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]

It also depends if there is a price segment between N33 and N32, that needs to be covered.

Don't expect N33 on desktop any time soon. Unless it is a bust on mobile. Like it's far enough away that it's a non factor in pricing of N32.

 

[Kaluan]

I said 300W as 'worst' case scenario (6750XT and 6800 are rated at the same TBP, so it's not without precedent), but I think seeing a similar 14 to 28% higher N22 to N21 clock delta between N32 and N31 is plausible.

So we'll probably see the "architected for 3GHz+" with Navi32 SKUs.

I don't think there is going to be a cut down desktop N32. They didn't do it for N22 until way way later. There's even less reason to do it because the N32 GCD is much smaller.

Only maybe sooner rather than later if N32 mobile doesn't sell great.

Well, no matter how good the yields are, they still need segmentation.

So it's just a question of time.

I know the Angstronomics leak said 3 and they were spot on with N31 but 3 SEs on N32 just does not make sense to me. It would mean just 96 rops in the 7800XT and 64 for the 7700XT which without a large clockspeed boost is a regression from 6800XT and 6700XT. Atleast with 4SEs the ROP count would be 128 and 96 respectively and then the a clockspeed boost is providing an improvement over prior gen.

Yeah so far Angstronomics' N32 leaks are the least convincing ones. Reminder that he also lead us to believe stacked MCDs will be a thing (future SKUs may still use them).

I just don't see shader engines going from 2048 ALU on N31 (N5) to 2560 ALU on N32 (N5) to 2048 again on N33 (N6). The scarcity of the texture units and raster units that would entail makes even less sense. Unless we're missing something.

I'll give 'em N33. But why would AMD have 3 different implementations of the same core architecture within the same product stack?

 

[Timorous]

The busted % is probally not that high. Mobile can take that volume.

I expect AMD will have more full N32 chips than they can sell as $700 7800XTs so they will be selling perfectly good dies as cut versions in both laptop and desktop skus.

You need to remember here that N32 is just 200mm². It is smaller than N23 and while 4 MCDs is around 150mm² the combined area for full N32 is comparable to N22 and far smaller than N21. This means margin in the $700 segment is a lot higher and margin in the $500 7700XT segment is probably similar to last year.

There is also the possibility that the 7800XT may also be a high bin for N32 meaning not all complete dies will pass.

 

[jpiniero]

I expect AMD will have more full N32 chips than they can sell as $700 7800XTs so they will be selling perfectly good dies as cut versions in both laptop and desktop skus.

As long as they forecast correctly, they shouldn't have that problem.

 

[Timorous]

As long as they forecast correctly, they shouldn't have that problem.

What else are they going to do with their rather large N5 supply? They have a part where they can match margin at the $500 level and make bonus margin at the $700 level.

AMD have enough supply to not be anywhere near as supply limited as they were with Zen3 and RDNA2.

 

[Rekluse]

What else are they going to do with their rather large N5 supply? They have a part where they can match margin at the $500 level and make bonus margin at the $700 level.

AMD have enough supply to not be anywhere near as supply limited as they were with Zen3 and RDNA2.

Is not the answer to this, MOAR SERVER CPU'S! Far better margin that GPU's after all

 

[maddie]

Is not the answer to this, MOAR SERVER CPU'S! Far better margin that GPU's after all

We're way past the time where the big companies need to prioritize high margin products before low margin ones. This by itself will cause average margins to fall and they're fighting hard to prevent this. They will fail as this is a reality that was interrupted by the recent troubles. The TAM includes lower margin products, that if produced, will lower your average margin.

The choice is to stay a small company only producing high margin products. Ultimately a failing strategy, as the bigger company, with a wider product line, and lower average margins, will have more revenue and R&D you into the ground, as they invade your high margin markets.

 

[TESKATLIPOKA]

I expect AMD will have more full N32 chips than they can sell as $700 7800XTs so they will be selling perfectly good dies as cut versions in both laptop and desktop skus.

You need to remember here that N32 is just 200mm². It is smaller than N23 and while 4 MCDs is around 150mm² the combined area for full N32 is comparable to N22 and far smaller than N21. This means margin in the $700 segment is a lot higher and margin in the $500 7700XT segment is probably similar to last year.

There is also the possibility that the 7800XT may also be a high bin for N32 meaning not all complete dies will pass.

N7: $8,000; N5: $16,000; Defect Density : 0.05 Link

	Dies per wafer	Good dies per wafer	Cost per chip	Cost per good chip
N21	98	76	$82	$105.3
N22	160	136	$50	$58.8
N23	233	208	$34.3	$38.5
N32	280	254	$57.1 + 4* $6.2 for MCD = $81.9	$63 + 4* $6.2 for MCD = $87.8
N32 Cutdown	280	254	$57.1 + 3* $6.2 for MCD = $75.7	$63 + 3* $6.2 for MCD = $81.6

The cost of N32 dies is a lot higher than N22, and It's not that much lower than N21. If we include the higher packaging cost, then they should be pretty close to each other.
BOM of the whole N32 card will be also pretty close to N21, considering It also uses 16GB Vram at higher speed, you will save a few $ on the cheaper cooler and weaker power delivery.
Full N32 will have better margin than RX6800($579) or RX6800XT($649) depending on the final price, but have worse margin than RX6900XT($999) or Rx6950XT($1099).
Cutdown N32 will need to have higher MSRP than RX6750XT($549) to end up more profitable.

 

[Rekluse]

We're way past the time where the big companies need to prioritize high margin products before low margin ones. This by itself will cause average margins to fall and they're fighting hard to prevent this. They will fail as this is a reality that was interrupted by the recent troubles. The TAM includes lower margin products, that if produced, will lower your average margin.

The choice is to stay a small company only producing high margin products. Ultimately a failing strategy, as the bigger company, with a wider product line, and lower average margins, will have more revenue and R&D you into the ground, as they invade your high margin markets.

True. While counterintuitive, I really think they ought to lean heavily on their low BOM and price as aggresively as they can just to penetrate the GPU Market deeper. It's going to require taking the performance crown from nVidia for AMD to shift past the 50% marketshare milestone. But pushing price/performance when they can (like now) should at least help them claw back some of that share.

 

[scineram]

We're way past the time where the big companies need to prioritize high margin products before low margin ones. This by itself will cause average margins to fall and they're fighting hard to prevent this. They will fail as this is a reality that was interrupted by the recent troubles. The TAM includes lower margin products, that if produced, will lower your average margin.

The choice is to stay a small company only producing high margin products. Ultimately a failing strategy, as the bigger company, with a wider product line, and lower average margins, will have more revenue and R&D you into the ground, as they invade your high margin markets.

It would be completely delusional trying to buy marketshare from a nonexistent consooomer base. There are hopeful sings the 4080 pricing was insanity. But that proves nothing about actual demand for Radeon graphics in any price segment.
Also such a competitor for AMD doesn't exist.

 

[Mopetar]

N32 pricing likely depends on whether whatever caused N31 to underperform was something AMD can and did fix with N32 or not. If it can clock closer to 3 GHz there's a good chance it could perform better at 1080p and even 1440p in a lot of titles where more shaders aren't as beneficial as faster shaders and there's no memory bottleneck.

NVidia will probably launch whatever they have for that slot first. We know they've got previously dubbed 4080 12 GB cards that will ultimately be pitted against N32, but what price they launch at is anyone's guess. I can't imagine them at anything above $800, but it would be weird for them to have a $400+ price gap in their product line.

 

[biostud]

I'm wondering how much performance/$ increase we will see from N32 vs the current 6800XT/6900XT prices. One thing is if they increase performance, but when you can get a 6800xt below $550 and a 6900xt below $670, they need to be priced right as well.

 

[eek2121]

A YouTuber claims to be benchmarking a 7900XTX and 4080, take with an absolute grain of salt:

EDIT: Clocks are too high. Probably fake.

 

[Hitman928]

A YouTuber claims to be benchmarking a 7900XTX and 4080, take with an absolute grain of salt:

EDIT: Clocks are too high. Probably fake.

Yep, just another YouTuber faking results for views.

 

[Timorous]

N7: $8,000; N5: $16,000; Defect Density : 0.05 Link

	Dies per wafer	Good dies per wafer	Cost per chip	Cost per good chip
N21	98	76	$82	$105.3
N22	160	136	$50	$58.8
N23	233	208	$34.3	$38.5
N32	280	254	$57.1 + 4* $6.2 for MCD = $81.9	$63 + 4* $6.2 for MCD = $87.8
N32 Cutdown	280	254	$57.1 + 3* $6.2 for MCD = $75.7	$63 + 3* $6.2 for MCD = $81.6

The cost of N32 dies is a lot higher than N22, and It's not that much lower than N21. If we include the higher packaging cost, then they should be pretty close to each other.
BOM of the whole N32 card will be also pretty close to N21, considering It also uses 16GB Vram at higher speed, you will save a few $ on the cheaper cooler and weaker power delivery.
Full N32 will have better margin than RX6800($579) or RX6800XT($649) depending on the final price, but have worse margin than RX6900XT($999) or Rx6950XT($1099).
Cutdown N32 will need to have higher MSRP than RX6750XT($549) to end up more profitable.

N6 was estimated around $10k and N5 $17K by Ian according to industry contacts. N7 is more expensive than N6 so $8k is far too low an estimate.

 

[TESKATLIPOKA]

N6 was estimated around $10k and N5 $17K by Ian according to industry contacts. N7 is more expensive than N6 so $8k is far too low an estimate.

I find $10K for N6 a bit too much especially in the current market situation.

I recalculated them with your prices for wafer.
N7: $11,000; N6: $10,000; N5: $17,000

	Dies per wafer	Good dies per wafer	Cost per chip	Cost per good chip
N21	98	76	$112.2	$144.7
N22	160	136	$68.8	$80.9
N23	233	208	$47.2	$52.9
MCD	1593	1564	$6.28	$7.03
N32	280	254	$60.7 + 4* $6.28 for MCD = $85.82	$66.9 + 4* $7.03 for MCD = $95.02
N32 Cutdown	280	254	$60.7 + 3* $6.28 for MCD = $79.54	$66.9 + 3* $7.03 for MCD = $87.99

 

[DisEnchantment]

AMD provided the estimates
16FF --> N7 : 2x
N7 --> N5 : 1.3-1.4x
But what do they know.

 

[The red spirit]

We're way past the time where the big companies need to prioritize high margin products before low margin ones. This by itself will cause average margins to fall and they're fighting hard to prevent this.

Not really, as long as there are people who pay, they will keep prices as high as they can. BTW 150k RTX 4090's have already been sold, so it's not like they don't sell.

 

[TESKATLIPOKA]

View attachment 71503
View attachment 71504
AMD provided the estimates
16FF --> N7 : 2x
N7 --> N5 : 1.3-1.4x
But what do they know.

If N5 cost pre mm2 is only 30-40% more than N7, then wouldn't It be better If even N33 was built using N5 instead of N6? Cost of the chip should be lower than or at least comparable to the one built on N6. Additional advantages would be higher clocks or lower power consumption.
I think N33 at N5 shouldn't be larger than 150mm2, I calculated ~103mm2 without memory controller or IC based on the table from Lacuza so 150mm2 is probably an overkill and size would be lower.

edit:
I calculated according to what Timorous said.

N6: $10,000; N5: $17,000

	Dies per wafer	Good dies per wafer	Cost per chip	Cost per good chip
N33: 200mm2 on N6	280	254	$35.71	$39.37
N33: 150mm2 on N5	380	352	$44.74	$48.3
N33: 140mm2 on N5	409	381	$41.56	$44.62

Price per chip would be only $9 higher.
Considering the advantage in higher clockspeed or lower power consumption which could be translated into increased price by $20 I do wonder If staying at N6 was the best bet. Maybe AMD feared they wouldn't have enough N5 wafer, who knows.

 

[maddie]

If N5 cost pre mm2 is only 30-40% more than N7, then wouldn't It be better If even N33 was built using N5 instead of N6? Cost of the chip should be lower than or at least comparable to the one built on N6. Additional advantages would be higher clocks or lower power consumption.

AMD is also claiming 2.65X transistor density in N31, at least for logic. Makes one wonder if all these "facts" can be true simultaneously.

 

[Kaluan]

N6 was estimated around $10k and N5 $17K by Ian according to industry contacts. N7 is more expensive than N6 so $8k is far too low an estimate.

Also who can attest these N7 prices were not even higher around October 2020 announcement and it's run-up, when AMD were still tabulating their margins and prices for RX 6000 SKUs.

I doubt node price, yield rate or general BOM cost for a 2020 product then vs now are the same.

Edit: If we're talking about margins on a historical scale, then maybe factor history in it as well. We don't live in 2020 anymore, but AMD's RX 6000 launch margins did.

 

[TESKATLIPOKA]

AMD is also claiming 2.65X transistor density in N31, at least for logic. Makes one wonder if all these "facts" can be true simultaneously.

That was only for WGPs. That increase is big, I have to wonder If they didn't use libraries for density instead of performance.

 

[Timorous]

If N5 cost pre mm2 is only 30-40% more than N7, then wouldn't It be better If even N33 was built using N5 instead of N6? Cost of the chip should be lower than or at least comparable to the one built on N6. Additional advantages would be higher clocks or lower power consumption.
I think N33 at N5 shouldn't be larger than 150mm2, I calculated ~103mm2 without memory controller or IC based on the table from Lacuza so 150mm2 is probably an overkill and size would be lower.

edit:
I calculated according to what Timorous said.

N6: $10,000; N5: $17,000

	Dies per wafer	Good dies per wafer	Cost per chip	Cost per good chip
N33: 200mm2 on N6	280	254	$35.71	$39.37
N33: 150mm2 on N5	380	352	$44.74	$48.3
N33: 140mm2 on N5	409	381	$41.56	$44.62

Price per chip would be only $9 higher.
Considering the advantage in higher clockspeed or lower power consumption which could be translated into increased price by $20 I do wonder If staying at N6 was the best bet. Maybe AMD feared they wouldn't have enough N5 wafer, who knows.

Disadvantage is that all new products are on the same node fighting for capacity/ margin optimisation.

With N33 on N6 and N32 on N5 AMD can make a lot more products because AMD only have so many N5 wafers.

 

[jpiniero]

Also who can attest these N7 prices were not even higher around October 2020 announcement and it's run-up, when AMD were still tabulating their margins and prices for RX 6000 SKUs.

I doubt node price, yield rate or general BOM cost for a 2020 product then vs now are the same.

Edit: If we're talking about margins on a historical scale, then maybe factor history in it as well. We don't live in 2020 anymore, but AMD's RX 6000 launch margins did.

If anything, things are worse now due to the TSMC price hikes.

 

[maddie]

If anything, things are worse now due to the TSMC price hikes.

Put up your current hard info. What was said a few months ago is now in flux. If you believe that wafer orders have been cut and TSMC will still stick with higher prices, then I have a bridge to sell. Interested?