Question Speculation: RDNA3 + CDNA2 Architectures Thread

Page 183 - 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

PJVol

Senior member
May 25, 2020
707
632
136
AMD releasing RDNA3 Radeon Pros. The W7900 is basically a 7900 XTX with 48 GB of memory; the W7800 is a new spec that has 70 CUs and 256 bit so 32 GB of memory. Possibly comparable to a theoretical 7900 non-XT.
Well, that's probably how consumer parts should have been called, 'cause giving the 84CU part 7900XT name seems to me beyond logic.

the only thing that we really don't know is where the clock speed will end up. I suppose the price is also unknown.
Yeah, just a couple of minor details )
Now it remains to find out whether N32 gpu also exists (whatever it's called), preferably not in the form of a logo
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136
Well, that's probably how consumer parts should have been called, 'cause giving the 84CU part 7900XT name seems to me beyond logic.

It should be obvious - they wanted room for a second cut N31 which would be called the RX 7900.

The amount of GCDs that don't meet the 7900 XT's specs is likely to be very low so a second cut doesn't really make sense unless they were going to pull an nVidia and basically discontinue the 7900 XT for a model which has less MCDs and memory for the current 7900 XT's price.
 

Timorous

Golden Member
Oct 27, 2008
1,748
3,240
136
Yeah, 128 bit bus, the same as NVidia 4060/4060 Ti, really limits them both to 8GB of RAM without making things more complicated.

128bit can handle 16GB okay.

AMD could make 7600 8GB and 7600XT 16GB or they could allow AIBs to offer a 16GB variant. Depends on the price point AMD want to hit. $300 or less and 8GB is fine, more than that and I think they need 16GB (well really 12GB would have been fine but that does not work cleanly with a 128 bit bus).
 

Ajay

Lifer
Jan 8, 2001
16,094
8,109
136
128bit can handle 16GB okay.

AMD could make 7600 8GB and 7600XT 16GB or they could allow AIBs to offer a 16GB variant. Depends on the price point AMD want to hit. $300 or less and 8GB is fine, more than that and I think they need 16GB (well really 12GB would have been fine but that does not work cleanly with a 128 bit bus).
Yeah, they'll be using GDDR6 (not X) to keep the cost down anyway.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,340
5,464
136
128bit can handle 16GB okay.

128 bits is only 4, 32 bit channels. Max size GDDR chips are 2GB on a 32 bit data bus.

2GB x4 = 8GB.

So 8GB is the maximum VRAM with one GDDR chip/channel.

They could do 2 chips/channel, but that's the complication I mentioned. It's a more expensive/complicated board design. So I really doubt they will do this on the lower end Navi 33 cards.

Navi 32 will have 256 bit, 16GB of RAM, and likely a cut down 192 bit, 12GB lower end cut.

Navi 33 with 128 bit would then logically be 8GB.
 

blckgrffn

Diamond Member
May 1, 2003
9,298
3,440
136
www.teamjuchems.com
128 bits is only 4, 32 bit channels. Max size GDDR chips are 2GB on a 32 bit data bus.

2GB x4 = 8GB.

So 8GB is the maximum VRAM with one GDDR chip/channel.

They could do 2 chips/channel, but that's the complication I mentioned. It's a more expensive/complicated board design. So I really doubt they will do this on the lower end Navi 33 cards.

Navi 32 will have 256 bit, 16GB of RAM, and likely a cut down 192 bit, 12GB lower end cut.

Navi 33 with 128 bit would then logically be 8GB.

Is that how the difficult-to-find 8GB 6500xt was able to exist?

It does seem like ram and flash prices have been in a relative free fall, I wonder if that will matter.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,340
5,464
136
Is that how the difficult-to-find 8GB 6500xt was able to exist?

It does seem like ram and flash prices have been in a relative free fall, I wonder if that will matter.

I'm surprised the did this for a consumer card, unless they repurposed the pro board.

The only tear down images I have seen for 6500 were for the 4GB model. It had 2 VRAM chips and no pads for any more chips front or back.
 

blckgrffn

Diamond Member
May 1, 2003
9,298
3,440
136
www.teamjuchems.com
I'm surprised the did this for a consumer card, unless they repurposed the pro board.

The only tear down images I have seen for 6500 were for the 4GB model. It had 2 VRAM chips and no pads for any more chips front or back.


¯\_(ツ)_/¯

I didn't chase it down to see if I could ever buy it.

 

DeathReborn

Platinum Member
Oct 11, 2005
2,776
781
136

Heartbreaker

Diamond Member
Apr 3, 2006
4,340
5,464
136

¯\_(ツ)_/¯

I didn't chase it down to see if I could ever buy it.


I saw the same stuff when I searched, but just out of curiousity, what I really wanted to see was a teardown of an 8GB board, to see where they put the extra VRAM. I guess with only two chips they won't have to resort to placing them on the back side of the board, though that may end up easier for Trace Routing, but I don't know.
 

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136

Laptop with the 7600S review. It's about 13% slower in the raster games they tested than the 4060 they compared it to, with both having the GPU power set to 80 W. I'd say it's not a disaster but not inspiring either.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,522
3,037
136

Laptop with the 7600S review. It's about 13% slower in the raster games they tested than the 4060 they compared it to, with both having the GPU power set to 80 W. I'd say it's not a disaster but not inspiring either.
So the full N33 will provide raster performance comparable to Ad107, but much worse RT performance.
Compared to 6800S(Full N23) the performance is also pathetic, barely any improvement.
Not seeing any reason to choose It over Ada. If It was built using 5(4)nm then at least perf/W would improve significantly, but It is using only the cheap 6nm process.

N32 could have pretty good raster performance, but RT will be another flop. I don't think a laptop with It will be that cheap either.
Screenshot_18.png
 
Last edited:

moinmoin

Diamond Member
Jun 1, 2017
5,064
8,032
136
In many results it seems taking 6800S over 7600S (both at 80W TDP) may actually be slightly preferable. RDNA3 is such an insignificant gen so far.
 

leoneazzurro

Golden Member
Jul 26, 2016
1,052
1,716
136
Quite strange results going on there... I think they made some errors in testing here and there. I.e. the 6800S 80W faster than a 6800M 120W in Cyberpunk? Yeah, sure. Also quite strange the fact that we see such a difference with the same memory amount on AMD vs NV cards. First time I see that, and we should have seen it plenty in the 3060 8Gb/3070 8gb vs the 6600 (which should be affected as well), as it is in more than one game. Quite unique.
In any case, yes, N33 is not a groundbreaking improvement over N23, even if this is not the full die.
 
  • Like
Reactions: Tlh97

ryanjagtap

Member
Sep 25, 2021
132
153
96
Not to be nitpicky, but the 7600S (28 CU) is a replacement for 6700S (28 CU) not 6800S (32 CU).
It has a TDP range of 50-75W. The 90W is already running the AMD equivalent of Nvidia dynamic boost. The 6800S has max TDP of 100W.
AMD has not changed much in N33 die with respect to N23 die. It's still 6nm, so the 25% improvement from just architecture change is ok (zen 2 to zen 3 on same node had 19% improvement). We are still getting the performance of the highest previous gen tier for this 'S' variant of GPU.

Not saying this is a great card, but it's not that bad a card to disparage as well. It's just..... Meh :disappointed:
 
  • Like
Reactions: Schmide and Tlh97

AtenRa

Lifer
Feb 2, 2009
14,003
3,361
136
70CUs would indicate 5 Shader engines enabled with 14CUs per shader engine instead of the full 6 SE + 16CUs per SE of the top model. That is quite a cut down part.

So I presume the spec would be as follows if this cut was also used as a GPU (possibly 7800XT IMO)

64MB L3
16GB GDDR6
70 CUs
70 RTcores
160 ROPs
280 TMUs.
640 GB/s bandwidth if 20 Gbps ram is used

Give it a 2.5-2.6ghz boost clock and it should perform a bit ahead of the 6950XT. Possible it gets released as a 7800XT but I expect AMD would much rather make N32 work as the 7800XT by increasing clocks to 3Ghz as they would be similar in performance and the N32 part would cost less to make.

No way NV32 can match the NV31 with the above specs.

NV32 will be used on the 7700XT and bellow.
 
  • Like
Reactions: Rigg

Timorous

Golden Member
Oct 27, 2008
1,748
3,240
136
No way NV32 can match the NV31 with the above specs.

NV32 will be used on the 7700XT and bellow.

60CUs Vs 70CUs is not that much different and if N32 clocks better performance could be quite close. We are talking a 15% delta so around 2.9Ghz for New Vs 2.5Ghz for the N31 design would give you equal compute, texture fillrate and RT performance.

4 MCDs means both would have the same cache and memory bandwidth to go with that compute.

3 SEs Vs 5SEs means N31 version would have more rops and even with the clockspeed delta it would be ahead in this department but that would only show in cases where you were pixel fillrate limited.

So yea. Both would be pretty close in performance with just a 15% clockspeed advantage for N32 Vs that spec N31.
 
  • Like
Reactions: Tlh97 and Mopetar

PJVol

Senior member
May 25, 2020
707
632
136
60CUs Vs 70CUs is not that much different and if N32 clocks better performance could be quite close. We are talking a 15% delta so around 2.9Ghz for New Vs 2.5Ghz for the N31 design would give you equal compute, texture fillrate and RT performance.
What your assumption that n32 can be clocked higher than n31 at all (let alone +15%) is based upon?
Haven't heard anything to support it yet, and in fact I'm more inclined to believe the initial n32 revision is DOA.

However, the question remains of how much, if any, can be improved with the next revision, and whether it's still commercially viable.
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136
At least at looking at 3dmark's results, the 7900 XTX can hit 2.9 Ghz just fine in a decent amount of samples... but there's a very high deviation. The 4080 and 4090 have a much tighter range.
 

Timorous

Golden Member
Oct 27, 2008
1,748
3,240
136
What your assumption that n32 can be clocked higher than n31 at all (let alone +15%) is based upon?
Haven't heard anything to support it yet, and in fact I'm more inclined to believe the initial n32 revision is DOA.

However, the question remains of how much, if any, can be improved with the next revision, and whether it's still commercially viable.

Smaller dies tend to clock higher in general. N22 clocks higher than N21 by 12.5% if you look at 6750XT vs 6950XT and 23.5% higher if you compare 6750XT vs 6800. If you want to look at original release the 6700XT has a 14.6% clock speed advantage vs the 6900XT

Then there is the fact the N31 seems to have been designed around a higher clock target that they cannot hit at sane power levels in games. If N32 goes some way to fix whatever causes that issue N32 will have a further improved v/f curve.

In addition 7900XTX has an 11% higher boost clock than the 6900XT (using this as neither are refresh parts)

Also checking the specs at Anandtech that W7800 config only has 128 rops so the delta is even less than I 1st though between that N31 config and the supposed N32 config.

So really given the fact smaller dies tend to clock better anyway and the supposed N31 clock vs power flaw and the N21 to N31 clock speed progression a 15% clock speed advantage for N32 vs N31 does not seem at all far fetched and is well within the. Just for reference an 11% increase in clock speed over the 6700XT would be 2.87Ghz which is about the same as a 15% hike from 2.5Ghz so there is that as well.

Given that it does not at all seem far fetched to think N32 can clock a bit higher than N31 and there is possibility to expect it could clock quite a lot higher. Obviously actual execution can be entirely different but ya know, speculation thread so that should always be a given.
 
  • Like
Reactions: Mopetar and Tlh97

TESKATLIPOKA

Platinum Member
May 1, 2020
2,522
3,037
136
Not to be nitpicky, but the 7600S (28 CU) is a replacement for 6700S (28 CU) not 6800S (32 CU).
It has a TDP range of 50-75W. The 90W is already running the AMD equivalent of Nvidia dynamic boost. The 6800S has max TDP of 100W.
AMD has not changed much in N33 die with respect to N23 die. It's still 6nm, so the 25% improvement from just architecture change is ok (zen 2 to zen 3 on same node had 19% improvement). We are still getting the performance of the highest previous gen tier for this 'S' variant of GPU.

Not saying this is a great card, but it's not that bad a card to disparage as well. It's just..... Meh :disappointed:
There was no cut-down N22 in that test, so we compared it to 6800S and both of them consumed ~80W.

I seriously don't know from where you got this nonsense about 25% improvement from just architecture, there is clearly no such thing.

1080pCyberpunk 2077Doom EternalF1 22Far Cry 6Ghostwire TokyoGuardiansSpider-ManAverage
RX 6800S113%118%103%105%101%103%106%107%
RX 7600S100%100%100%100%100%100%100%100%

RX 6800S has 14% more CU(Shader,TMU), but RX 7600S clocks higher.
What I found:
F1 22: 2350/2407MHz avg/max for 7600S at 90W. That should mean ~2300-2325MHz at the lowest.
F1 21: ~2400MHz is max frequency at 105W and then falls by 350MHz to 2050MHz at 80W, at 90W It looks to be ~2150-2175MHz. ComputerBase
So ~2300-2325Mhz vs ~2150-2175Mhz at comparable 90W, that's 7% difference.
My conclusion so far is that, If you compared the same configuration N33 vs N23 at comparable clocks, then performance would be just a bit better.

N33 looks like a FLOP. Why they even designed It is beyond me, at least If they used 5nm process, we would see some improvement in performance and better perf/W at lower TGP.
 
Last edited:

PJVol

Senior member
May 25, 2020
707
632
136
At least at looking at 3dmark's results, the 7900 XTX can hit 2.9 Ghz just fine in a decent amount of samples... but there's a very high deviation
In TimeSpy the frequency is more like ~ specified "boost clock", i.e. 2.5 ghz.
3dmark results tell you nothing about power consumed, so 2.9 may be for heavily OC-ed/tuned/cooled GPU.