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

 

[beginner99]

They just need to stop clinging onto the fantasy of pandemic margins and get into the mindset of growing marketshare, even if it isn't immediately that profitable to do so

fully agree especially if AMD has too many wafers allocated. Better to gain market share than pay TSMC a penalty.

That would mean AMD squandered 2 quarters of potentially good sales, mostly due to poor marketing decisions.

While I agree with you that if they have the wafers for a "price war" they should make use of that ASAP, planing of releases happens way, way earlier. AMD might have some bandwidth limitations in certain areas that make it hard to bring multiple products to the market at the same time. eg. N31 and N32 don't run fully parallel and at one point you must give one of the priority and that decision didn't happen last months but more likley a year ago where this whole inflation thing was om nobodies horizon yet.

One bottlebeck could be driver optimization, I^m sure they will vary a bit for each specific chip.

Phoenix has 6 WGPs so an 8WGP N34 would be kinda pointless.
Why would 8WGP, 16CU, 2048SP need 128 bit bus + 8x PCIe 4.0?
N33 also uses only 128 bit bus + 8x PCIe 4.0, yet It has 16WGP, 32CU, 4096SP. That doesn't make sense.
I also wonder If 8WGP N34 would really perform close to 16WGP N23(RX 6600XT).

my 2 cents are that an N34 would still make sense even if it is not that much faster than an APU. Simply because it can be put in PCs that do not have an APU. like somebody that does CPU heavy work and only does light gaming or shares the machines with kids but for them a low end gpu is good-enough and not willing to spend >$500 on it. (I admit given zen4 iGPU reduces need for low end gpus as well but still I think for upgrading older machines, budget machines and so forth it is still worth it. APUs in general haven't been perfect for budget. A dedicated combo used to give you 50%+ more for just $150 (before shortage)

 

[RnR_au]

So, imagine that Zen 4 V-Cache and N32 are released in March. That would mean AMD squandered 2 quarters of potentially good sales, mostly due to poor marketing decisions.

Zen 4 vcache is outside of AMD's control.

Edit: and I just saw this...

https://twitter.com/x/status/1583934983332249600

Don't understand how it can't be ready yet since N32 and N31 shares arch?

 

[TESKATLIPOKA]

my 2 cents are that an N34 would still make sense even if it is not that much faster than an APU. Simply because it can be put in PCs that do not have an APU. like somebody that does CPU heavy work and only does light gaming or shares the machines with kids but for them a low end gpu is good-enough and not willing to spend >$500 on it. (I admit given zen4 iGPU reduces need for low end gpus as well but still I think for upgrading older machines, budget machines and so forth it is still worth it. APUs in general haven't been perfect for budget. A dedicated combo used to give you 50%+ more for just $150 (before shortage)

You are actually right, N34 wouldn't be pointless. Phoenix will be great, but It will be limited to 8 cores. If you want to play a bit, but need more than 8 cores or have an older machine, then you would need a dedicated GPU. I would still prefer N34 with the specs I mentioned than one with only 8 WGPs in full config.

 

[TESKATLIPOKA]

Edit: and I just saw this...

https://twitter.com/x/status/1583934983332249600

Don't understand how it can't be ready yet since N32 and N31 shares arch?

It's ready, just the production didn't start in my opinion, and Kepler already mentioned that AMD still has a lot of N21 to sell.

 

[jpiniero]

The 60k wafers for Q4 that Graymon mentioned are N5 wafers.

If they started N32 production in Q4, March is a perfectly reasonable release date. You're probably underestimating production timelines.

 

[Timorous]

Ian put out an interesting video about 7950X production cost.

Key info here is the $17K / wafer estimate for 5nm and the $10k / wafer estimate for 6nm.

For Navi 2x the N22 die was roughly square and the N21 die was elongated. I will make the same assumption here which makes N32 around 14.2x14.2 and N31 around 14.2x 21.4.

With TSMCs 7nm that means N31 gets 149 good dies a wafer and 35 defects. N32 gets 255 good dies a wafer and 38 defects. With the MCDs having 1,615 good dies per wafer and 41 defects.

I we just use good dies it gives us a cost of $115 per N31 die and $67 per N32 die. Each MCD comes in at about $6.20 each.

So for N31 that would total $152.20 + packaging costs and for N32 it would be $91.80 + packaging costs. In the video Ian estimated packaging costs for the 7950X as around $7.50 so with 7 chips instead of 3 lets just go with a high end of $30 for packaging and then round up to the nearest $5 and that gives N31 a total cost of $185 and N32 a total cost of $115.

AD102 on the other hand comes in at 58 good dies per wafer which is $293.

AD103 comes in at 112 good dies for a cost per die of $152.

To me this shows that AMD would have the ability to price lower and make more margin that NV. I could easily see the 7900XT coming in at $1,200 ish while offering 4090 or very close to performance and still being a higher margin part than the 4080 16GB.

 

[H T C]

Ian put out an interesting video about 7950X production cost.

Key info here is the $17K / wafer estimate for 5nm and the $10k / wafer estimate for 6nm.

For Navi 2x the N22 die was roughly square and the N21 die was elongated. I will make the same assumption here which makes N32 around 14.2x14.2 and N31 around 14.2x 21.4.

With TSMCs 7nm that means N31 gets 149 good dies a wafer and 35 defects. N32 gets 255 good dies a wafer and 38 defects. With the MCDs having 1,615 good dies per wafer and 41 defects.

I we just use good dies it gives us a cost of $115 per N31 die and $67 per N32 die. Each MCD comes in at about $6.20 each.

So for N31 that would total $152.20 + packaging costs and for N32 it would be $91.80 + packaging costs. In the video Ian estimated packaging costs for the 7950X as around $7.50 so with 7 chips instead of 3 lets just go with a high end of $30 for packaging and then round up to the nearest $5 and that gives N31 a total cost of $185 and N32 a total cost of $115.

AD102 on the other hand comes in at 58 good dies per wafer which is $293.

AD103 comes in at 112 good dies for a cost per die of $152.

To me this shows that AMD would have the ability to price lower and make more margin that NV. I could easily see the 7900XT coming in at $1,200 ish while offering 4090 or very close to performance and still being a higher margin part than the 4080 16GB.

Question: does any of this take in to account the costs of the video card's PCBs?

I have NO IDEA how much it costs, for either company.

 

[beginner99]

Kepler already mentioned that AMD still has a lot of N21 to sell.

Maybe. But if that were the case, why are prices here in Europe still way too high? 6900xt if it even is in stock is still $1000, 6800 xt is $750. They could very easily sell atom of n21 here but not at these prices.

 

[Timorous]

Question: does any of this take in to account the costs of the video card's PCBs?

I have NO IDEA how much it costs, for either company.

No but a 450W TBP card with 24GB GDDR6X VRAM is going to be similar to another 450W TBP card with 24GB GDDR6 VRAM. If the AMD card comes in under then the BOM for the cooler, vrms etc will be lower and I think GDDR6X is more expensive than GDDR6 but not sure if that applies to the 24gbps GDDR6.

NV might get some economies of scale benefits because they order more parts but I think calling it a rough wash with parts in a similar TBP range and memory amounts is probably ballpark correct and we are only going to get ballpark numbers here.

 

[Timorous]

Maybe. But if that were the case, why are prices here in Europe still way too high? 6900xt if it even is in stock is still $1000, 6800 xt is $750. They could very easily sell atom of n21 here but not at these prices.

Strong $ makes it more profitable for AMD to sell parts in $ rather than other currencies hence the high price outside of USA and low price in the USA to pick up demand in the more profitable market.

 

[TESKATLIPOKA]

Ian put out an interesting video about 7950X production cost.

Key info here is the $17K / wafer estimate for 5nm and the $10k / wafer estimate for 6nm.

For Navi 2x the N22 die was roughly square and the N21 die was elongated. I will make the same assumption here which makes N32 around 14.2x14.2 and N31 around 14.2x 21.4.

With TSMCs 7nm that means N31 gets 149 good dies a wafer and 35 defects. N32 gets 255 good dies a wafer and 38 defects. With the MCDs having 1,615 good dies per wafer and 41 defects.

I we just use good dies it gives us a cost of $115 per N31 die and $67 per N32 die. Each MCD comes in at about $6.20 each.

So for N31 that would total $152.20 + packaging costs and for N32 it would be $91.80 + packaging costs. In the video Ian estimated packaging costs for the 7950X as around $7.50 so with 7 chips instead of 3 lets just go with a high end of $30 for packaging and then round up to the nearest $5 and that gives N31 a total cost of $185 and N32 a total cost of $115.

AD102 on the other hand comes in at 58 good dies per wafer which is $293.

AD103 comes in at 112 good dies for a cost per die of $152.

To me this shows that AMD would have the ability to price lower and make more margin that NV. I could easily see the 7900XT coming in at $1,200 ish while offering 4090 or very close to performance and still being a higher margin part than the 4080 16GB.

Difference in wafer cost + packaging won't be so high.
You calculate only the good dies, but not the defective ones, and most of them can be used for a cutdown version in my opinion.

	Die size [mm2]	Good dies	Defective dies	Price per wafer	Cost per die (good dies)	Cost per chip (all dies)
N31	308	146	34	$17,000	$117	$95
N32	200	255	38	$17,000	$67	$58
N33	203	241	36	$10,000	$41.5	$36
MCD	37.5	1565	43	$10,000	$6.4	$6.2
AD102	608.5	58	30	$17,000	$293	$194
AD103	378.6	112	33	$17,000	$152	$118
AD104	294.5	152	33	$17,000	$112	$92

In the "worst case", all dies can be used and sold. Here is the best and worst case:
N31: (95 or 117)+6*6.4+30 = ~$164-186
N32: (58 or 67)+4*6.4+20 = ~$104-113
Not sure how much would the packaging cost be, so I set It at $5 for single chip packaging.
N33: ~$41-46.5
Ad102: ~$199-298
Ad103: ~$123-157
Ad104: ~$97-117

If the full N31 cost the same as RTX 4080 16GB($1199), then AMD would have lower margin than Nvidia, because the chip costs more.

P.S. of course, I am excluding the cost of the whole card(memory,cooler,power delivery etc.).

 

[Yosar]

Strong $ makes it more profitable for AMD to sell parts in $ rather than other currencies hence the high price outside of USA and low price in the USA to pick up demand in the more profitable market.

Actually it works in the other way. It is only more profitable if AMD has costs in let's say EUR. And that's only if you don't take into consideration that those costs will be higher due to inflation also. And I doubt there are many costs they pay in EUR.
So strong $ means only one thing for AMD (or any corporation from US), their sales will tank in any market that currency was weakend in relation to $. Because it will be much more expensive there than it was.
And tanking sales means less profit not more.
Basically now most corporations from US has decisions to make. Will they pretend that 550$ for 7900X (the same as for 5900X) is good price in any market outside US.
As long as $ is strong it's not a good price. Of course it's not that simple, you cannot simply lower prices in let's say Europe just like that, because Americans will start to buy from Europe instead of American sellers.
And that's a big no for American corporations.

So it's quite quite delicate matter of balance. But if American corporations will act like $ is not strong (too strong for their good) their sales and profit will tank outside of US (tax for inflation + strong $ is too high there).

And AMD is only example. It concerns nVidia (I could buy here RTX 4090 no problem, no ques), intel, apple et cetera et cetera.

That's why China always tried to keep its currency artificially low in relation to $. It makes export/sales more profitable, no the other way.

 

[Timorous]

Difference in wafer cost + packaging won't be so high.
You calculate only the good dies, but not the defective ones, and most of them can be used for a cutdown version in my opinion.

	Die size [mm2]	Good dies	Defective dies	Price per wafer	Cost per die (good dies)	Cost per chip (all dies)
N31	308	146	34	$17,000	$117	$95
N32	200	255	38	$17,000	$67	$58
N33	203	241	36	$10,000	$41.5	$36
MCD	37.5	1565	43	$10,000	$6.4	$6.2
AD102	608.5	58	30	$17,000	$293	$194
AD103	378.6	112	33	$17,000	$152	$118
AD104	294.5	152	33	$17,000	$112	$92

In the "worst case", all dies can be used and sold. Here is the best and worst case:
N31: (95 or 117)+6*6.4+30 = ~$164-186
N32: (58 or 67)+4*6.4+20 = ~$104-113
Not sure how much would the packaging cost be, so I set It at $5 for single chip packaging.
N33: ~$41-46.5
Ad102: ~$199-298
Ad103: ~$123-157
Ad104: ~$97-117

If the full N31 cost the same as RTX 4080 16GB($1199), then AMD would have lower margin than Nvidia, because the chip costs more.

P.S. of course, I am excluding the cost of the whole card(memory,cooler,power delivery etc.).

I did state I was only looking a good dies because we don't know how many of the defective dies are actually usable, could be all of them could be some could be none, depends where the defect is.

This also does not take into account parametric yield so of those good dies not all will be good enough quality to be used in full N31 because power consumption at the required clock speed will be too much or it just won't clock that high at all.

All this does is give you an indication.

I also never tried to compare full N31 to the 4080 16GB. I said the 7900XT which is the 20GB cut N31 model.

Where this becomes more interesting is comparing RDNA3 to RDNA2. Full N21 is > 500mm^2 so a wafer can get about 67 good dies which went into $1,000 and $1,100 SKUs. Cost per die with packaging depends on N7 cost but it is meant to be more expensive than N6 so lets just give it a 10% higher price ($11,000) and that works out to $170 / full chip.

If a cut N31 can sell in a $1,200 then it is obvious that the margin RDNA3 is going to be higher than RDNA2 at the high end.

Comparing Full N32 to cut N21 gets you $113 for full N32 and $115 for cut N21 but the power requirements for full N32 will be lower so the cooler, vrms, PCB etc can all be a bit cheaper giving it a lower BOM. I do expect 7800XT pricing to be closer to the $800 range than it is the 6800XT price point but performance should be a lot better so perf/$ will still improve a lot.

Again if you compare full N22 to cut N32 then we have $90 for N22 and ~$90 for N32 but I also expect the 7700XT to go up in price vs the 6700XT. $550-600 for 7700XT for above 6950XT performance in a ~250W power envelope does seem pretty impressive and again increases AMDs margins while also offering a better perf/$ proposition.

N33 is even more obvious because it is smaller than N23 and on a cheaper node.

What it shows is that AMD can both increase margins for their GPU stack and offer a meaningful perf/$ increase over their old parts and vs 4000 series.

 

[TESKATLIPOKA]

.....
I also never tried to compare full N31 to the 4080 16GB. I said the 7900XT which is the 20GB cut N31 model.

Where this becomes more interesting is comparing RDNA3 to RDNA2. Full N21 is > 500mm^2 so a wafer can get about 67 good dies which went into $1,000 and $1,100 SKUs. Cost per die with packaging depends on N7 cost but it is meant to be more expensive than N6 so lets just give it a 10% higher price ($11,000) and that works out to $170 / full chip.

If a cut N31 can sell in a $1,200 then it is obvious that the margin RDNA3 is going to be higher than RDNA2 at the high end.
.....

You mentioned $1,200 7900XT offering RTX 4090 performance or close to It, so I assumed RX 7900XT was a typo, and you actually meant RX 7950XT.

 

[Timorous]

You mentioned $1,200 7900XT offering RTX 4090 performance or close to It, so I assumed RX 7900XT was a typo, and you actually meant RX 7950XT.

The math works on it.

A claimed >50% perf/watt increase and AMD have historically compared top SKU to top SKU. Last time out they claimed 50%, achieved 54% with the 6800XT and 64% with the 6900XT both vs the 5700XT. We also know the 6800 has far better perf/w than both the 6800XT and 6900XT since cut parts at lower TBPs tend to have better perf/watt than the bigger brothers. This leads me to think if AMD are claiming >50% which is likely 6950XT vs 7950XT then the 7900XT might be a bit ahead of that and a 60% perf/watt gain for the 7900XT vs the 6950XT with similar TBP would put the 7900XT at ~95% of a 4090. That is pretty close in my book.

There is also the rumour that mobile N32 has an ~80% perf/watt increase over mobile N22 but no idea how much truth there is in that.

So it seems to be in the ballpark of possible. If they under hit they will charge less and have lower margins, if they over hit they may charge more and have higher margins but I think that kind of estimate is not far fetched.

 

[TESKATLIPOKA]

The math works on it.

A claimed >50% perf/watt increase and AMD have historically compared top SKU to top SKU. Last time out they claimed 50%, achieved 54% with the 6800XT and 64% with the 6900XT both vs the 5700XT. We also know the 6800 has far better perf/w than both the 6800XT and 6900XT since cut parts at lower TBPs tend to have better perf/watt than the bigger brothers. This leads me to think if AMD are claiming >50% which is likely 6950XT vs 7950XT then the 7900XT might be a bit ahead of that and a 60% perf/watt gain for the 7900XT vs the 6950XT with similar TBP would put the 7900XT at ~95% of a 4090. That is pretty close in my book.

There is also the rumour that mobile N32 has an ~80% perf/watt increase over mobile N22 but no idea how much truth there is in that.

So it seems to be in the ballpark of possible. If they under hit they will charge less and have lower margins, if they over hit they may charge more and have higher margins but I think that kind of estimate is not far fetched.

RTX 4090 is 82.5% faster than RX 6950XT in Cyberpunk 2077 for comparison.


I made a table based on Greymon55's claims.
Greymon55 claims TBP of reference is amazing and that raster is 2x better.

https://twitter.com/x/status/1582599964567314432

https://twitter.com/x/status/1582630645674934272

I got up to 100% better efficiency to be sure, because AMD seriously understated Zen4's perf. Improvement, so I don't know If they won't do It again.

TBP	Efficiency Increase	Perf. at 335W	Perf. at 350W	Perf. at 375W	Perf. at 400W	Perf. at 450W
6950XT	-	100%	-	-	-	-
7950XT	50%	150%	156.7%	167.9%	179%	201.5%
7950XT	55%	155%	161.9%	173.5%	185%	208%
7950XT	60%	160%	167%	179%	191%	215%
7950XT	65%	165%	172%	184.7%	197%	221.6%
7950XT	70%	170%	177.5%	190%	203%	228%
7950XT	75%	175%	183%	196%	209%	235%
7950XT	80%	180%	188%	201.5%	215%	242%
7950XT	85%	185%	193%	207%	221%	248.5%
7950XT	90%	190%	198.5%	212.5%	227%	255%
7950XT	95%	195%	204%	218%	233%	262%
7950XT	100%	200%	209%	224%	239%	269%

Not sure what TBP he considers amazing, so everyone can choose what they think.
RX 7900XT will likely lose ~15% of performance compared to Full N31.

P.S. Nvidia also achieved serious improvement in efficiency with RTX 4090.

P.S.2 That rumor about 80% improvement of mobile N32 was from me. It was based on a leak from Greymon55 about RX 6950XT level of performance. So I just compared mobile N22 against RX 6950XT.

 

[HurleyBird]

So for N31 that would total $152.20 + packaging costs and for N32 it would be $91.80 + packaging costs. In the video Ian estimated packaging costs for the 7950X as around $7.50 so with 7 chips instead of 3 lets just go with a high end of $30 for packaging and then round up to the nearest $5 and that gives N31 a total cost of $185 and N32 a total cost of $115.

Probably a bit low because you also need to add the the board, memory, heatsink, accessories, etc.

 

[Timorous]

RTX 4090 is 82.5% faster than RX 6950XT in Cyberpunk 2077 for comparison.
View attachment 69727

I made a table based on Greymon55's claims.
Greymon55 claims TBP of reference is amazing and that raster is 2x better.

https://twitter.com/x/status/1582599964567314432

https://twitter.com/x/status/1582630645674934272

I got up to 100% better efficiency to be sure, because AMD seriously understated Zen4's perf. Improvement, so I don't know If they won't do It again.

TBP	Efficiency Increase	Perf. at 335W	Perf. at 350W	Perf. at 375W	Perf. at 400W	Perf. at 450W
6950XT	-	100%	-	-	-	-
7950XT	50%	150%	156.7%	167.9%	179%	201.5%
7950XT	55%	155%	161.9%	173.5%	185%	208%
7950XT	60%	160%	167%	179%	191%	215%
7950XT	65%	165%	172%	184.7%	197%	221.6%
7950XT	70%	170%	177.5%	190%	203%	228%
7950XT	75%	175%	183%	196%	209%	235%
7950XT	80%	180%	188%	201.5%	215%	242%
7950XT	85%	185%	193%	207%	221%	248.5%
7950XT	90%	190%	198.5%	212.5%	227%	255%
7950XT	95%	195%	204%	218%	233%	262%
7950XT	100%	200%	209%	224%	239%	269%

Not sure what TBP he considers amazing, so everyone can choose what they think.
RX 7900XT will likely lose ~15% of performance compared to Full N31.

P.S. Nvidia also achieved serious improvement in efficiency with RTX 4090.

P.S.2 That rumor about 80% improvement of mobile N32 was from me. It was based on a leak from Greymon55 about RX 6950XT level of performance. So I just compared mobile N22 against RX 6950XT.

Why would you use 1 game as a baseline when there are perfectly good multi game averages.

Or

+ 50% of a 6950XT gets 92% vs the 4090 at TPU and 89% at Techspot/HUB.

+ 60% gets 98% and 94% respectively.

If 7900XT has the same TBP as the 6950XT then +50% is just the low end of AMDs perf/watt claim so it adds up.

The 7950XT being about 15%/20% faster would put it between 10% to 20% ahead of the 4090 which is exactly what the 420W TBP calcs work out to and what it would need to be to actually compete with the 4090Ti.

 

[TESKATLIPOKA]

Why would you use 1 game as a baseline when there are perfectly good multi game averages.

https://tpucdn.com/review/nvidia-geforce-rtx-4090-founders-edition/images/relative-performance_3840-2160.png

Or

Nvidia GeForce RTX 4090 Review Photo Gallery - TechSpot

View 60 detailed pictures from our Nvidia GeForce RTX 4090 Review, close-up photos in high resolution.

www.techspot.com

Here you have 83% better average performance in 17 games. ComputerBase

It's not RX 6950XT, but the difference against RX 6900XT is almost the same.

edit: I just found RX 6950XT review at ComputerBase and XFX RX 6900XT Black performs the same as RX6950XT. Link

 

[biostud]

Here you have 83% better average performance in 17 games. ComputerBase
View attachment 69734
It's not RX 6950XT, but the difference against RX 6900XT is almost the same.

I think the the 4090 was held back by the 5800X they used @TPU which is why the do not have so much improvement.

 

[TESKATLIPOKA]

I think the the 4090 was held back by the 5800X they used @TPU which is why the do not have so much improvement.

You just nailed It.

 

[Timorous]

You just nailed It.

PCGH tested 20 games with a 12900K and quite fast ram and go this.

So + 50% there is 85% of 4090 performance and + 60% is 90% performance.

Even with the CB results (assuming that 6900 ~= the 6950) gets you 83% with a +50% perf/watt bump and 89% with a +60% perf/watt bump.

Also what is the speed up of the 6950XT with the 12900K at 4K? probably less than 6.5% but not 0% either.

So I think the Techspot / HUB / PCGH numbers are about right to be honest. ComputerBase seems like an upper end outlier and TPU a lower end outlier.

 

[TESKATLIPOKA]

PCGH tested 20 games with a 12900K and quite fast ram and go this.

So + 50% there is 85% of 4090 performance and + 60% is 90% performance.

Even with the CB results (assuming that 6900 ~= the 6950) gets you 83% with a +50% perf/watt bump and 89% with a +60% perf/watt bump.

Also what is the speed up of the 6950XT with the 12900K at 4K? probably less than 6.5% but not 0% either.

So I think the Techspot / HUB / PCGH numbers are about right to be honest. ComputerBase seems like an upper end outlier and TPU a lower end outlier.

I think the difference in reviews is not the biggest problem we have.
We have too little info about N31 and RDNA3. We probably know the performance, which is supposedly 2x better for 7950XT.
Yet, we don't know TBP and efficiency. As shown in my table, we can reach 2x better performance at different TBPs depending on which perf/W improvement we choose.
For this reason, I won't bother guessing the performance of RX 7900XT.
Considering Greymon55's claim about TBP of reference being amazing, I would expect TBP to stay the same.
That would mean with 2x better performance you have 2x better perf/W, but that's too unrealistic, so I will lower my expectations to 375W TBP and 75-80% better perf/W.

 

[Timorous]

I think the difference in reviews is not the biggest problem we have.
We have too little info about N31 and RDNA3. We probably know the performance, which is supposedly 2x better for 7950XT.
Yet, we don't know TBP and efficiency. As shown in my table, we can reach 2x better performance at different TBPs depending on which perf/W improvement we choose.
For this reason, I won't bother guessing the performance of RX 7900XT.
Considering Greymon55's claim about TBP of reference being amazing, I would expect TBP to stay the same.
That would mean with 2x better performance you have 2x better perf/W, but that's too unrealistic, so I will lower my expectations to 375W TBP and 75-80% better perf/W.

That would be ~10% faster thana 4090 making the 7900XT 90% or more of a 4090.

Not sure it will get there at 375W, I am thinking more 400W to get to that level but it is not impossible it could happen at 375W.

Looking forward to finding out how wrong my guesses are on Nov 3rf.

 

[Joe NYC]

While I agree with you that if they have the wafers for a "price war" they should make use of that ASAP, planing of releases happens way, way earlier. AMD might have some bandwidth limitations in certain areas that make it hard to bring multiple products to the market at the same time. eg. N31 and N32 don't run fully parallel and at one point you must give one of the priority and that decision didn't happen last months but more likley a year ago where this whole inflation thing was om nobodies horizon yet.

One bottlebeck could be driver optimization, I^m sure they will vary a bit for each specific chip.

It's possible that there was an original plan, before the mining crash, inflation end economic problems, but the N31 and N32 are very similar, and there was plenty of time to switch priorities.

I don't think AMD needs to wage a price war. RDNA3 will have cost advantage. If AMD targets its "normal" gross margin of ~50%, this will translate to MSRP noticeably below comparable NVidia chip.

Here are my price estimates I posted elsewhere, in July, of what AMD could do to gain market share (and still maintain 50% gross margin). This was before Angstronomics reveal. With that info in mind, N33 could go even lower - $300:

N33: $400-500
N32: $550-800
N31: $850-1200
N31x3D: $1499