Question Speculation: RDNA2 + CDNA Architectures thread

[TESKATLIPOKA]

Then what will replace RX 5500XT 8GB?
If a cutdown N23 with 8GB will be sold for $229 then OK, I think $30 more for much higher performance is worth It.

 

[Glo.]

Then what will replace RX 5500XT 8GB?

I wouldn't worry about that .

Altho, im not sure it will happen in 1H 2021.

 

[uzzi38]

The only thing I can confidently say is Navi22 and Navi23 should both be there for laptops at CES for show if nothing else. The only question is if AMD will bring them to desktops straight away.

 

[TESKATLIPOKA]

I am most interested in laptops, so that would be great for me. The question is, when will we be able to buy the first laptops with them.

 

[soresu]

Got to be at least liquid cooled.

I'd wager more exotic than that - even a chilled liquid cooler might have problems keeping that stable.

It will definitely be interesting though given this achievement even with exotics just how far RDNA3 and 4 will be able to go on air alone.

 

[uzzi38]

I am most interested in laptops, so that would be great for me. The question is, when will we be able to buy the first laptops with them.

Don't know, but I would guess around March

 

[Glo.]

Well, I have to be fair with something.

The only thing that was told to me was literally: "RX 6500 XT will have 24 CUs and 6 GB VRAM with 96 bit bus, and it will be up to 15% faster than GTX 1660 Ti."

So there only two possibilities. Either this is heavily cut down Navi 23 die, or its completely different die. And the only option for that is Navi 24 for which we already have "traces" that it exists.

That 6500 XT is based on N24 with that specific CU/memory bus config is just my assumption, based on known stuff.

 

[TESKATLIPOKA]

Here is some info from Yuko Yoshida about Infinity and memory configs of N22, N23 and N24..

64/48/32
192/128/64

Link

And another one about N24.

And it's 24+64 Still has Infinity Cache.

Link

What you wrote could be really just a cutdown N23, but It would have the same number of CUs as N24.
N24 could end up with more Vram, but that was already the case with 5500XT vs 5600XT, so It wouldn't be such a surprise.

 

[Kepler_L2]

I believe it's 192/128/96 bit bus for Navi 22/23/24 respectively, with Navi 22 also having a 160 bit cutdown config.

IC config should be 96/64/48 MB for N22/23/24, with N22 cutdown having 80MB.

 

[Glo.]

Here is some info from Yuko Yoshida about Infinity and memory configs of N22, N23 and N24..

Link

And another one about N24.

Link

What you wrote could be really just a cutdown N23, but It would have the same number of CUs as N24.
N24 could end up with more Vram, but that was already the case with 5500XT vs 5600XT, so It wouldn't be such a surprise.

Im pretty certain that on both ocasions in his info Yuko is wrong in one way or another.

N24 won't end up with more VRAM, than 6500/XT because 4 GB memory chips have eye watering prices, compared to 2 GB ones.

So its always 6 GB's VRAM.

 

[Glo.]

I believe it's 192/128/96 bit bus for Navi 22/23/24 respectively, with Navi 22 also having a 160 bit cutdown config.

IC config should be 96/64/48 MB for N22/23/24, with N22 cutdown having 80MB.

Im fairly certain that what you wrote, compared to Yuko, is correct.

Apart from maybe that 80 MB IC part .

 

[Kepler_L2]

Im fairly certain that what you wrote, compared to Yuko, is correct.

Apart from maybe that 80 MB IC part .

Drivers suggest IC is tied to memory channels, if there is a cutdown 160 bit N22 (6600 XT 10 GB?) then it should have 80 MB IC.

 

[Glo.]

Drivers suggest IC is tied to memory channels, if there is a cutdown 160 bit N22 (6600 XT 10 GB?) then it should have 80 MB IC.

If so then Yes, you are correct.

Interesting thought about possibility of 6600XT being 10 GB and based on Navi 22 die.

I know where this might come from.

 

[TESKATLIPOKA]

Im pretty certain that on both ocasions in his info Yuko is wrong in one way or another.

N24 won't end up with more VRAM, than 6500/XT because 4 GB memory chips have eye watering prices, compared to 2 GB ones.

So its always 6 GB's VRAM.

RTX 3090 has 384bit GDDR6x and 24x 1GB GDDR6X chips.
Link
So you actually can have 4x 2GB GDDR6 chips for a total of 8GB Vram with only 64bit width memory bus and you don't need 4GB per chip.

 

[Glo.]

RTX 3090 has 384bit GDDR6x and 24x 1GB GDDR6X chips.
Link
So you actually can have 4x 2GB GDDR6 chips for a total of 8GB Vram with only 64bit width memory bus and you don't need 4GB per chip.

Well, sure, but that has its own drawbacks. You need to provide cooling on the backside of the GPU.

Which is a no - go for cheap, sub-170$ GPU, that is supposed to be made as cheaply as possible, to save margins.

 

[TESKATLIPOKA]

I just said It's possible, not that It's the cheapest idea.

BTW I don't understand why N24 has 24CU and not be just 1/4 of the full N21.
Like this -> 20CU, 1280SP, 80TMU, 32ROPs, 64bit GDDR6 and 32MB IC.
Ok, one explanation would be 96bit bus and 6GB Vram so It's capable of feeding more CU, but Navi14 also had 24CU and the rest was half of Navi10.

 

[Mopetar]

Drivers suggest IC is tied to memory channels, if there is a cutdown 160 bit N22 (6600 XT 10 GB?) then it should have 80 MB IC.

That raises an interesting question about what's being done with all of the N21 dies with a defective memory controller or infinity cache that necessitates disabling a memory controller. Just from the information we have available the IF and memory controllers account for about 25% of the die space. Although the yields are good, it's still a big die so they should be getting 30-40% defective dies per wafer depending on how much the defect density has changed since TSMC last disclosed it. That's potentially as much as 10 dies per wafer that aren't being sold as even a 6800. I can't see AMD just throwing them out.

They also didn't leave themselves any room in the naming system unless they're going to make a 6800ZZ where the letters designate a lower performance part as opposed to higher performing as is the case of XT. Although AMD has used XT internally for code names for a long time, they hadn't actually used it as part of a product name going back to the days of when they were still ATI up until the release of Navi. Both GT and SE seemed to be popular labels that were used around that time. Do you think we're going to see a 6800SE or something along those lines?

My other theory is that all of those are going to someone like Apple that probably doesn't mind having cut-down silicon since they don't want to put a 300W GPU in an iMac. If that's the case I suspect we might see a 6800 Pro.

 

[Glo.]

BTW I don't understand why N24 has 24CU and not be just 1/4 of the full N21.

All I can tell you at this point is that it is related to the layout of the dies.

I think you'll see the obvious, technical reason for this.

 

[TESKATLIPOKA]

All I can tell you at this point is that it is related to the layout of the dies.

I think you'll see the obvious, technical reason for this.

If the layout is similar to Navi14, then it's because of the placement of dualCU, which is 3 columns and 4 rows of dualCUs for a total of 12 dualCUs. Am I right?
Navi14 dieshot

 

[Glo.]

If the layout is similar to Navi14, then it's because of the placement of dualCU, which is 3 columns and 4 rows of dualCUs for a total of 12 dualCUs. Am I correct?
Navi14 dieshot

No, its not similar.



Infinity cache makes the dies "longer".

 

[TESKATLIPOKA]

Then I am very curious how It looks.

 

[lightmanek]

Then I am very curious how It looks.

A Triangle - more area efficient placement on round wafer

 

[Glo.]

https://twitter.com/i/web/status/1334927336198770690

Yeah, Navi 24 exists, and mobile part really is sub 50W TDP.

 

[maddie]

A Triangle - more area efficient placement on round wafer

Your post might sound very funny, but it reminded me of this article from semiaccurate several years ago.

Disco makes hexagonal and non-regular chips possible

Disco makes hexagonal and non-regular chips possible

Disco makes devices to process wafers such as saws, grinders, and processing tools but the results are quite exceptional.

semiaccurate.com

 

[TESKATLIPOKA]

NV24 XM -> 35-50W
NV24 XML -> 25W

GTX 1650(Max-Q)GDDR5 -> 30-50W
GTX 1650Ti(Max-Q) GDDR6 -> 35-55W
GeForce MX450 GDDDR5,6 -> 25W

Radeon RX 5300M(5500M) -> 85W, clocks: 1000-1445(1375-1645)MHz
Radeon Pro 5300M(5500M) -> 50W, clocks: 1000-1250(1300)MHz

Now the question is, how high can It clock compared to other mobile GPUs within a given power budget.