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

 

[exquisitechar]

You are assuming the 6700XT will be the 40CU part, what if its not?

What else would it be? Full fat Navi22 is 40CUs, below the 6800 should be the 6700XT.

 

[Glo.]

You are assuming the 6700XT will be the 40CU part, what if its not?

If its not, it will have, then, higher performance than 40 CU die .

If 6700XT is based on N21, with lets say 56 or 48 CUs, then it will still be faster than 40 CU/192 bit bus Navi 22 die.

I'd love to see that happen, cause that will push down the stack the Navi 22 and 23 dies. But I know this is business, and that ain't happening.

The performance targets do not change. 40 CU - between RTX 2080 Super and RTX 2080 Ti(10Above RTX 2080 Super).

I've just remembered about something.

6900XT - XTX Die.
6800XT - XT Die.
6800 - XL Die.
Unknown - XE Die.

So there might be an SKU below 6800 on the road based on N21, still.

That is interesting . And if we take this into context, the pricing of 6800 now makes absolute sense, for this context.

 

[leoneazzurro]

If its not, it will have, then, higher performance than 40 CU die .

If 6700XT is based on N21, with lets say 56 or 48 CUs, then it will still be faster than 40 CU/192 bit bus Navi 22 die.

I'd love to see that happen, cause that will push down the stack the Navi 22 and 23 dies. But I know this is business, and that ain't happening.

The performance targets do not change. 40 CU - between RTX 2080 Super and RTX 2080 Ti(10Above RTX 2080 Super).

I've just remembered about something.

6900XT - XTX Die.
6800XT - XT Die.
6800 - XL Die.
Unknown - XE Die.

So there might be an SKU below 6800 on the road based on N21, still.

That is interesting . And if we take this into context, the pricing of 6800 now makes absolute sense, for this context.

58/60CU, 192 bit, a bit lower clocks, reuses 6700XT's PCB?

 

[Mopetar]

If it isn't 40 CU as speculated then it would probably be 48. It's certainly possible for it to be even larger at something like 56, but I doubt that since it pushes up too closely against the 6800, which is a bit more cut down than I expected.

The clocks on the 6800 are low enough that at 48 CU die with clock speeds like the 6800XT would slot in well at $450, especially if it has 12 GB VRAM so even if it isn't quite as good as a 3070 on average it will be cheaper and with more VRAM which I think is enough for it to do well in the market.

 

[Shivansps]

If its not, it will have, then, higher performance than 40 CU die .

If 6700XT is based on N21, with lets say 56 or 48 CUs, then it will still be faster than 40 CU/192 bit bus Navi 22 die.

I'd love to see that happen, cause that will push down the stack the Navi 22 and 23 dies. But I know this is business, and that ain't happening.

The performance targets do not change. 40 CU - between RTX 2080 Super and RTX 2080 Ti(10Above RTX 2080 Super).

I've just remembered about something.

6900XT - XTX Die.
6800XT - XT Die.
6800 - XL Die.
Unknown - XE Die.

So there might be an SKU below 6800 on the road based on N21, still.

That is interesting . And if we take this into context, the pricing of 6800 now makes absolute sense, for this context.

Well thats exactly what i think, but if this is case i dont think it is going to push anything down, lets say 52CU at $500 for the 6700XT and $420 for the 40CU die RX 6700.

AMD has increased every price point this year so i can easily see this happening...

 

[exquisitechar]

I don't see why AMD would want to sell N21 with even more of it disabled. 6800 is already the black sheep of sorts.

 

[Kedas]

.....
And agreed, 6800 is nice, but a little bit more $, u get 6800XT with much more perf. 6900XT price is nuts for another 10%. No thanks. 6800XT is where it's at.

The 6900XT price is not based on 10% better performance but based on yield of 'gold' dies.

 

[moinmoin]

I mean hardware optimizations. The consoles use custom Zen 2, so whatever customization is going on in their boards is going to somehow emulate this Zen 3 + RDNA2 + 550 optimizations in a different way. I guess, if regular Zen2 can't work with DXR1.1 (I have no idea if that is the case), then that is already something...

AMD presented (SAM and RAGE) as a specific AMD ecosystem feature that requires Zen3. I'm not sure if that has any relation to RT performance or not, but I'm really curious about those console Zen 2 chips with RDNA2, and how it compares to Zen2 or Zen3 with RDNA2 on PC.

I highly doubt the AMD tech in the consoles significantly deviates from what's in PCs aside the memory system rebalance and custom audio system. SAM for example very likely is a solution to an issue that is first being solved on the consoles. With all these tech consoles are actually first, semi custom business allows AMD to be paid by others for R&D it can then introduce in its own products as well. This is what's happening here imo.

Maybe you glossed over my previous posts, RDNA2 by all indications is now also built onto the Infinity Fabric hierarchy (having SCF and SDF). In consoles both the CPU and GPU are on one die so both are within the same Infinity Fabric hierarchy. Not so on PC with dGPUs, so that's where SAM is being introduced to connect both separate device's Infinity Fabric across PCIe 4, enabling each device to act on its own on the other side. It's likely that this capability (which is likely part of the earlier announced Infinity Architecture) was only introduced in Zen 3 (likely Infinity Control Fabric, extending the local SCF beyond the package), thus necessitating that.

 

[uzzi38]

The 6900XT price is not based on 10% better performance but based on yield of 'gold' dies.

No the price of the 6900XT is based on the hopes of fishing for very large whales.

 

[YBS1]

The 6900XT price is not based on 10% better performance but based on yield of 'gold' dies.

Exactly......beyond that a lot of the type of people interested in the 6900XT are going to likely be the types to squeeze on it even harder. Watercooling, unlocked BIOS, etc. The performance difference between it and the 6800XT isn't what they are buying (and it's not in fact worth it), they are buying top binned chip possibilities.

 

[zinfamous]

CPU architecture has nothing to do with DXR 1.1

...well if that's a response to me, that's why I wasn't asking about it.

 

[Kenmitch]

How many of you guys were in the hype-train that flew over my house last night?

The EKWB water blocks look kind of nice. I'm not into it anymore, but I'd imagine a few members still are.

EKWB Has Special Edition Watercooling Parts for AMD's Latest CPUs and GPUs

Must-have parts for a high-end, custom liquid all-AMD build

www.tomshardware.com

 

[Glo.]

If it isn't 40 CU as speculated then it would probably be 48. It's certainly possible for it to be even larger at something like 56, but I doubt that since it pushes up too closely against the 6800, which is a bit more cut down than I expected.

The clocks on the 6800 are low enough that at 48 CU die with clock speeds like the 6800XT would slot in well at $450, especially if it has 12 GB VRAM so even if it isn't quite as good as a 3070 on average it will be cheaper and with more VRAM which I think is enough for it to do well in the market.

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.

The full stack would look like this:
6900XT - 80 CUs
6800XT - 72 CUs
6800 - 60 CUs
6700XT - 48 CUs
1st card below 6700 XT - 40 CUs(Navi 22 dies)
cut down N22 - 36 CUs
Next is N23 with 32 CUs and god only knows how many as Cut down.

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

 

[Glo.]

A side note.

Based on what I am hearing, AMD's marketing team underplayed the efficiency of RDNA2 GPUs, again.

That 250 and 300W of power TBP are supposedly numbers of when the GPUs are pushed absolutely to the wall.

In normal circumstances we should see 6900XT averaging around 275W of power.

6800XT will be more like 250-260W of power.

 

[MrTeal]

I can't see them releasing a 56CU N21 die, let alone a 48CU one. Even 56 is only 70% of a full die, which is a crazy amount to cut on a mature high yielding process. That's basically what the GTX 465 was at 69% of a full GF100, and that was a disaster chip that they were never able to release a full die version of before respinning it.

 

[andermans]

AMD may just release more cut-downs for market segmentation reasons totally unrelated to yield. At which point the question really is whether it is worth it vs. trying to capture some of the part of the market with a smaller and cheaper die.

 

[MrTeal]

AMD could stick a good but cut down N21 into a 6700 XT that they sell for ~$450, but that seems rather wasteful when they could sell it in a $570 6800 or $650 6800XT. If Nvidia released a really competitive GA106 lineup AMD might be pushed into doing so if N22 isn't ready, but right now there's no real need to push out a massively cut down card. A 60% enabled 48CU / 8GB card would be tops $400, that's a lot of money left on the table.

 

[blckgrffn]

AMD may just release more cut-downs for market segmentation reasons totally unrelated to yield. At which point the question really is whether it is worth it vs. trying to capture some of the part of the market with a smaller and cheaper die.

The smaller die makes sense, I think.

There is a lot of talk other places that the 6800 is advertised as having “up to” 16 GB of memory which implies a lower priced 8 GB version is possible/inbound.

Having a further cut down card have more memory creates confusion and it seems like they’d sidestep that for now.

The much lower clocks on the 6800 leave a lot of room for a “jump up” with fewer CUs but aggressive clocks, like PS5+ clocks and similar TBP for a 16GB 6700xt.

I am really interested to see if the next tier down retains the infinity cache. Full size, half size, or not at all. It seems like they could keep the cache and go to a 128 bit bus (is there ram sizing to support that at 16GB of GDDR6?) or drop/reduce the cache and stick to 256 bit bus width. There should still be a clear hierarchy there and the 5700xt faired well with the 256 bit bus it seemed.

 

[beginner99]

I can't see them releasing a 56CU N21 die, let alone a 48CU one. Even 56 is only 70% of a full die, which is a crazy amount to cut on a mature high yielding process. That's basically what the GTX 465 was at 69% of a full GF100, and that was a disaster chip that they were never able to release a full die version of before respinning it.

that's the issue that they don't have a die between 40 and 80 CUs. I agree 48CU cut seems like a lot it's too close to a 40CU full die and far from the 60CU one but 56 it too close too 60. More likley is 40CU has higher clocks so it acts more like 50CUs making the difference to 60CU a little smaller. But AMD for sure has a "hole" in their stack around the 50CU mark that is hard to fill economically.

 

[Tuna-Fish]

- The fact that they're going to offer a 60 CU N21 at all also strongly suggests they are severely supply constrained at the moment and are harvesting even garbage dies to sell. Its weird because what happens when supply opens up? AMD isn't going to be lopping off 20CUs off a die to sell forever, are they?

They are not just cutting off 20CU, they are cutting off a complete SE, with rops. This allows them to harvest dies where there is any fault in rops or other SE-level resources.

The memory controllers, last-level cache (Infinity Cache), display connectors, codec engines, etc can be on an IO die. This approach pretty much mirrors how it's done on Zen 2/3.

You cannot move the cache off-die or you get murdered by the off-die transmit power. Any potential future chiplet CPU is going to have a huge cache on every die. The cache is a huge part of exactly what makes splitting a large design into chiplets possible in the first place.

Yes, but is it an $800 increase? You can buy two 3080's* for the price of one 3090*.

Because it can fit larger models, a 3090 right now is the gpu to get if you want to train AI, and don't have the budget of a huge company. For that purpose, the extra memory is easily worth the extra price. IMHO that's it's main market, it's really only sold as a gaming card to get some cash out of the few whales who buy the best no matter how much it costs.

 

[alexruiz]

6700XT will be Navi 22.
Navi 22 will be 40CU, same as Navi 10.

As several of my fellow ATers pointed out, the 6800 is already cutting the Navi 21 die by a lot. Those TSMC 7nm wafers are in short supply, and if deciding on what to make, a Zen3 chiplet or a heavily cut down Navi 21, it is obvious what AMD will pick
The bulk of the market in below the high products, and AMD will cover it with a smaller die.
What I am curious to speculate about is how much faster a Navi22 card will be vs a Navi 10? Now much vRAM?

 

[Ajay]

They are not just cutting off 20CU, they are cutting off a complete SE, with rops. This allows them to harvest dies where there is any fault in rops or other SE-level resources.

The clocks are lower as well - so AMD is likely fusing off based on lower speed bins.

 

[MrTeal]

6700XT will be Navi 22.
Navi 22 will be 40CU, same as Navi 10.

As several of my fellow ATers pointed out, the 6800 is already cutting the Navi 21 die by a lot. Those TSMC 7nm wafers are in short supply, and if deciding on what to make, a Zen3 chiplet or a heavily cut down Navi 21, it is obvious what AMD will pick
The bulk of the market in below the high products, and AMD will cover it with a smaller die.
What I am curious to speculate about is how much faster a Navi22 card will be vs a Navi 10? Now much vRAM?

That's an interesting question. A 40CU class card probably doesn't need more than 8GB even going forward, but for 8GB it would be running either a 128b or 256b bus. Even with Infinity Cache 128b would be woefully inadequate, so if they went that route you think they'd need 256b/8GB. A 6/12GB card with a 192b bus would be a possibility, and would give some marketing points in the larger config. In either case, I'd expect to see 64MB of IC on the 40CU die.

 

[Ajay]

6700XT will be Navi 22.
Navi 22 will be 40CU, same as Navi 10.

As several of my fellow ATers pointed out, the 6800 is already cutting the Navi 21 die by a lot. Those TSMC 7nm wafers are in short supply, and if deciding on what to make, a Zen3 chiplet or a heavily cut down Navi 21, it is obvious what AMD will pick
The bulk of the market in below the high products, and AMD will cover it with a smaller die.
What I am curious to speculate about is how much faster a Navi22 card will be vs a Navi 10? Now much vRAM?

12GB vRAM for N22 according to _rogame, so 192b bus.

https://twitter.com/x/status/1306655000454725636

I wonder if MALL will be 96MB?

 

[GodisanAtheist]

that's the issue that they don't have a die between 40 and 80 CUs. I agree 48CU cut seems like a lot it's too close to a 40CU full die and far from the 60CU one but 56 it too close too 60. More likley is 40CU has higher clocks so it acts more like 50CUs making the difference to 60CU a little smaller. But AMD for sure has a "hole" in their stack around the 50CU mark that is hard to fill economically.

Frankly I'd be amazed if AMD does not end up with something like an 80/60/40/20 CU full stack to cover the entire range, maybe we'll see a 60 CU chip as a mid-gen refresh part. RDNA3 isn't showing up until (probably mid-late) 2022, so I cannot see AMD dropping 20CUs (one full shader engine as was mentioned earlier) off an N21 die way out in dole drums of 2021 as yields/supply improve.

I figure the full RDNA 2 stack will end up looking like 80/72 (N21) 60/52 (???) 40/32 (N22) 20/12 (???) for full top to bottom coverage.