new spec that has 70 CUs and 256 bit so 32 GB of memory. Possibly comparable to a theoretical 7900 non-XT.