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]

The real cost is in the initial design and the validation of the initial mask set, which should already have been done. Replicating masks is not that expensive, although if they need a respin, that is more costly. But it still seems unlikely that they would scrap an entire design until it would actually sell worse than to keep selling Navi 2x.

But I agree that a node shrink of Navi 2x doesn't make much sense.

If AMD had done a shrink of Navi21 from N7 to N5 I think it would have worked out fine. The chip would have been significantly smaller and faster.

IMO AMD was a bit too ambitious with their chiplet design.

Both AMD and NVIDIA need to find a way to make smaller, faster, less expensive chips/GPUs, or at the very least, find a way to make the technology automatically scale.

I can’t imagine what a flagship GPU will cost 5 years from now.

 

[soresu]

IMO AMD was a bit too ambitious with their chiplet design.

On the contrary.

They didn't go full chiplet, it's still got a lot of room to separate various functions of the GPU into its own die.

Like with Zen 1/+ they had to start somewhere to better inform design of later generations.

 

[Heartbreaker]

IMO AMD was a bit too ambitious with their chiplet design.

IMO, AMD really did a smart, optimal first step for GPU chiplets. Too ambitions would have been dividing up compute.

The problem multi-chip compute GPUs was really going to be needing massive BW, and super low latency comparable to being on the same chip. If not you then need to resort to horrible SLI/Xfire software solutions.

But the Memory Controller/Cache chiplets sidestep that issue, since all the compute is still together, plus both memory controller and cache take a lot of space, and don't really respond much to process shrinks anymore so is ideal to put on a cheaper less advanced process node.

But it's mostly a cost savings move, so they don't get more performance out of it, and NVidia made a bigger leap this generation because they were getting a much bigger boost moving away from inferior Samsung Process node, and AMD a smaller upgrade from an already good TSMC process.

 

[PJVol]

Why would anyone believe N21 is cheaper to produce? We know the die sizes already.

yield*7nm-price/520 vs yield*5nm-price/200 + yield*6nm price/146?
~ 90 vs 85.
What about adding interconnect and advanced packaging cost premium?

 

[Aapje]

But it's mostly a cost savings move

If they don't pass a large part of that on to consumers, their cards become fairly unattractive.

I'm wondering what the 7800 XT and 7700 XT will end up costing after some months (rather than at first, since AMD will likely ask too much, as usual).

 

[Ajay]

Clearly N31 was intended to be a 4090+ competitor.

AMD said they were aiming for the expected RTX 4080 performance. Of course, they could have just said that after they realized they had come up short.

 

[Aapje]

AMD said they were aiming for the expected RTX 4080 performance. Of course, they could have just said that after they realized they had come up short.

It's quite plausible that they underestimated how much benefit Nvidia would have from switching to TSMC. I think that we all did.

But I don't think that RDNA3 is working as well as they expected.

 

[leoneazzurro]

yield*7nm-price/520 vs yield*5nm-price/200 + yield*6nm price/146?
~ 90 vs 85.
What about adding interconnect and advanced packaging cost premium?

Yield for 5nm are similar to those of 7nm according to TSMC and that means defect per cm^2.
Thus, going up in size cost is superlinear because same level of defects means killing more wafer area. And a bigger die uses more wafer edge area because you try to accomodate a square shape in a circular wafer. 6N should cost less than 7N because less EUV steps needed. So, all in all, it should be more 100/110 vs 85 for silicon only. OFC package costs more.

 

[gdansk]

Why would anyone believe N21 is cheaper to produce? We know the die sizes already.

Development costs and economic forecasts as usual. There is basically no point to finish development of N32 unless they are sure that (margin of N32 - margin of N21) * expected number of units > cost of finishing N32.

It's probably true given how similar N32 is to N31. The cost of finishing it should have been very low. But it isn't preordained when the design is such a dud as RDNA3, AMD is selling only about 10% of what Nvidia is selling and the economy is in a downturn. These factors conspire to limit number of units.

But these factors, to me, imply no respin. We're getting unfixed N32 because it was too far along to save any money by cancelling or nothing.

 

[Kaluan]

Smaller RDNA3 arriving in June is just sad. AMD still using RDNA2 kind of reminds me of AMD reusing Polaris for a few gens, while not having something very competitive in the high end because Vega was just barely competitive with the 1080. Sure, N31 is much more competitive with the 4080 than what Vega64 was with the 1080, but it's just bad. I hope that AMD can fix the supposed bugs that exist with N33 and N32 or a N31v2 (if it's not just that the architecture is fundamentally bad), but I'm doubtful. Hopefully RDNA4 will be better, but it's surely far away in the future.

Well, then there's the whole aspect of their oversupply of RDNA2 cards...
And right now they're selling very well, so I doubt AMD is hurting with this division too much.

That being said, still, I hope they also show something about N32/further cut N31 at Computex.

 

[jpiniero]

Well, then there's the whole aspect of their oversupply of RDNA2 cards...
And right now they're selling very well, so I doubt AMD is hurting with this division too much.

I dunno about very well... but way better than RDNA 3. The problem is that RDNA 2 is selling at prices AMD isn't happy about. Why make things worse by bringing out N32, esp if mobile OEMs aren't interested?

 

[Rigg]

It looks to me like stock of N21 is rapidly dwindling. PC partpicker only shows 7 of 61 (7/43 if you remove the 6900 XT) models of N21 cards available for purchase. My local MC echos this as well. 6 models of N21 cards available with only one showing 25+ stock. There were 2 cards with 25+ availability yesterday when I looked. The other card is now showing 21 in stock.

I was under the impression they have a bunch of 5nm capacity booked and would assume they need to make something besides Zen 4 and N31 to utilize it all. If that's the case all of this speculation about N32 being canceled seems kind of baseless.

 

[PJVol]

Yield for 5nm are similar to those of 7nm according to TSMC and that means defect per cm^2.
Thus, going up in size cost is superlinear because same level of defects means killing more wafer area. And a bigger die uses more wafer edge area because you try to accomodate a square shape in a circular wafer. 6N should cost less than 7N because less EUV steps needed. So, all in all, it should be more 100/110 vs 85 for silicon only. OFC package costs more.

Yeah, using this online calculator for simplicity

Silicon Cost Calculator

This calculator gives a rough estimate of the manufacturing cost of unpackaged semiconductor chips. Wafer prices aren’t something that is generally available, but doing an internet search of …

www.adapteva.com

and 17000 and 9350 for the 5nm and 7nm wafer price resp.,
which gave me $97 for N21 and $84 for N32 ( I assumed 6nm the same price as 7nm)
Still curious though, what extra the new interconnect and packaging adds to the cost.

 

[Timorous]

Yeah, using this online calculator for simplicity

Silicon Cost Calculator

This calculator gives a rough estimate of the manufacturing cost of unpackaged semiconductor chips. Wafer prices aren’t something that is generally available, but doing an internet search of …

www.adapteva.com

and 17000 and 9350 for the 5nm and 7nm wafer price resp.,
which gave me $97 for N21 and $84 for N32 ( I assumed 6nm the same price as 7nm)
Still curious though, what extra the new interconnect and packaging adds to the cost.

Where does 9350 come from for 7nm? I have seen 17k ballparked for 5n before but 7N was always around the 12k mark when Ian spoke with industry contacts and 6N was a bit cheaper since it goes through fewer process steps.

This means in the very best case laid out above N21 is 15% more expensive than N32 and I think with more realistic N7 and N6 wafer prices that gap widens further.

This is what Locuza came up with for N7 and N5 onwards.

 

[coercitiv]

This is what Locuza came up with for N7 and N5 onwards.

Take that image down immediately! You're endangering the narrative that N5 is twice as expensive as N7. /s

The $9350 per N7 wafer probably came from here:

 

[maddie]

I dunno about very well... but way better than RDNA 3. The problem is that RDNA 2 is selling at prices AMD isn't happy about. Why make things worse by bringing out N32, esp if mobile OEMs aren't interested?

You seem to know some inside information. What price will AMD be comfortable selling at, for the various cards? Also any firm info (not guess) on the margins at that price?

 

[jpiniero]

What price will AMD be comfortable selling at, for the various cards?

The MSRP?

Speaking of which, Amazon is selling an 7900 XT for $762 now.

 

[Kaluan]

I dunno about very well... but way better than RDNA 3. The problem is that RDNA 2 is selling at prices AMD isn't happy about. Why make things worse by bringing out N32, esp if mobile OEMs aren't interested?

Maybe, but I'm not sure selling virtually none and having continuous oversupply problems on their hands is preferable.

Currently RDNA3 and RDNA2 SKUs don't really overlap in price categories, so that point is moot. And anyway, RDNA3 is far from selling terrible, all things like market slowdown and AMD's market and mind share considered.

 

[dr1337]

If AMD had done a shrink of Navi21 from N7 to N5 I think it would have worked out fine. The chip would have been significantly smaller and faster.

IMO AMD was a bit too ambitious with their chiplet design.

I can't help but wonder if they would have been better off keeping n31 at 16gb/256bit just like n21. 2 less MCDs and 4 less RAM chips would have increased their margins and helped ease the price drops down the road. n31 as is doesn't seem completely bandwidth starved but perhaps cutting down to 64mb of cache might be a bit too thin for 4k performance. Seems kinda like AMD have simply boxed themselves in by having only 16mb of cache per chaplet instead of 24 or 32.

 

[Timorous]

What if N32 is actually good so now AMD are in the position where a fully enabled N32 that clocks well in the desired power envelope is actually very very close to a 7900XT already and it is a lot cheaper to make.

Could AMD be waiting for an N31 respin so they can launch 7800XT and 7950XT and 7950XTX all at once? 7900XT and 7900XTX can gain a good 10-15% from being OC'd and if AMD could respin N31 so those OC clocks can be hit at stock within the 300W and 355W power envelopes it would make for a pretty good refresh.

If N32 7800XT does come in at around 6950XT performance (and this is quite wide because you have TPU that use the reference 6950XT performance and you have HUB that use an AIB 6950XT and the difference between AIB 6950XT and reference 6950XT can be 10% or more) then they could charge $550 or something and still make far more margin than selling 6950XTs at $600.

 

[coercitiv]

What if N32 is actually good so now AMD are in the position where a fully enabled N32 that clocks well in the desired power envelope is actually very very close to a 7900XT already and it is a lot cheaper to make.

This makes little sense considering Nvidia is already selling new gen cards in the N32 price range. They would be protecting 7900XT sales while losing even more potential revenue with a successful N32 launch.

 

[Timorous]

This makes little sense considering Nvidia is already selling new gen cards in the N32 price range. They would be protecting 7900XT sales while losing even more potential revenue with a successful N32 launch.

They are offering a $600 4070 that has specs more befitting a 4060 that the market seems to be rejecting.

Currently AMD churning through the remaining N21 inventory is offering better perf/$ than NVs new apart and with 16GB of Vram is also offering something with more longevity.

 

[PJVol]

Currently RDNA3 and RDNA2 SKUs don't really overlap in price categories

If the rumored n33 performance is close to final, then RDNA 3 is really about performance per area, not $.
Perhaps in the current state of market it has little value to offer over RDNA 2.

 

[Aapje]

N33 seems to be a low-risk, low cost chip that is primarily intended to take a lot of laptop market share.

By using TSMC 6, they hedged their bets against huge TSMC 5 shortages (which turned out to not be an issue). TSMC 6 is also cheaper than both TSMC 7 and 5. By having a smaller die, they again save money.

On the desktop market I expect that N33 cards will be a little faster than N23 cards and a little cheaper. Nothing spectacular, but with the poor performance/$ improvement that Nvidia is bringing, it may be plenty competitive. And Nvidia may not bring out really cheap cards, so AMD may be the only one to have cards for gamers with a budget below $250.

 

[Timorous]

N33 seems to be a low-risk, low cost chip that is primarily intended to take a lot of laptop market share.

By using TSMC 6, they hedged their bets against huge TSMC 5 shortages (which turned out to not be an issue). TSMC 6 is also cheaper than both TSMC 7 and 5. By having a smaller die, they again save money.

On the desktop market I expect that N33 cards will be a little faster than N23 cards and a little cheaper. Nothing spectacular, but with the poor performance/$ improvement that Nvidia is bringing, it may be plenty competitive. And Nvidia may not bring out really cheap cards, so AMD may be the only one to have cards for gamers with a budget below $250.

That is my thinking and if AMD themselves or AIBs offer a 16GB variant it could be a good card to give you console like performance until the PS6 generation comes out of the cross gen phase.