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

 

[Timorous]

That's what I mean, I'm not sure it'd be near 6900 XT performance. The cut model definitely wouldn't be.

Keep in mind it would need to be 90% faster than the 6650 XT and 103% at 1440p to get to 6900 XT levels.

It has 100% the shaders and higher clocks than a 6650XT. Bandwidth may be an issue but AMD will have modelled fast cache vs more cache to see what works best. For 1080p I have no doubt fast is probably better. 1440p not so sure and 4K is obviously more but a 7600XT with 8GB vram won't be targeting 4K, it will be a 1080p card with the ability to do well at 1440p on a budget.

 

[DisEnchantment]

I think the cache amounts are wrong for the given MCD die size.

32bit PHY in N21 is around 6mm. Two of them are 12mm which means we have 25mm left of the 37.5mm die size. The Zen 3 cache dies are 64MB for 36mm of die area so 32MB would fit in 18mm which is a total of 30mm leaving 7.5mm for other bits n pieces that may be needed like links to the GCD.

Actually they seem OK.
Since they are on N6, you can go by N21 IF sizes. ~88 mm2 for 128 MiB or 22mm2 for 32MiB
2x 32 Bit PHY + UMC = 2 * 6.9mm2 = 13.8mm2
Per MCD = 22+14 -> 36mm2 , round about the values given in that article

 

[Timorous]

Actually they seem OK.
Since they are on N6, you can go by N21 IF sizes. ~88 mm2 for 128 MiB or 22mm2 for 32MiB
2x 32 Bit PHY + UMC = 2 * 6.9mm2 = 13.8mm2
Per MCD = 22+14 -> 36mm2 , round about the values given in that article

Article is saying 16MB with 64bit PHY though and as you say 32MB + 64bit PHY seems to fit in the 37.5mm die area given. It does also claim AMD have optimised for speed and reducing the hit when accessing VRAM.

 

[DisEnchantment]

Article is saying 16MB with 64bit PHY though and as you say 32MB + 64bit PHY seems to fit in the 37.5mm die area given. It does also claim AMD have optimised for speed and reducing the hit when accessing VRAM.

Oops, Well it looks like we share similar opinion on this then

 

[Frenetic Pony]

The arrangement of chiplets here makes some sense, a lot of memory chiplets means if one goes bad in placement you can easily bin the chip while the rest still work and they'd be cheaper to lose than graphic dies, especially if they're so small.

The thing that doesn't sit right with me is the limited GPU resources. 384bit bus for a mere 20% increase in GPU resources doesn't work for the math, especially when the reworked LLC should increase hit rate in some scenarios and so you'd be getting even more bandwidth efficient.

This would be a pointless card, the clockspeed of this thing is not hitting 4ghz+ and so bandwidth would be vastly undersatured, AMD would just be throwing money away. If these details are at all accurate, and not educated guesswork off the leaks so far which they very well could be, I'd be willing to bet any 384bit bus card would see 2 graphics compute dies. That would be much more amenable to a 384bit bus, some extra bandwidth efficiency, and even Samsung's 24gbps ram which they're sampling and intent on producing soon.

The other thing that doesn't sit right with me is the "60CU" card. If the other two are to be believed then the repeating block size is 16CUs, not 12. No chip designer is going to change that up just on one product in the lineup just for funsies. So either that's a weird typo, or this is just guesswork dressed up as a leak for the sake of attention seeking. These predictions look almost exactly like mine except they didn't check the math.

 

[Karnak]

It's a +140% ALU increase going from 5120 all the way up to 12288.

 

[Frenetic Pony]

It's a +140% ALU increase going from 5120 all the way up to 12288.

Another math error, "legacy" CU for RDNA2 was 80CUs for Navi 21, thus 96CUs would be a 20% increase. Or that they're quadrupling the SIMD32 per block, but there's no indication of that, just doubling, and the blog post doesn't indicate that either.

Real or fake this has a bunch of typos and errors.

 

[Karnak]

Don't see "typos" or data mentioned which I would call fake. Seems all reasonable to me.

And now kepler's tweet from a few days ago makes even more sense regarding his "N33 CUs are different from N31/N32 CUs".

Besides that it's Skyjuice. Nothing against like greymon but... he's on another level.

 

[Joe NYC]

Then the desktop version is probably going to be cancelled because of The Flood. Same thing with AD106, I don't expect it on desktop any time soon.

Navi 33 is going to be super competitive part, with smaller die size than 6600 XT. So AMD can price these competitively and let NVidia worry about The Flood.

 

[jpiniero]

It has 100% the shaders and higher clocks than a 6650XT.

That's what I mean. The shrink to N6 is decent but not that decent where they can double the shaders, increase clocks, etc... and be 15% smaller than N23.

 

[Timorous]

That's what I mean. The shrink to N6 is decent but not that decent where they can double the shaders, increase clocks, etc... and be 15% smaller than N23.

They have rearchitected the WGPs by halving FP64, halving L1 and doubling shader count. They have also increased L2 allegedly.

So it looks like there are a lot of changes under the hood so strong statements like this are foolish.

 

[jpiniero]

They have rearchitected the WGPs by halving FP64, halving L1 and doubling shader count. They have also increased L2 allegedly.

So it looks like there are a lot of changes under the hood so strong statements like this are foolish.

Let me put it this way... a straight shrink of N23 to N6 would be bigger than this. The stuff you are talking about would have to take up a ton of space.

I think our leaker friend forgot this is on N6 instead of N5.

 

[DisEnchantment]

They have rearchitected the WGPs by halving FP64, halving L1 and doubling shader count. They have also increased L2 allegedly.

So it looks like there are a lot of changes under the hood so strong statements like this are foolish.

Doubling FP32 ALUs per SIMD would halve FP32:FP64 ratio if FP64 remains same.
But this one is interesting,

https://twitter.com/x/status/1558266427949432834

It could mean there are 6 VGPR banks for N31 but 4 for N33?
Full blown 1 cycle Wave64 support would be possible if that is the case for N31.
It is very odd to have such big difference between N31 and N33 but we shall see.

 

[Leeea]

Navi 33 is going to be super competitive part, with smaller die size than 6600 XT. So AMD can price these competitively and let NVidia worry about The Flood.

One thing we have seen over the years is AMD is capable of producing price competitive parts even in the face of extreme odds.

The whole rx500 series is a case of AMD cannot out perform or compare against Nvidia, so just make it cheap to manufacture and sell a ton.


Competing against used parts on an older process that have no warranty? I think AMD will have no problem staying on top the flood. AMD knows how to make cheap BOMs that perform good enough.

 

[exquisitechar]

The whole rx500 series is a case of AMD cannot out perform or compare against Nvidia, so just make it cheap to manufacture and sell a ton.

I don’t think Polaris was particularly cheap compared to Nvidia’s equivalent. AMD was far behind and needed larger dies and cranked up power to compete. Gamers might remember it fondly because of how cheap it got for the good performance it provided, but Polaris wasn’t much of a winner for AMD.

The RDNA3 GPUs seem to be quite different. Navi33 will be incredible and the perf/area of the whole lineup in general looks like it will be great. I expect Nvidia to have the performance crown with full fat AD102, but AMD should have the more efficient and cheaper to manufacture GPUs. 7600XT could be a killer mid range GPU, but I think AMD won’t be very generous in regards to the pricing and most of Navi33 will go to laptops anyway.

 

[Karnak]

Well at ~200mm² for N33 that's most likely a $3xx GPU. On the MSRP side though since it's on the cheap AF N6 node (if Angstronomics is right, still thinking it's the same N5 for all of RDNA3 like AMD said) and it's that small that they can produce a lot of them. And I mean a lot if they want to.

Performance wise I think it'll sit in between the 6800 and 6800XT. Give or take. Which would be outstanding considering the die size (~200mm² vs. 519mm²) and almost the same node. Power consumption somewhere around 150W-180W (prob the latter ) and there you go.

Small, cheap to manufacture, great efficiency and great performance as a potential $3xx GPU. I'm pretty sure AMD won't sell a ~200mm² die for anything above $400. Even N23 is cheaper (MSRP wise) although it has a larger die.

 

[TESKATLIPOKA]

That's what I mean, I'm not sure it'd be near 6900 XT performance. The cut model definitely wouldn't be.

Keep in mind it would need to be 90% faster than the 6650 XT and 103% at 1440p to get to 6900 XT levels.

RX6900 XT is only ~53% faster than 6600XT at FullHD, Link.


50% more performance should be possible for N33 with 2x more shaders and higher clocks, but I think 32MB IF is really too low, my bet is 64MB.

 

[exquisitechar]

RX6900 XT is only ~53% faster than 6600XT at FullHD, Link.
View attachment 65825

50% more performance should be possible for N33 with 2x more shaders and higher clocks, but I think 32MB IF is really too low, my bet is 64MB.

That has to be because of CPU bottlenecks at 1080p. The difference is actually quite a bit larger than that.

Small, cheap to manufacture, great efficiency and great performance as a potential $3xx GPU. I'm pretty sure AMD won't sell a ~200mm² die for anything above $400. Even N23 is cheaper (MSRP wise) although it has a larger die.

Yeah, I agree.

 

[Grabo]

Navi33 is sometime 2023, "Navi33 is the mobile-first push for AMD. They expect robust sales of AMD Advantage laptops with it, as the design is drop-in compatible with Navi23 PCBs, minimizing OEM board re-spin headaches. They aim to ship more Navi33 silicon for mobile than to desktop AIB cards". Would be great for Euro winter electricity prices 2022 but I suspect it'll be nice for Euro winter electricity prices 2023 as well

The Ångström post is pretty detailed and he posted for the first time a couple of months back, at least on that site. It'll be easy to see how accurate everything is and as such how trustworthy he is. I didn't quite understand "Besides that it's Skyjuice." as if he's been posting for a lot longer than two months and indeed been proven correct a number of times, maybe he has and has been active elsewhere? I don't know : )

 

[jpiniero]

RX6900 XT is only ~53% faster than 6600XT at FullHD, Link.

I am basing it upon Compubase's review. Compared to the 6750 XT, it would need to be 43% faster at 1080p and 49% at 1440p to get to 6900 XT levels.

AMD Radeon RX 6650 XT, RX 6750 XT und RX 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT

RX 6650 XT, 6750 XT, 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT / Testsystem und Testmethodik

www.computerbase.de

 

[TESKATLIPOKA]

That has to be because of CPU bottlenecks at 1080p. The difference is actually quite a bit larger than that.

Does It really matter If It is because of a CPU bottleneck or not? N33 is supposedly aimed for 1080p. If It manages comparable FPS as RX 6900XT at lower cost and power consumption, It will be a good card.

I am basing it upon Compubase's review. Compared to the 6750 XT, it would need to be 43% faster at 1080p and 49% at 1440p to get to 6900 XT levels.

AMD Radeon RX 6650 XT, RX 6750 XT und RX 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT

RX 6650 XT, 6750 XT, 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT / Testsystem und Testmethodik

www.computerbase.de

Not sure why is the difference between N23 vs N21 so large between TPU and CB reviews.

edit: I see, TPU only uses R7 5800x, but still the difference is huge.

 

[jpiniero]

Not sure why is the difference between N23 vs N21 so large between TPU and CB reviews.

Different games, newer drivers, faster CPU (Compubase used the 12900K with DDR5 5400 )

 

[Timorous]

I am basing it upon Compubase's review. Compared to the 6750 XT, it would need to be 43% faster at 1080p and 49% at 1440p to get to 6900 XT levels.

AMD Radeon RX 6650 XT, RX 6750 XT und RX 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT

RX 6650 XT, 6750 XT, 6950 XT im Test: Taktraten, Benchmarks (FHD, WQHD & UHD) mit und ohne RT / Testsystem und Testmethodik

www.computerbase.de

So you don't think 2x the ALUs running at higher clocks can improve performance 50%?

RV770 managed more than that over RV670 so AMD have prior for it.

Considering the changes though performance speculation is going to be tricky because I can't imagine AMD are getting 100% scaling although I do think they will see more than 50% scaling.

 

[jpiniero]

So you don't think 2x the ALUs running at higher clocks can improve performance 50%?

6750 XT already basically runs at 2.7. Plus it has much more IF$ and memory bandwidth than N33 allegedly has. I am also wondering about what the ROPs and TMUs configuration is.

I wouldn't rule it out but it's looking more like N33 is more of an AD106 competitor rather than 4070. That may not really be a problem given the mobile focus.

 

[Kepler_L2]

6750 XT already basically runs at 2.7. Plus it has much more IF$ and memory bandwidth than N33 allegedly has. I am also wondering about what the ROPs and TMUs configuration is.

I wouldn't rule it out but it's looking more like N33 is more of an AD106 competitor rather than 4070. That may not really be a problem given the mobile focus.

2.7 GHz is base clocks territory for RDNA3. And ofc it's an AD106 competitor, it's literally less than half of xtors of AD104.