Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[PhoBoChai]

Yep. 48 fits the bill.

60 CUs require you to disable one full Shader Array, of the 4 available. So you are left with 3 SA's. Disabling one Dual CU unit from each SA will give you 54 CUs. Possible, but IMO - more likely is exactly 48 CUs, with 192 bit bus. Unless... AMD will want to break the game more and offer full 256 bit bus, and 16 GB of VRAM as selling point.

I presume, the room below 6800 gives AMD the flexibility to adapt to what Nvidia releases.

There is no situation where a further cut N21 makes sense, particularly as you move down the stack and into higher volume markets. AMD doesn't want to sell 500mm+ dies for cheap. The 6800 existing is already non-optimal, and the price gap to 6800XT shows AMD doesn't intend to sell many 6800. They prefer ppl to target the 6800XT instead.

You have to view it as a limited wafer supply situation and how do you maximize revenue & profits for the wafers you have.

Factor in TSMC yields are awesome levels. Now you do the reanalysis.

If there's anymore cut N21, it will be low volume, OEM only SKU.

Its weird that N22 is 40 CU, and N23 is what, 32/36 CU? Why so close.. This never made sense.

N22 should be ~48 CU, to fit better in the stack.. :/

 

[PhoBoChai]

Expect more such picks in the future as soon as consoles are involved, you won't see RT effects being used by default that tank the FPS.

You've seen the PS5 GI & reflections in their upcoming games? That's running on 36 CUs. Look at Digital Foundry coverage of Spider Man, even those guys were mighty impressed. They claim its native 4K30, with RT GI, Reflections and Shadows. All 3 effects. With max detail settings, they even zoom in for special attention to model & texture quality and its impressive.

For comparison, Fortnite RTX also use 3 effects. Do you know what the FPS is at 4K on a 3080? 23 FPS (Techspot). It's disgusting how bad optimized it is.

I think people have a serious case of under estimating RDNA2 RT.

We will know a lot more once next-gen games come with RT designed for RDNA2 consoles.

 

[Glo.]

There is no situation where a further cut N21 makes sense, particularly as you move down the stack and into higher volume markets. AMD doesn't want to sell 500mm+ dies for cheap. The 6800 existing is already non-optimal, and the price gap to 6800XT shows AMD doesn't intend to sell many 6800. They prefer ppl to target the 6800XT instead.

You have to view it as a limited wafer supply situation and how do you maximize revenue & profits for the wafers you have.

Factor in TSMC yields are awesome levels. Now you do the reanalysis.

If there's anymore cut N21, it will be low volume, OEM only SKU.

Its weird that N22 is 40 CU, and N23 is what, 32/36 CU? Why so close.. This never made sense.

N22 should be ~48 CU, to fit better in the stack.. :/

Well, we all assume that Yields on 7 nm large die are perfect .

Which may not be true.

The fact that there are 4 SKUs based on N21 die, and we have yet seen only three suggests that there is one below 6800 SKU.

32 CU die is perfectly understandable. Its Mobile First die.

 

[Shivansps]

Im thinking about infinity cache. i wonder if they could apply this for a APU in some kind of combined CPU+IGP cache, like Broadwell had...

 

[PhoBoChai]

Well, we all assume that Yields on 7 nm large die are perfect .

Which may not be true.

The fact that there are 4 SKUs based on N21 die, and we have yet seen only three suggests that there is one below 6800 SKU.

32 CU die is perfectly understandable. Its Mobile First die.

Not perfect, just pretty damn good. Cut 6800XT already going to cover most of the defective dies, 6800 is overkill for harvesting reasons. Can't go below 60CU, so much die wasted.

"Mobile First" just means lower voltage & clks. Unless u mean different lower power focused node or something. Making an entire die design for "mobile" is silly when N22, downvolted will cater to that market fine. They are both too close to each other.

 

[PhoBoChai]

Im thinking about infinity cache. i wonder if they could apply this for a APU in some kind of combined CPU+IGP cache, like Broadwell had...

For sure. 32MB of cache will do APUs a miracle in perf leap, as they are big time banwidth bottleneck from system RAM.

 

[CastleBravo]

For sure. 32MB of cache will do APUs a miracle in perf leap, as they are big time banwidth bottleneck from system RAM.

Yeah, between "infinity cache" and DDR5, I bet we are going to see some impressive APUs in a year or two.

 

[BarkingGhostar]

Wonder if Nvidia and AMD bid to try and get a Cyberpunk bundle for the holidays, and if so, who won?

Didn't they just delay that game another month? Its like some dude if that nut on the ground came from his Harley.

 

[Shamrock]

Ok, one more. LOL

 

[Midwayman]

So RT and DLSS are now the most critical features?
If so, you can go and by the RTX 3070. Some other of us will get a Navi and "put up" with the "not so great drivers"

AMD won't cut the price, they really don't want to sell this one.
The price was set that high to push people to the RX 6800 XT.

No, but its silly to pretend that they aren't significant features both companies are competing on now. Same for the drivers. I don't think anyone can honestly say they have as good a history. The point isn't to bag on the 6800, but that there are considerations other than raster performance in pricing. I'm most likely going with a 6800XT, but I know I'm making concessions in some areas. The 10gb of ram just worries me on the 3080 especially since RT really can bump up ram usage.

 

[raghu78]

For sure. 32MB of cache will do APUs a miracle in perf leap, as they are big time banwidth bottleneck from system RAM.

Navi 21 has sixteen 64 byte AMD Infinity Fabric channels connecting the cache to the graphics engine. Navi 21 has eight 32 bit GDDR6 memory controllers. So I think AMD is maintaining a ratio of 1:2 (memory channels to cache channels)

N21 - 16 channels x 8 MB = 128 MB cache ~ 530 sq mm (256 bit G6 = 8 x 32 bit G6)
N22 - 12 channels x 8 MB = 96 MB cache ~ 340 sq mm (192 bit G6 = 6 x 32 bit G6)
N23 - 8 channels x 4 MB = 32 MB cache 235 sq mm (

https://twitter.com/x/status/1317851087777443841

) (128 bit G6 = 4 x 32 bit G6)

N23 die with 32 CU at 235 sq mm can accomodate only 32 MB cache.

N24 - 4 channels x 4 MB = 16 MB cache 150 sq mm ? (

https://twitter.com/x/status/1318116657470996482

)

Rembrandt with 8C/16T Zen 3 and 12 CU RDNA2 is claimed to have a die size of 208sq mm. ( https://twitter.com/patrickschur_/status/1316059243829657600 )
Van Gogh with 4C/8T Zen 2 and 8 CU RDNA2 GPU is claimed to have a die size of 162 sq mm ( https://twitter.com/patrickschur_/status/1315788414877347850 )

Rembrandt and Van Gogh both support 128 bit LPDDR5. From Renoir die size we can estimate a 32 MB Infinity cache in both RDNA2 APUs. So AMD's Infinity Cache is going to be present in all RDNA2 GPUs and APUs.

 

[Mopetar]

Navi 22 CANNOT come as 48 CU die.

The specs for N22: 40CUs/192 bit bus.

My question was how sure we were on the specs. I think someone else said it was taken from Apple software updates that have checked out, so I'll trust it unless there's a good reason to think maybe Apple isn't getting any of the full die cards so don't have an entry, but Illthat's unlikely.

Well, we all assume that Yields on 7 nm large die are perfect .

Which may not be true.

The fact that there are 4 SKUs based on N21 die, and we have yet seen only three suggests that there is one below 6800 SKU..

Why assume it's below? I'll admit that there's certainly room since there aren't any N21 cards with disabled cache or memory controllers which is a little suspect.

However, there's also no 6900 and there's a big price hole between the 6800 XT and 6900 XT that AMD might want to hold off on filling until they see what NVidia will do. It's odd that there aren't any 80 CU parts that can't hit 6900XT clocks. Sure many can be made into serviceable 6800XT, but I think there's another bin in there. An 80 CU part with a 1915 MHz game clock is about 5% on either side of the two XT cards. Sounds like a possible $750 card.

 

[DaaQ]

So RT and DLSS are now the most critical features?
If so, you can go and by the RTX 3070. Some other of us will get a Navi and "put up" with the "not so great drivers"

AMD won't cut the price, they really don't want to sell this one.
The price was set that high to push people to the RX 6800 XT.

You have to remember, the only reason "some" people want AMD competitive, is they think it will force Nvidia to lower prices.

 

[insertcarehere]

N22 - 12 channels x 8 MB = 96 MB cache ~ 340 sq mm (192 bit G6 = 6 x 32 bit G6)
N23 - 8 channels x 4 MB = 32 MB cache 235 sq mm (

https://twitter.com/x/status/1317851087777443841

) (128 bit G6 = 4 x 32 bit G6)

The correlation isn't perfect, but 32CU/48ROP (?) Navi 23 being so much smaller than 40CU/64ROP Navi 22 must be largely down to the difference in Cache size, it clearly takes up a big chunk of Navi 21's die and from the looks of things seems to be a large reason why Navi 22 could be so "large" for a 40CU part.

 

[ModEl4]

The point is, if your calculation is correct, from this post:
... then 40CU die has lost IPC compared to not even RDNA1, but also Navi 21.

Performance targets have been already discussed.

According to you, a 2.5GHz 40CU RDNA2 part with +20% IPC in relation with RDNA1 would be at +15% of a 2080Ti, which means that if RDNA2 has no IPC benefit vs RDNA1 then we would have:
2080Ti=100
+20% IPC part= 100+15%= 115
0 IPC part= 115/120*100= 95.8
In the last 3 weeks you changed the tune saying that it will be at 2080S+10% level which is way slower than 95.8 (Navi22 with 0 IPC improvements versus RDNA1 according with your calculations)
Who is saying again that RDNA2 lost IPC?
I really didn't expect all this controversy, I redone my calculations carefully just to check exactly and I revised my predictions by just 1% so really no essential change:
2.5GHz 40CU part is 87% 3070 at 4K (86% in my first post)
and regarding the 2.3GHz 32CU part, 5700 will only be 7-8% slower than this part at QHD (6-7% in my first post)
Regarding IPC, 2.25GHz 72CU 6800XT = 2 X 1.9GHz 40CU 5700XT according to AMD, so RDNA2 lost IPC?
This was a rhetorical question...
I feel confident regarding the above predictions, but if I'm wrong I'll have no problem admitting it when the reviews come.
In the mean time let's agree to disagree.

 

[mohit9206]

When do we expect the lowest end desktop parts to come out? You know the 20CU and 24CU parts. Hopefully it will be significantly faster than the 5500XT 8gb which costs $200.

 

[TESKATLIPOKA]

Navi 21 has sixteen 64 byte AMD Infinity Fabric channels connecting the cache to the graphics engine. Navi 21 has eight 32 bit GDDR6 memory controllers. So I think AMD is maintaining a ratio of 1:2 (memory channels to cache channels)

N21 - 16 channels x 8 MB = 128 MB cache ~ 530 sq mm (256 bit G6 = 8 x 32 bit G6)
N22 - 12 channels x 8 MB = 96 MB cache ~ 340 sq mm (192 bit G6 = 6 x 32 bit G6)
N23 - 8 channels x 4 MB = 32 MB cache 235 sq mm (

https://twitter.com/x/status/1317851087777443841

) (128 bit G6 = 4 x 32 bit G6)

N23 die with 32 CU at 235 sq mm can accomodate only 32 MB cache.

N24 - 4 channels x 4 MB = 16 MB cache 150 sq mm ? (

https://twitter.com/x/status/1318116657470996482

)

Rembrandt with 8C/16T Zen 3 and 12 CU RDNA2 is claimed to have a die size of 208sq mm. ( https://twitter.com/patrickschur_/status/1316059243829657600 )
Van Gogh with 4C/8T Zen 2 and 8 CU RDNA2 GPU is claimed to have a die size of 162 sq mm ( https://twitter.com/patrickschur_/status/1315788414877347850 )

Rembrandt and Van Gogh both support 128 bit LPDDR5. From Renoir die size we can estimate a 32 MB Infinity cache in both RDNA2 APUs. So AMD's Infinity Cache is going to be present in all RDNA2 GPUs and APUs.

So according to you a 40CU N22 has 50% more bandwidth and 3x more cache compared to a 32CU N23 while having only 1/3 more CUs?
So according to you a 32CU N23 has the same amount of Infinity cache as a 8-12CU APUS?
So according to you a 16-20CU N24 has only half of the infinity cache of an 8-12CU APU?
Are you so sure about all this?

 

[Tuna-Fish]

What chance that there's some built in redundancy here, that this is actually an 88 or 96 CU chip? Something to meet the Ti card that nvidia will surely drop in the coming months.

Zero, they showed pictures and you can count 40 WGPs

So according to you a 40CU N22 has 50% more bandwidth and 3x more cache compared to a 32CU N23 while having only 1/3 more CUs?
So according to you a 32CU N23 has the same amount of Infinity cache as a 8-12CU APUS?
So according to you a 16-20CU N24 has only half of the infinity cache of an 8-12CU APU?
Are you so sure about all this?

The demand for the cache is primarily driven by display resolution, as that determines the size of the render targets, not by how much computing power the card has. The cards are therefore targeted at specific resolutions. 128MB is for ~ 4k, 64MB ~ 1440p, 32MB ~ 1080p, 16MB ~ 720p. The wildcard is the N22, which has a 192-bit bus -- there might be 96MB on it if the cache needs to match up with the memory channels.

There is a 100m^2 difference in size between N22 and N23, with only a 8CU difference, while N23 has half the bus width of N10, yet is not that much smaller than it (so it's not cacheless). I also think the numbers lean very strongly to what Raghu posted.

N24 will probably just be a very weak bottom-tier card.

 

[Shmee]

Any news yet on how N21 will mine?

 

[coercitiv]

So RT and DLSS are now the most critical features?

Get used to this, RT and DLSS will be pushed hard as competitive advantage. Expect reflection galore in super-sampled concrete slabs.

 

[Dribble]

Get used to this, RT and DLSS will be pushed hard as competitive advantage. Expect reflection galore in super-sampled concrete slabs.

The problem is that without RT you don't really need one of these very high end gpu's.

 

[Kuiva maa]

Get used to this, RT and DLSS will be pushed hard as competitive advantage. Expect reflection galore in super-sampled concrete slabs.

I still believe that the strategy of having a big chunk of the die dedicated to just RT and DLSS is not sustainable in the long run. It locks nvidia to big die chips and is detrimental to general rasterization performance. Especially with consoles going the RDNA2 way.

 

[Olikan]

Since the confirmation of the Infinity cache, i am wondering, what the L2 cache is used for?
Infinity and L2 cache seem to fit the same role...

 

[Veradun]

I believe Cyberpunk 2077. Whichever card has the best performance in that game will get sold fast...

Provided it will release :>

 

[Veradun]

Its weird that N22 is 40 CU, and N23 is what, 32/36 CU? Why so close.. This never made sense.

N22 should be ~48 CU, to fit better in the stack.. :/

It only makes sense if N22 is HBM, designed with density in mind, and targeting lower frequencies to make a mobile beast