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]

Put up your current hard info. What was said a few months ago is now in flux.

TSMC's October 2022 revenue was up 56% YoY. Some of that is favorable currency but TSMC is not doing bad by any means. Sept and Oct were lower than August but only a bit.

N3 does have issues but that's N3.

https://investor.tsmc.com/english/monthly-revenue/2022

 

[scineram]

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.

Clearly $799 for the 16GB and $699 for the 12GB card make obvious sense. Maybe a little less for the latter. Or they will have 2 sku with 16, but I doubt that.

 

[biostud]

Clearly $799 for the 16GB and $699 for the 12GB card make obvious sense. Maybe a little less for the latter. Or they will have 2 sku with 16, but I doubt that.

That seems too expensive compared to current generation and to the 7900XT(X)

 

[Kaluan]

Clearly $799 for the 16GB and $699 for the 12GB card make obvious sense. Maybe a little less for the latter. Or they will have 2 sku with 16, but I doubt that.

I could see full N32 (7800XT?) at up to $750.
But no chance we'd see cut down (7700XT or 7800-esque) at $700.

500mm2+ (and more complex packaging/double the RAM) GPUs $900-1000, but 300mm2+ at $700? No freaking way. WTH

Why are some people here acting like we don't already know the MSRPs of N31 SKUs? 😂

 

[scineram]

So you say $699 and maybe $599 instead? I suppose if they want to be aggressive.

 

[Joe NYC]

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).

It doesn't appear Navi31 is bandwidth starved to need stacked MCDs at the GCD frequencies.

There may be some SKU configurations with less memory (fewer MCDs), possibly higher GCD clock speeds, where there would be more bandwidth demand on each MCD, so maybe the stacked MCDs will still emerge.

For example, a configuration I would not mind to have ( ) is full Navi 31 with 4 MCDs and stacked infinity cache.

 

[Kaluan]

So you say $699 and maybe $599 instead? I suppose if they want to be aggressive.

A soft maybe.

$649 and $549 would make even more sense to me. They could market price consistency between generations (while also mocking nVidia lol) AND great gen-on-gen gains (6750 XT vs 7700XT/cut N32, 6800XT vs 7800XT/full N32).

AND the gap between them will probably be smaller than 6700/6750XT and 6800XT but greater than the price difference, so they can upsell the 7800XT a-la 7900XT to XTX.

Edit: They'll probably want to do a N32 refresh at some point, so some price spacing between 7800XT and 7900XT/XTX leaves more room for $699-$749 7850XT or XTXs that could shore up their margins better.

 

[SteveGrabowski]

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.

So maybe don't expect it until say summer? Ugh I'm tempted to grab an RX 6700 10GB right now since I'm not sure we'll see them any lower, especially with it not seeming like a whole lot of 6700 non XT were made. Really would love to get a 130W card instead of a 180W one though. And nothing I really want to play on PC until Like a Dragon Ishin comes out in February.

 

[tajoh111]

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.

People still have expectation in line with pre Nov 3 info release and don't have an objective view of market conditions and pricing.

The 7800xt performance is going to be very close to the 6950xt's or 6900's. the 1.54 performance per watt shows this. With the small size of navi 32, AMD will have to be very careful about heat and clocks since the die is so small so heat transfer will be inefficient.

If this card is 3ghz but emitting 300 watts from a 200mm2, your going to have a chip with terrible minimums as the card thermal throttles with nothing short of water on top. Memory and a moderately clocked memory controller are relatively cool which is why they can be cooled passively. Pure logic is hot and AMD is packing it altogether tighter than ever.

A 700 dollar 6800xt has the same pitfalls as a 1200 rtx 4080, but without the criticism because of the bias in this thread. A small chip being upmarketed and up priced.

Remember the 5700xt was 399 and the 6700xt was 480 with the latter being up priced as a result of being launched in the middle of the mining boom.

This card should be priced at 499 considering there is only 200mm of 5nm chip there. The rest of it is cheap 6nm. Add this with the AMD brand and this should be 499 much like how the RTX 4080 should be priced more along the lines of 799.

 

[maddie]

People still have expectation in line with pre Nov 3 info release and don't have an objective view of market conditions and pricing.

The 7800xt performance is going to be very close to the 6950xt's or 6900's. the 1.54 performance per watt shows this. With the small size of navi 32, AMD will have to be very careful about heat and clocks since the die is so small so heat transfer will be inefficient.

If this card is 3ghz but emitting 300 watts from a 200mm2, your going to have a chip with terrible minimums as the card thermal throttles with nothing short of water on top. Memory and a moderately clocked memory controller are relatively cool which is why they can be cooled passively. Pure logic is hot and AMD is packing it altogether tighter than ever.

A 700 dollar 6800xt has the same pitfalls as a 1200 rtx 4080, but without the criticism because of the bias in this thread. A small chip being upmarketed and up priced.

Remember the 5700xt was 399 and the 6700xt was 480 with the latter being up priced as a result of being launched in the middle of the mining boom.

This card should be priced at 499 considering there is only 200mm of 5nm chip there. The rest of it is cheap 6nm. Add this with the AMD brand and this should be 499 much like how the RTX 4080 should be priced more along the lines of 799.

Isn't 1.5 W/mm^2 around Zen 4 power density?

 

[beginner99]

This card should be priced at 499 considering there is only 200mm of 5nm chip there. The rest of it is cheap 6nm. Add this with the AMD brand and this should be 499 much like how the RTX 4080 should be priced more along the lines of 799.

Given "older" pricing it should be more like $399 as flagships were around $600 max. But these days are gone.

Pricing entirely depends on clocking. with the 7900xt being $899 and a high clocking N32 would come rather close to it in terms of performance, I think even $699 is wishful thinking for a 7800xt. that would be a $200 gap or almost 30% jump. 7900xt is already unattractive right now and a 7800xt with say 15% less performance for 30% less money would make it useless. Therefore I expect it to be priced either higher or not really hitting these 3 ghz clocks as well. I think the perfromance gab will need to be at least 20% for a $699 price.

 

[eek2121]

neeeeed benchmarks…🤤

 

[Timorous]

People still have expectation in line with pre Nov 3 info release and don't have an objective view of market conditions and pricing.

The 7800xt performance is going to be very close to the 6950xt's or 6900's. the 1.54 performance per watt shows this. With the small size of navi 32, AMD will have to be very careful about heat and clocks since the die is so small so heat transfer will be inefficient.

If this card is 3ghz but emitting 300 watts from a 200mm2, your going to have a chip with terrible minimums as the card thermal throttles with nothing short of water on top. Memory and a moderately clocked memory controller are relatively cool which is why they can be cooled passively. Pure logic is hot and AMD is packing it altogether tighter than ever.

A 700 dollar 6800xt has the same pitfalls as a 1200 rtx 4080, but without the criticism because of the bias in this thread. A small chip being upmarketed and up priced.

Remember the 5700xt was 399 and the 6700xt was 480 with the latter being up priced as a result of being launched in the middle of the mining boom.

This card should be priced at 499 considering there is only 200mm of 5nm chip there. The rest of it is cheap 6nm. Add this with the AMD brand and this should be 499 much like how the RTX 4080 should be priced more along the lines of 799.

The math says otherwise.

On the updated AMD info the 7900XTX is 54% faster than a 6950XT and the 7900XT is 30% faster.

Using this chart that would put the 7900XTX at 131 fps and the 7900XT at 111 fps.

That would mean 7900XTX is $7.63 / frame and the 7900XT is $8.1 per frame.

To match the XTX $/frame a $700 7800XT would need 92fps so 3090Ti performance. Usually though the $/frame improves as you go down a stack so I could see it be as low as $7/frame which would put it at 100fps on this chart.

It might seem very close to the 7900XT in terms of performance but AMD are already not pricing the 7900XT in line with the XTX so I don't think AMD are worried about a lower tier part hurting 7900XT sales. I mean between now and the 7800XT launch AMD can adjust the price of the 7900XT to make it line up better in the perf/$ charts.

I would also say that anything much below 100 fps would feel like a lacklustre increase from the 6800XT. Matching the 3090Ti would just be a 30% improvement over the 6800XT which is not great.

The other angle you can look at it from is that the 7900XT with 84CUs and a slight clock regression vs the 6900XT is going to have around 38% more performance. A 7800XT with 60CUs but 40% higher clocks than a 6800 should quite comfortably offer a 60% gain over it which also puts it at 100fps in the above chart.

So I think that at worst the 7800XT will match the perf/$ of the 7900XTX which means 3090Ti as a low end. At best I can see it offering a ~10% value advantage at the cost of not hitting that 4k120 threshold the XTX does which would put it at around 100fps. Of course the 3rd option is that it does perform like a 3090Ti but AMD keep the $650 price point rather than pushing it to $700.

So really I guess I am saying that I expect between 3090Ti level performance and 3090Ti + 10% with a price between $650 - $700 to keep it the right side of the perf/$ curve set by the 7900XTX.

 

[Kaluan]

I think most people would also like to know how these cards and especially N32 cards perform at < 4K resolutions.

 

[TESKATLIPOKA]

One disadvantage for N32 compared to N31 is that It has supposedly only 3 shader engines, which is half of N31. We will have to see, how It affects performance.

 

[Timorous]

One disadvantage for N32 compared to N31 is that It has supposedly only 3 shader engines, which is half of N31. We will have to see, how It affects performance.

Assuming the ROP to SE ratio remains the same 96 rops feels too few for 7800XT tier cards which makes me think 4SEs with 16CUs each makes more sense.

EDIT: What makes me think 3SEs could be correct though is the layout in N31. It is two bank of 3SEs so if you just chopped half the SEs out and left the rest as is the layout is already there.

 

[TESKATLIPOKA]

Assuming the ROP to SE ratio remains the same 96 rops feels too few for 7800XT tier cards which makes me think 4SEs with 16CUs each makes more sense.

EDIT: What makes me think 3SEs could be correct though is the layout in N31. It is two bank of 3SEs so if you just chopped half the SEs out and left the rest as is the layout is already there.

4SE -> 64CU with 128 ROPs, 64MB IC and 256bit bus would be 2/3 of N31.
3SE -> 48CU with 96 ROPs, 48MB IC and 192bit bus would be 1/2 of N31.

What is interesting about RX 7900XT is that It probably still has 6 shader engines active.
If It had only 80CUs active, then It would be likely with 5 SE active.

 

[Heartbreaker]

GN talk with a AMD engineer about chiplets:

 

[PJVol]

GN talk with a AMD engineer about chiplets

Not a single question that goes beyond of what is already known.
Pretty useless one.

 

[Stuka87]

Not a single question that goes beyond of what is already known.
Pretty useless one.

Well, obviously he cannot ask about anything that is under NDA or not announced.

 

[PJVol]

Well, obviously he cannot ask about anything that is under NDA or not announced.

Yeah, I understand not much can be revealed until 13.12.
Anyway I'd personally like to know, for example,
how they dealt with an obviously increased power density for the GCD,
or did they manage to eliminate Cac overhead from toggling half of ALU's when SIMD can't issue two instructions in parallel.

 

[TESKATLIPOKA]

Yeah, I understand not much can be revealed until 13.12.
Anyway I'd personally like to know, for example,
how they dealt with an obviously increased power density for the GCD,
or did they manage to eliminate Cac overhead from toggling half of ALU's when SIMD can't issue two instructions in parallel.

Zen4's CCD is ~72.5 mm² and 7700x has TDP: 105W(PPT: 142W), remove IO die(~15W) and you end up with 127W for CCD alone. That's up to ~1.75W/mm² for a single CCD.

RX 7900XTX has 355W TBP, just the GCD is ~240-255W, maybe even less.
That would mean only ~0.8-0.85 W/mm² for GCD.
So where is the problem?

 

[PJVol]

So where is the problem?

Simple math so far. I took Navi21 - 520mm2/300W TBP , Navi31 530mm2/355W TBP. Considering 225mm2 of L3 macros and ddr phy relatively low power compared to the GCD combinational logic, it looks like almost double the power density of Navi21 (I've got ~ 0.5 and 0.9 Wmm-2)

Navi21 stock cooler struggle to handle 300W+ PPT (stock 255)

 

[TESKATLIPOKA]

Simple math so far. I took Navi21 - 520mm2/300W TGP , Navi31 530mm2/355W TGP. Consider 225mm2 of L3 macros and ddr phy relatively low power compared to the GCD combinational logic, it looks like almost double the power density of Navi21 (I've got ~ 0.5 and 0.9 Wmm-2)

Navi21 stock cooler struggle to handle 300W+ PPT (stock 255)

Are you talking about heat(power) density of the chip or If the cooler for N31 will be able to cool It?
TBP is only 20W more than RX 6950XT and the cooler is bigger, so I don't think the cooler won't keep up. GDC is smaller than Navi21 so the heat needs to be removed from a smaller area, but as I pointed out Zen4 is a lot worse in this regard, so I don't think It should be a problem.

TGP(total graphic power) and TBP(total board power) is not the same.
RX 7900XTX has 355W TBP.
RX 6900XT has 300W TBP.
RX 6950XT has 335W TBP.

 

[PJVol]

Are you talking about heat(power) density of the chip or If the cooler for N31 will be able to cool It?

Both, cause 1st (actually Pd(x,y)) should define the hotspot range under a TDP workload and 2nd - will it manage to sustain max GFX clock in these conditions.

TGP(total graphic power) and TBP(total board power) is not the same.

Yes, TGP of Navi21 is 255W indeed (202W core logic in TimeSpy), corrected prev. post. That gives 0.5 and ~0.8 W*mm-2.

but as I pointed out Zen4 is a lot worse in this regard

Yes, I'd say it's a nightmare, given half of a CCD is low power SRAM, and Pd(x,y) at the FPU logic should be enourmous. And yet I think it's more suitable to compare prev and next gen gpus.