Discussion AMD Cezanne/Zen 3 APU Speculation and Discussion

Page 20 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

dr1337

Member
May 25, 2020
199
333
106
Fresh leak out today, not much is known but at least 8cu's is confirmed. Probably an engineering sample, core count is unknown and clocks may not be final.

This is very interesting to me because cezanne is seemingly 8cu only, and it seems unlikely to me that AMD could squeeze any more performance out of vega. A cpu only upgrade of renoir may be lackluster compared to tigerlake's quite large GPU.

What do you guys think? Will zen 3 be a large enough improvement in APU form? Will it have full cache? Are there more than 8cus? Has AMD truly evolved vega yet again or is it more like rdna?
 

QueBert

Lifer
Jan 6, 2002
21,992
452
126
I would love to test out one of these, I just got a laptop with a 5700u last month and the apu's really impressed me so far. And I know since it's Zen2+ the Zen3 ones would perform a decent amount better. 900p on my 13.3" screen's fine with me. Since integrated graphics came into existence, I remember them always being junk for gaming unless it was super old ones. Apparently, you can play Cyberpunk 2077 on a 5700U with low'ish, but playable framerates, which is absolutely crazy shit to me lol. Will probably upgrade to a 5800U laptop when they're actually available.
 

Insert_Nickname

Diamond Member
May 6, 2012
4,825
1,494
136
ill post results probably tomorrow. One thing i can say for sure is that Hardware Unboxed was spot on with the memory, dual rank is the way to go, everything that is 3200 or faster will perform too close to each other, not worth the money. Dual rank, or use 4 single rank sticks. I only had one dual rank 3200 kit, and for a short while today, after that i started using 2, 2x8 DDR4-3600 kits, that was the faster combination, faster than DDR4-4000 single rank. Oh, DDR4-4000 is the max you can use, after that it goes async and performance drops.
Have you tried triple or quad rank? I'd be interested in how that performs compared to regular dual rank. You may not be able to get much more then 3000MHz'ish out of it, but it should perform very well in practice. Nobody ever seems to test that.

AMD CPU/APUs have always loved dual rank memory. So I'm not surprised it performs well on Cezanne. I'm running triple rank (16+8GB on each channel) on my 3600 at 3066MHz, it's a nice compromise between speed and the number of ranks.
 

Shivansps

Diamond Member
Sep 11, 2013
3,641
1,329
136
Have you tried triple or quad rank? I'd be interested in how that performs compared to regular dual rank. You may not be able to get much more then 3000MHz'ish out of it, but it should perform very well in practice. Nobody ever seems to test that.

AMD CPU/APUs have always loved dual rank memory. So I'm not surprised it performs well on Cezanne. I'm running triple rank (16+8GB on each channel) on my 3600 at 3066MHz, it's a nice compromise between speed and the number of ranks.
The thing is i dont have two fast 16GB sticks to test unfortunately, i cant just go and take wharever ram i want. So i only had 8GB sticks and they are all single rank these days, unless someone have a old kit laying around with 2R 8GB sticks, this is going to be the case for everyone running Cezanne, so, going 4 ranks means going to 32GB.

I tested:
-2x8 DDR4-3200 (2R) + IGPU Stock
-2x8 DDR4-3200 (2R) + IGPU OC
-2x8 DDR4-3466 (2R) + IGPU Stock
-2x8 DDR4-3466 (2R) + IGPU OC
-2x8 DDR4-3600 (2R) + IGPU OC
-2x8 DDR4-4000 (2R) + IGPU OC
-4x8 DDR4-3600 (4R) + IGPU Stock
-4x8 DDR4-3600 (4R) + IGPU OC
 

NTMBK

Diamond Member
Nov 14, 2011
9,975
4,356
136
This really does support Shivas' point that Vega on Renoir/Cezanne was shrunk too much. It might be barely sufficient on mobile, with the usually more constrained ram configurations, but it's not doing the desktop APUs any favors at all. I hope that the next generation of APUs on mobile get far more generous configurations. Having significantly more bandwidth and interleaving of transfers on DDR5 should justify at least 8 CUs of RDNA2 (16 Vega equivalents).
The market doesn't really care about desktop APU performance. If you want graphics power, you put in a GPU.

It's massively important in a power, heat, and size constrained laptop, so that is rightly where they put the effort. On desktop they just need it to work well enough for non-gaming systems.
 

Abwx

Diamond Member
Apr 2, 2011
9,814
2,250
136
5600G and 5700G review :


With a specific review for the iGPU :

 

Shivansps

Diamond Member
Sep 11, 2013
3,641
1,329
136
Looks like Computerbase reached to the same opinion as i, DDR4-4000(the max you can use) is faster, but not worth it with these IGPs. The same with the OC potential, its there, but it does very little...

But it is definatelly worth it to use faster than DDR4-3200(DDR4-3466 or 3600 is what i would recomend) AND OC at the same time, it is not a big jump but having both it helps a lot to keep some games over 30.

As for the other thing, lets not start that again, Computerbase is right to point out these APU are around 30% below the RX 460, just think about what could have been. Now AMD is going big again with RMB, no point in arguing this, wharever the reason for the shrunk on Renoir, they are now undoing it.
 
Last edited:

maddie

Diamond Member
Jul 18, 2010
4,334
3,873
136
Looks like Computerbase reached to the same opinion as i, DDR4-4000(the max you can use) is faster, but not worth it with these IGPs. The same with the OC potential, its there, but it does very little...

But it is definatelly worth it to use faster than DDR4-3200(DDR4-3466 or 3600 is what i would recomend) AND OC at the same time, it is not a big jump but having both it helps a lot to keep some games over 30.

As for the other thing, lets not start that again, Computerbase is right to point out these APU are around 30% below the RX 460, just think about what could have been. Now AMD is going big again with RMB, no point in arguing this.
What are you on about again?

All I see is strong evidence that dual channel DDR4 is a roadblock to iGPU performance gains, barring an infinity cache solution.
 
  • Like
Reactions: spursindonesia

LightningZ71

Golden Member
Mar 10, 2017
1,483
1,694
136
With stacking, that's not impossible. That being said, the improvement in ram efficiency in RDNA, combined with the higher bandwidth and greater utilization of it in DDR5, should make even a modestly expanded iGPU achieve similar performance to an RX560. At 1080p, that's not bad, and it won't even require an infinity cache setup.

One thing that is absolutely holding back the APUs is the L3 size. Doubling it again should make a big difference all-around, and with N5, that's not impossible.
 

blckgrffn

Diamond Member
May 1, 2003
8,598
2,121
136
www.teamjuchems.com
With stacking, that's not impossible. That being said, the improvement in ram efficiency in RDNA, combined with the higher bandwidth and greater utilization of it in DDR5, should make even a modestly expanded iGPU achieve similar performance to an RX560. At 1080p, that's not bad, and it won't even require an infinity cache setup.

One thing that is absolutely holding back the APUs is the L3 size. Doubling it again should make a big difference all-around, and with N5, that's not impossible.
Having used an RX460 at 1080p I can tell you that's not a very satisfying bar to clear, especially heading in 2022. Let alone heading into, what 2023, when we actually see N5 APUs?

I hear what you are saying, but the ~35% or so increase in bandwidth moving to DDR5 is "meh" compared to what the IC could do when coupled with that same bandwidth at 1080p. We'll see what the 6600XT can do with it.

FSR like tech is probably going to be key to many titles and 1080p APU performance in AAA titles... probably indefinitely? I am sure that AMD likes selling GPUs too.
 
  • Like
Reactions: Tlh97

NTMBK

Diamond Member
Nov 14, 2011
9,975
4,356
136
Well . . . some downmarket consumers do. They don't bring as much cash to the table, so under the present circumstances where dice need to be sold at a premium, their needs will not be met first (or perhaps at all).
Is it worth compromising the design of a laptop part to suit that tiny, low profit niche? I suspect not.
 

Shivansps

Diamond Member
Sep 11, 2013
3,641
1,329
136
What are you on about again?

All I see is strong evidence that dual channel DDR4 is a roadblock to iGPU performance gains, barring an infinity cache solution.
How do you explain that the 5700G is faster than the 5600G in IGP then? Check Computerbase numbers, they used 2x8 DDR4-3200 what is the worst possible option for bandwidth, and the 5700G is still ahead. It does not means you are not going to gain FPS with faster ram or dual rank sticks, you do, but it is not as much as you may think, the 5700G has a lead, small, but still a lead, specially considering it should also gain more from faster ram.
 

maddie

Diamond Member
Jul 18, 2010
4,334
3,873
136
How do you explain that the 5700G is faster than the 5600G in IGP then? Check Computerbase numbers, they used 2x8 DDR4-3200 what is the worst possible option for bandwidth, and the 5700G is still ahead. It does not means you are not going to gain FPS with faster ram or dual rank sticks, you do, but it is not as much as you may think, the 5700G has a lead, small, but still a lead, specially considering it should also gain more from faster ram.
Tiny disproportionate gains really is grasping at straws. The RX 560 is about 50-75% faster than this APU. Do you really see that performance easily achievable with DDR4? For any meaningful increase in APU graphics performance, there has to be a way to significantly increase effective memory access, either through increased raw memory bandwidth ( high speed DDR5, Quad channel memory) or through caching mechanisms, such as the recently introduced infinity cache ( probably best choice).

Expecting large gains by doing more of the same is wishful thinking in my opinion.
 
  • Like
Reactions: Tlh97 and prtskg

Shivansps

Diamond Member
Sep 11, 2013
3,641
1,329
136
Tiny disproportionate gains really is grasping at straws. The RX 560 is about 50-75% faster than this APU. Do you really see that performance easily achievable with DDR4? For any meaningful increase in APU graphics performance, there has to be a way to significantly increase effective memory access, either through increased raw memory bandwidth ( high speed DDR5, Quad channel memory) or through caching mechanisms, such as the recently introduced infinity cache ( probably best choice).

Expecting large gains by doing more of the same is wishful thinking in my opinion.
I only said RX 460, but look at this, im going to put all data on a graph later, but this does not looks like ONLY memory bound to me:

DDR4-3200 w/IGP STOCK

DDR4-3200 w/IGP OC

DDR4-4000 w/IGP OC

4xDDR4-3600 w/IGP OC

4xDDR4-3600 w/IGP STOCK
 
Last edited:
  • Like
Reactions: lightmanek

DrMrLordX

Lifer
Apr 27, 2000
20,285
9,335
136
Is it worth compromising the design of a laptop part to suit that tiny, low profit niche? I suspect not.
Low profit, probably, tiny niche, not really. The world is a much bigger place than the US or Europe and silicon shortages won't last forever.
I'm inclined to agree with @maddie . Eventually AMD will find themselves with access to N7/N6 wafers and a shrinking family of products that need it. That day is not today, though.
 
  • Like
Reactions: Tlh97

Abwx

Diamond Member
Apr 2, 2011
9,814
2,250
136
I'm inclined to agree with @maddie . Eventually AMD will find themselves with access to N7/N6 wafers and a shrinking family of products that need it. That day is not today, though.
5700G die is 180 mm2, not sure what need to be shrinked since they can get something like 320 functional dies out of a waffer, dunno the exact waffer cost but seems that it got below 8000$ when Vermeer was released.



 

moinmoin

Diamond Member
Jun 1, 2017
4,043
6,080
136
5700G die is 180 mm2, not sure what need to be shrinked
To reach the low budget "niche" market AMD would need to shrink the profit margin which they'll try to avoid doing as long as the products sell perfectly fine at the current prices and margins. ;)

Once supply beats demand (whenever that may be) lower margin products will reappear.
 

blckgrffn

Diamond Member
May 1, 2003
8,598
2,121
136
www.teamjuchems.com
To reach the low budget "niche" market AMD would need to shrink the profit margin which they'll try to avoid doing as long as the products sell perfectly fine at the current prices and margins. ;)

Once supply beats demand (whenever that may be) lower margin products will reappear.
Also, the "shrinking need" need for N7/N6 will be because they've moved tons of products to N5 but there is wafer capacity available on the previous Node.

AMD in growth mode needs to be smart about growing their market share with the wafers they can get, not the wafers they necessarily want.
 
  • Like
Reactions: Tlh97

Asterox

Senior member
May 15, 2012
960
1,626
136
5700G die is 180 mm2, not sure what need to be shrinked since they can get something like 320 functional dies out of a waffer, dunno the exact waffer cost but seems that it got below 8000$ when Vermeer was released.



For comparisson, Renoir R7 Pro 4750G has 156mm2 die size.

 
  • Like
Reactions: Abwx and Tlh97

KompuKare

Senior member
Jul 28, 2009
853
558
136
5700G die is 180 mm2, not sure what need to be shrinked since they can get something like 320 functional dies out of a waffer, dunno the exact waffer cost but seems that it got below 8000$ when Vermeer was released.



I though earlier estimates had it at 175mm², but I guess that 180mm² is "rund 180 mm² große Chip".

Anyway, plugging the SqrtRt of 180 (13.4mm) into this:
wafer yield calculator with a defect rate of 0.09 per cm² gives us this (TSMC gave that defect rate a good while back so it should hopefully be a lot better now):

Your 320 figure was assuming no defects.
The calculator expects 269 good dies per wafer although not all the 47 'defect' dies would be waste as with some of them they could disable a CU or core, etc.
On the other hand, not all good dies will bin at the kind of clocks or voltages they want, but since Cezanne isn't a single SKU console those should be useable somewhere.
$8000 is a bit optimistic, 7nm wafer are meant to be more like $12000+.
Anyway while wafers are scares, I guess the comparison for AMD is how many CCDS could they make instead?
 
  • Like
Reactions: Mopetar

DrMrLordX

Lifer
Apr 27, 2000
20,285
9,335
136
5700G die is 180 mm2, not sure what need to be shrinked since they can get something like 320 functional dies out of a waffer, dunno the exact waffer cost but seems that it got below 8000$ when Vermeer was released.
I think you misunderstand what I was saying. Someday they're gonna move all their high-profit stuff off N7/N6. That's what I meant by "shrinking family of products". I did not imply that anything in particular needed a die shrink.
 

Abwx

Diamond Member
Apr 2, 2011
9,814
2,250
136
To reach the low budget "niche" market AMD would need to shrink the profit margin which they'll try to avoid doing as long as the products sell perfectly fine at the current prices and margins. ;)

Once supply beats demand (whenever that may be) lower margin products will reappear.
APUs prices are somewhat unrealistic these days given that they are supposed to cover the low cost segment.

Hope thIs will filled either with Renoir, wich would allow for a 8C, but that s doubtfull since it would eat on the 7nm budget, or more likely with Van Gogh.
I though earlier estimates had it at 175mm², but I guess that 180mm² is "rund 180 mm² große Chip".

Anyway, plugging the SqrtRt of 180 (13.4mm) into this:
wafer yield calculator with a defect rate of 0.09 per cm² gives us this (TSMC gave that defect rate a good while back so it should hopefully be a lot better now):

Your 320 figure was assuming no defects.
The calculator expects 269 good dies per wafer although not all the 47 'defect' dies would be waste as with some of them they could disable a CU or core, etc.
On the other hand, not all good dies will bin at the kind of clocks or voltages they want, but since Cezanne isn't a single SKU console those should be useable somewhere.
$8000 is a bit optimistic, 7nm wafer are meant to be more like $12000+.
Anyway while wafers are scares, I guess the comparison for AMD is how many CCDS could they make instead?
Actually i did put the recoverable dies at 90%, and when i said functional dies this include the 6C/7CU parts where either a core or a CU has been disabled due to a defect in the silicon.

FI if the GPU is 40% of the area then 40% of the defects will be concentrated in the GPU, wich can be neutered by disabling a CU, and so on with defects that are on a core out of 8.

As for the waffer price it was stated by some exec (Lisa Su or some TSMC head) as shifting below 8k from the 12k or so when Zen 2 was released.
 

moinmoin

Diamond Member
Jun 1, 2017
4,043
6,080
136
I though earlier estimates had it at 175mm², but I guess that 180mm² is "rund 180 mm² große Chip".
"rund 180 mm²" translates to "around 180 mm²" which means it's a fuzzy rounded number, 175 mm² may still be more correct.

APUs prices are somewhat unrealistic these days given that they are supposed to cover the low cost segment.
That's your expectation. These APUs are clearly not part of the low cost segment though. Why would they? Honestly it's insane enough that people expect a bigger monolithic chip with iGPU to be cheaper than a similar featured MCM chip without iGPU. The only reason this still makes somewhat sense for these particular chips is the significantly cut L3$ and lack of PCIe 4, and they are cheaper than the X-counterparts as a result.
 

ASK THE COMMUNITY