Question Speculation: RDNA3 + CDNA2 Architectures Thread

Page 160 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Platinum Member
Oct 16, 2019
2,705
6,427
146

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
It matters because AMD will sell more N32 than N31 due to there being more buyers at $700 than at $900+
So this was your point. Now I understand.

I also made a comparison for N33.
RX 7600 XT: $449
N33 monolith 204mm2 -> 250 dies. 7000/250 = $28, with packaging maybe $40.
BOM: $140
70% of MRSP is $314.

This is the price AMD could charge manufacturers.
RX 7600 XT: $314 - $140 = $174 (39% of MRSP)

Profit per chip:
RX 7600 XT: $314 - $140 - $40 = $134
That's just $25 less than N32.

Product margin = (selling price – cost of product) / selling price.
RX 7600 XT: (174-40)/174 = 77% :eek:

If the MRSP is only $399:
RX 7600 XT: $279-$140=$139 -> 139-40=$99 -> (139-40)/139=71%
If the MRSP is only $349:
RX 7600 XT: $244-$140=$104 -> 104-40=$64 -> (104-40)/104=61.5%
If the MRSP is only $299:
RX 7600 XT: $209-$140=$69 -> 69-40=$29 -> (69-40)/69=42%
If the MRSP is only $269:
RX 7600 XT: $188-$140=$44 -> 44-40= $4 -> (44-40)/44=9%
If the MRSP is only $249:
RX 7600 XT: $174-$140=$34 -> 34-40= -$6 loss

If AMD is really greedy, then RX 7600 XT will cost $449.
If AMD is greedy, then RX 7600 XT will cost $399.
If AMD is sensible, then RX 7600 XT will cost $349.
If AMD wants to make a very good impression, then RX 7600 XT will cost $299.
If AMD wants to start a price war, then RX 7600 XT will cost $269.
If AMD wants to lose money, then RX 7600 XT will cost $249. :p :D
 
Last edited:

Timorous

Golden Member
Oct 27, 2008
1,748
3,240
136
So this was your point. Now I understand.
It's better to have higher margin for the cheaper product because you sell more of them and profit more.

I also made a comparison for N33.
N33: $449
N33 monolith 204mm2 -> 250 dies. 7000/250 = $28, with packaging maybe $40.
BOM: $140
70% of MRSP is $314.

This is the price AMD could charge manufacturers.
RX 7600 XT: $314 - $140 = $174 (39% of MRSP)

Profit per chip:
RX 7600 XT: $314 - $140 - $40 = $134
That's just $25 less than N32.

Product margin= (selling price – cost of product) / selling price.
RX 7600 XT: (174-40)/174 = 77%

If the MRSP is only $399:
RX 7600 XT: $279-$140=$139 -> 139-40=$99
RX 7600 XT: (139-40)/139=71%

If the MRSP is only $349:
RX 7600 XT: $244-$140=$104 -> 104-40=$64
RX 7600 XT: (104-40)/104=61.5%

If the MRSP is only $299:
RX 7600 XT: $209-$140=$69 -> 69-40=$29
RX 7600 XT: (69-40)/69=42%

I expect desktop N33 to be more like $350 but I expect AMD are charging less for laptop OEMs to get more design wins which is probably good in the long run.
 
  • Like
Reactions: Kaluan

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Why would be the margin for N32 higher than N31?
Price: $699 vs $999
That's a difference of $300.

I calculated the cost for making N31 and N33.
Wafer prices are 7nm at that time $8000, 5nm currently $15000, 6nm currently $7000.
MCD is 37mm2 -> 1547 good dies. 7000/1547 = ~ $4.5
N31 GCD is 300mm2 -> 160 good dies. 15000/160 = $94 + 6*$4.5 = $121, with packaging maybe $155.
N31(cut) GCD is 300mm2 -> 160 good dies. 15000/160 = $94 + 5*$4.5 = $116.5, with packaging maybe $148.
N32 GCD is 200mm2 -> 251 good dies. 15000/251 = $60 + 4*$4.5 = $78, with packaging maybe $100.
GDDR6 is $20 per 2GB GDDR6 module included in BOM.

MRSP price will be divided: 8% to shop, 3% shipping, 19% manufacturer margin, you will be left with 70% of MRSP for BOM and AMD's asked price for the chip.
Chip costBOMMSRP70% of MRSP
RX 7900 XTX$155$330$999$699
RX 7900 XT$148$290$899$629
RX 7800 XT$100$230$649-699$489
AMD can set the price for manufacturers per chip at:
RX 7900 XTX: $699 - $330 = $369 (36.9% of MRSP)
RX 7900 XT: $629 - $290 = $339 (37.7% of MRSP)
RX 7800 XT: $489 - $230 = $259 (37% of MRSP)

Profit per chip:
RX 7900 XTX: $699 - $330 -$155 = $214
RX 7900 XTX: $629 - $290 -$148 = $191
RX 7800 XT: $489 - $230 - $100 = $159

Both N31 based models are more profitable than N32 for AMD.

This doesn't make sense in reality. You aren't including the amount required to recoup R&D, or marketing, or any of the other costs involved with selling a product. It also assumes 100% chip yields. The profit is likely no where close to what you have listed here.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
This doesn't make sense in reality. You aren't including the amount required to recoup R&D, or marketing, or any of the other costs involved with selling a product. It also assumes 100% chip yields. The profit is likely no where close to what you have listed here.
I didn't assume 100% chip yields, but It's true I used only the good ones, the faulty ones were considered unusable, although in reality many should be reusable.

This is pure speculation, It's not like I know how much wafer or packaging cost, the same with BOM per card. How much is transport, cut for manufacturer, shop etc.
This was just an illustration how profit could look per different GPU at different price.
Even If It's not realistic, I think It's still an interesting read.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
Tldr; "wait for the drivers" or in other words "buy the performace in future (61tflops)". Hope driver will learn to pack vopds by the rdna4 launch )
I think everyone is already accustomed to how mature AMD drivers are at launch. ;)
This is especially true for this generation, where they did a mayor change inside their execution units.
Performance will be certainly better, but we don't know by how much and when.:D

Nvidia supposedly working on AI-optimized drivers. I think AI could help AMD do faster and better driver updates. :p
 
Last edited:

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
Chips and Cheese posted an article where they microbenched RDNA 3. Enjoy!
Fantastic article. This is why I love Chips and Cheese. :)

RDNA3 is not a bad architecture, but that dual issue is simply useless If It's left to the compiler.
Games will need heavy optimizations done by AMD's driver team to make this architecture "shine". This doesn't look optimistic.
I really have to wonder why AMD choose this path, when they knew how much work It needs to work correctly.

If the code is optimized correctly for VOPD instructions, then the improvement in TFLOPs is ~100%.
Screenshot_33.png
Curiously, RDNA 3 actually regresses in FP64 throughput. We already saw a hint of this earlier with OpenCL, where one RDNA 2 WGP could execute eight FP64 operations per cycle. RDNA 3 cuts throughput in half, meaning a WGP can do four FP64 operations – probably one per SIMD, per cycle.

Naturally, this whole optimization is by no means an easy task, but It should result in significant improvements in games in my opinion.
 
Last edited:

Aapje

Golden Member
Mar 21, 2022
1,515
2,065
106
RDNA3 is not a bad architecture, but that dual issue is simply useless If It's left to the compiler.
Games will need heavy optimizations done by AMD's driver team to make this architecture "shine". This doesn't look optimistic.
I really have to wonder why AMD choose this path, when they knew how much work It needs to work correctly.

Perhaps they are also working on an AI compiler :p

Seriously though, it might work very well to run games with a dynamic compiler like Java commonly uses and then have that dynamic compiler try out different optimizations and then measure the effect on frametimes. Then they can use this to create 'compiler hints' that are used when compiling the code that is actually shipped.

Unless they already do compilation on the consumer's machine, which the "shader compilation" messages that we commonly see nowadays suggests. If they do that, they can potentially ship the 'compiler hints' with the driver.
 
  • Like
Reactions: TESKATLIPOKA

leoneazzurro

Golden Member
Jul 26, 2016
1,052
1,716
136
Now when they already launched them? I don't think they will unlaunch them just because I think their naming sense is pretty bad. :D


4.5% difference in clockspeed 7600M XT vs 7700S is because that is at their highest TDP 100W vs 120W.
Their TDP range is:
7600M XT: 75-120W
7700S: 75-100W
That means 7600M XT can also be limited to 100W.

If you compared 2 laptops, one with 7600M XT 100W vs 7700S 100W. How will they perform? Won't they perform the same?
The same with 7600M vs 7600S
Unless those S models are binned N33 then these models have no real reason to exist.
For 3 models you can guess the specs, but 7700S has simply wrong name. For no good reason, It's a higher tier.

Yes they will perform the same (if no clock limitation on one of the parts will kick in, that is, 7700S may be even more limited than a 7600M in this regard) and no one said two N33 at same TDP and clocks will perform differently lol. What I said is that these two variants have different TDP range settings and slight game clock differences, so while they can be set at the same value and have the same performance, in general they are not so generally they will perform (slightly) differently. And not only, they are in theory aimed at different market segments in the same way Nvidia Max-Q does (with similar issues in some cases as many Max-Q cards are quite TDP limited). Now this for us users may mean nothing, but for OEMs could mean different requirements about cooling solutions, voltage supply characteristics, and so on. This may be enough for a different name? Maybe not, but this is more a commercial issue. Also, it is clear that N33 is the replacement for N23, but N23 was also the 6800S, while here we have only up to the 7700S. This means that if we will have a 7800S, this will be N32 based (and so a 7700M and 7800M, if they will come).
 
Last edited:

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
Yes they will perform the same (if no clock limitation on one of the parts will kick in, that is, 7700S may be even more limited than a 7600M in this regard) and no one said two N33 at same TDP and clocks will perform differently lol. What I said is that these two variants have different TDP range settings and slight game clock differences, so while they can be set at the same value and have the same performance, in general they are not so generally they will perform (slightly) differently. And not only, they are in theory aimed at different market segments in the same way Nvidia Max-Q does (with similar issues in some cases as many Max-Q cards are quite TDP limited). Now this for us users may mean nothing, but for OEMs could mean different requirements about cooling solutions, voltage supply characteristics, and so on. This may be enough for a different name? Maybe not, but this is more a commercial issue. Also, it is clear that N33 is the replacement for N23, but N23 was also the 6800S, while here we have only up to the 7700S. This means that if we will have a 7800S, this will be N32 based (and so a 7700M and 7800M, if they will come).
I can agree with this.
 

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
Fantastic article. This is why I love Chips and Cheese. :)

RDNA3 is not a bad architecture, but that dual issue is simply useless If It's left to the compiler.
Games will need heavy optimizations done by AMD's driver team to make this architecture "shine". This doesn't look optimistic.
I really have to wonder why AMD choose this path, when they knew how much work It needs to work correctly.

If the code is optimized correctly for VOPD instructions, then the improvement in TFLOPs is ~100%.
View attachment 74208


Naturally, this whole optimization is by no means an easy task, but It should result in significant improvements in games in my opinion.
This is an area where using AI to optimize could be a big help. Train a model to recognize a frame, then implement various tweaks and see if output = input while fps > old fps.
By default shaders are compiled in the driver while the assets are loading. Sometimes they are compiled on the fly (which is usually bad and causes stuttering), the compiled shaders are also often cached locally to speed up future loads.

By default, sure, but precompiled, distributed shader assets are a thing.

The amazing things AMD could do if they focused more on software. Imagine if they created a Radeon specific shader modification/replacement toolkit? They include a bunch of highly optimized shaders out of the box and let the community do the rest. They could distribute it via steam and use Steam workshop for user content.

God forbid they think outside the box.

Side note: Maybe the x50 refresh will be RDNA3+?
 

Kepler_L2

Senior member
Sep 6, 2020
537
2,199
136
This is an area where using AI to optimize could be a big help. Train a model to recognize a frame, then implement various tweaks and see if output = input while fps > old fps.


By default, sure, but precompiled, distributed shader assets are a thing.

The amazing things AMD could do if they focused more on software. Imagine if they created a Radeon specific shader modification/replacement toolkit? They include a bunch of highly optimized shaders out of the box and let the community do the rest. They could distribute it via steam and use Steam workshop for user content.

God forbid they think outside the box.

Side note: Maybe the x50 refresh will be RDNA3+?
RDNA3+ is just for APUs.
 

PJVol

Senior member
May 25, 2020
708
632
136
Last edited:
  • Like
Reactions: Vattila

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,038
136
If RDNA3+ is only for APUs, then a good question is why?
Because It's not financially or based on performance worth It to make new GPUs based on RDNA3+ or RDNA4 would be too close to It to make any sense.
 
Last edited:
  • Like
Reactions: Kaluan

Kaluan

Senior member
Jan 4, 2022
504
1,074
106
Where did you get that? N31 is 1100, N32 - 1102, N33 - 1103.
Anyway, 1101 and 1104 are APUs:
https://elixir.bootlin.com/linux/v6.2-rc3/source/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c#L2213
C++:
switch (adev->ip_versions[GC_HWIP][0]) {
    case IP_VERSION(9, 1, 0):
         ...
    case IP_VERSION(10, 3, 7):
    case IP_VERSION(11, 0, 1):
    case IP_VERSION(11, 0, 4):
        adev->flags |= AMD_IS_APU;
        break;
    default:
        break;
}
Ah, my mistake.

But do note that "IP version" and "GFX ID" are not the same from what we can see.

N31 may be both GFX1100 and IPver 11.0.0, but Phoenix is GFX1103 but IPver 11.0.1