Discussion AMD Cezanne/Zen 3 APU Speculation and Discussion

Page 23 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

dr1337

Senior member
May 25, 2020
311
515
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?
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
There is a Linux driver patch that points out to a new AMD APU (AMD Cyan Skillfish) with a RDNA1 IGP, that looks like the new Athlon to me. Now, why use RDNA1 when they are already making RDNA2 apus i have no idea. A 4C/8T with 6 RDNA1 CUs looks good to me to replace Dali. Keep in mind that RMB is likely to come in 12,10 and 8CU SKUs.

I feel that this is heavily based on the lower transistor count required for an RDNA1 CU and the pointlessness of having ray tracing capabilities in such a low end part. 6CU of RDNA1 on 12LP (which will result in lower clock rates) should be more than capable of keeping pace with the 48eu base configs of the TigerLake and expected Alderlake parts that wind up in the low end of the market. It might have issues with the 80/96EU Xe parts though, however, competitiveness there will depend on price, and it remains to be seen if Intel will discount the 11th gen quad core parts as lower level 12th gen parts.

I feel that this will be a decent successor to the 3400G/3500U/3700U parts with a reasonable performance uplift all around.
 

Panino Manino

Senior member
Jan 28, 2017
813
1,010
136
There is a Linux driver patch that points out to a new AMD APU (AMD Cyan Skillfish) with a RDNA1 IGP, that looks like the new Athlon to me. Now, why use RDNA1 when they are already making RDNA2 apus i have no idea. A 4C/8T with 6 RDNA1 CUs looks good to me to replace Dali. Keep in mind that RMB is likely to come in 12,10 and 8CU SKUs.

Seems obvious to me.
RDNA1 CUs are smaller than RDNA2 CUs, it's cheaper to make.
As much as I wish to see APUs with RDNA2 I believe that they'll be very rare and late to come.
 

blckgrffn

Diamond Member
May 1, 2003
9,111
3,029
136
www.teamjuchems.com
Seems obvious to me.
RDNA1 CUs are smaller than RDNA2 CUs, it's cheaper to make.
As much as I wish to see APUs with RDNA2 I believe that they'll be very rare and late to come.

Makes sense, especially with a larger process that is value oriented.

I am going to trust that this is more RNDA 1.1 and they tweaked some of the issues they fixed moving from RDNA1 to RDNA2 given the chance to lay it out again. We've been told the console APUs are somewhere between RNDA 1 and RDNA 2 so there is likely some wiggle room to the actual iGPU version and features. Maybe. :)

I guess I am not that worried about competitiveness against Xe. If it can be as good or better than 3400G iGPU performance but be available at good (low!) prices and isn't just some embedded SFF OEM only part I'll take that as a win.
 

tomatosummit

Member
Mar 21, 2019
184
177
116
Seems obvious to me.
RDNA1 CUs are smaller than RDNA2 CUs, it's cheaper to make.
As much as I wish to see APUs with RDNA2 I believe that they'll be very rare and late to come.
Aren't the gpu feature sets the decider for rdna1/2 branding. So the igpu could still have various improvements from rdna2 design but not meet all the criteria. Same as the ps5 not being all the way to rdna2 and it not particularly mattering.
Late era GCN graphics cards where a hodgepodge of features sets in a similar example.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
Makes sense, especially with a larger process that is value oriented.

I am going to trust that this is more RNDA 1.1 and they tweaked some of the issues they fixed moving from RDNA1 to RDNA2 given the chance to lay it out again. We've been told the console APUs are somewhere between RNDA 1 and RDNA 2 so there is likely some wiggle room to the actual iGPU version and features. Maybe. :)

I guess I am not that worried about competitiveness against Xe. If it can be as good or better than 3400G iGPU performance but be available at good (low!) prices and isn't just some embedded SFF OEM only part I'll take that as a win.

The only thing that I find to be a head scratcher is that, due to the Zen2 core being a good bit larger in transistor count than Zen1/+, and the RDNA CUs being larger than Vega CUs, this die must be nearly as large as Zeppelin (Zen1/+ 8 core). Unless the cost per wafer for 12LP wafers has come down a lot, or this is a 12LP+ product with the increased transistor density it brings, it's not going to be especially inexpensive to make.

The other question is, what are the target clocks? If we go by what was achieved by the 3400g, this will likely cap at 4.3Ghz or so and the iGPU won't make it much past about 1500mhz.
 
  • Like
Reactions: Tlh97 and blckgrffn

blckgrffn

Diamond Member
May 1, 2003
9,111
3,029
136
www.teamjuchems.com
The only thing that I find to be a head scratcher is that, due to the Zen2 core being a good bit larger in transistor count than Zen1/+, and the RDNA CUs being larger than Vega CUs, this die must be nearly as large as Zeppelin (Zen1/+ 8 core). Unless the cost per wafer for 12LP wafers has come down a lot, or this is a 12LP+ product with the increased transistor density it brings, it's not going to be especially inexpensive to make.

The other question is, what are the target clocks? If we go by what was achieved by the 3400g, this will likely cap at 4.3Ghz or so and the iGPU won't make it much past about 1500mhz.

I guess I am just assuming that it is at the latest process that GloFo can bring to bring to the table because I would think that GloFo would want a good customer for that investment and AMD would want the density improvements.

The way I see it, this is a market share play more than a filling a niche in the current product offering type move.

Given the global climate getting more sellable CPUs from any source possible gives you the ability to simply sell more units and use the wafers available across all sources in a way that creates the most revenue.

I'll be interested to see if any current motherboards get support for this CPU officially. I don't think we've seen it called out yet, right? I feel like if we do start seeing it in firmware releases - especially for B550 and 570 chipsets, we'll have a reasonable shot to see this in the DIY space in some volume.

Otherwise it might just be devoured by some OEM. I have no idea on the type of volume GloFo can offer here.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
The only thing that I find to be a head scratcher is that, due to the Zen2 core being a good bit larger in transistor count than Zen1/+, and the RDNA CUs being larger than Vega CUs, this die must be nearly as large as Zeppelin (Zen1/+ 8 core). Unless the cost per wafer for 12LP wafers has come down a lot, or this is a 12LP+ product with the increased transistor density it brings, it's not going to be especially inexpensive to make.

The other question is, what are the target clocks? If we go by what was achieved by the 3400g, this will likely cap at 4.3Ghz or so and the iGPU won't make it much past about 1500mhz.

Just a little note, there is no mention of the IGP CU count, 6CU is what i think should be right considering RMB CU count and what it should face at low end. It could be a lot lower than that.

I think it could be a backport of Van Gogh to 12LP, with a RDNA1 GPU instead of RDNA2 and probably a little smaller, that why i said 6CU. But it could be 5 or 4.
The max they will be able to sell it on DIY for a fully enabled part would be around $110 vs the I3s, a.k.a 3200G replacement. The lowest RMB should take the 3400G place, 4C/8T are not going to be enoght for that anymore.

AMD always uses the older socket for the low end for the first year of a new platform, what would probably means the 5300G and 3000G will be the low end for the first year of AM5. So the new low end APU is likely DDR5 only.
 
Last edited:
  • Like
Reactions: Tlh97

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
That's a lot of "what if". This has to be a very low end product or a very custom part for an OEM with a big order. I'm thinking that this is their next gen for their Chromebook and sub $350 notebook product space. What they have there now is just technically uninteresting vs. the 1005G1 and any of the TigerLake products. Given that 12LP+ is what GF considers leading edge, I doubt that they are giving it away. Regular 12LP is good for its era, but, eh, for a low end product it should be fine. But, that leaves this in a weird position. I think that AMD's choice here was to use precious N7/N6 volume for a cut-down part (perhaps with the same general specs) that can give a lot of units per wafer, or what we expect here. The difference is, I guess, that N7/N6 wafer area brings a greater return on investment when used for higher end products combined with using 12LP for this low end part, then the alternative. But, unless they are getting these parts for less than half of what they paid for Zeppelin die, it just isn't a break-even proposition. If it's less than 3 RDNA CUs, it probably isn't even worth it as a product for anything but the lowest end solutions, and it's a shame if it's less than 5-6. However, with even 6 RDNA 1 CUs, dual channel DDR4 should provide more than enough bandwidth, especially since RDNA1 is more memory efficient.
 

Panino Manino

Senior member
Jan 28, 2017
813
1,010
136
Aren't the gpu feature sets the decider for rdna1/2 branding. So the igpu could still have various improvements from rdna2 design but not meet all the criteria.

I thought about this, but the only sure thing I can think is that it'll not have RT and RT is the "basic killer" feature that makes RDNA2 deserve it's name.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
QUOTE="uzzi38, post: 40561958, member: 448427"]
Pretty sure RDNA2 CUs are marginally smaller than RDNA1 CUs actually.
[/QUOTE]

At roughly 75+million transistors more per CU, the transistor counts for the RDNA2 GPUs seem to indicate otherwise. Excluding the infinity cache still leaves them millions of transistors larger. You don't get the ray tracing units and expanded IPC for free.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
QUOTE="uzzi38, post: 40561958, member: 448427"]
Pretty sure RDNA2 CUs are marginally smaller than RDNA1 CUs actually.

At roughly 75+million transistors more per CU, the transistor counts for the RDNA2 GPUs seem to indicate otherwise. Excluding the infinity cache still leaves them millions of transistors larger. You don't get the ray tracing units and expanded IPC for free.[/QUOTE]

I mean I was just going off the Series X die shot provided by Microsoft a while back, which is still to this day the only real die shot we have, and those were just ever so slightly smaller than RDNA1 CUs iirc.
 

Asterox

Golden Member
May 15, 2012
1,026
1,775
136
Hm, what is going on here?Where is the root or reason, for this Cezanne Zen3/CPU performance regression in this 3DPM test. :mask:


R5 4650G, default setings/PBO is enabled, singlecore boost is 4.3ghz, allcore boost is 4.1-4.3ghz depending on the CPU load

R5 5600G, default setings/PBO is enabled, singlecore boost it should be around 4.4-4.5ghz, allcore boost it should be around 4.2-4.4ghz depending on the CPU load


2021-08-07_152228.jpg

R7 2700X vs R7 1800X.We now that R7 2700X has a higher IPC+higher Singlecore Turbo.

2021-08-07_141251.jpg
 
Last edited:

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
Look at the Geekbench benchmark subscores. There are two particular ones that show notable regressions, and both have to do with a particular set of math functions. In all likelihood, Zen3 is either generating too much heat in those functions and having to roll back clock speed or, there are different timings for them to accommodate a more frequent use case that hurts them.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
phttps://www.anandtech.com/show/16824/amd-ryzen-7-5700g-and-ryzen-5-5600g-apu-review/12

Isn't Ian IGP tests nonsensical?
Very low resolutions, even 360p all lowest quality. Who plays like that? Won't "real people" not use a mix of settings to get maximum quality possible at playable framerates? Aren't these new IGPs enough to 720p-900p medium?
Them he tests 1080p and even 4K at high and ultra. Why? We already know that it's possible with these IGPs, what the public wants to know is what sort of experience they can actually get with these APUs, but these tests don't answer this question, they mean nothing.


Ouch this is indeed nonsensical, nobody plays with 360p or 480p - 720p is the bare minimum. 720p and/or 1080p is the way to go and preferably 1080p. Actually there are more flaws in this test, many or most of the other iGPU scores seems to be a a copy and paste from old tests. For example the i7-1185G7 Deus Ex and Final Fantasy XIV fps were used from their Tigerlake September 2020 review. The driver in this test therefore is 1 year old which is crazy, that's why Anandtech never speaks open about the driver versions used in the test, this info is always hidden.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
Well here they are, they are way too close to each other.

RZasMUy.png

0wLU9Fw.png

RReGxxC.png

Y9fEroI.png

HcmzVI1.png

dJB0dfI.png

PiZjUwJ.png

EZtJcV1.png

EZtJcV1.png

PKGFgOy.png

EZtJcV1.png

RSKi0Xo.png

ybD6ydT.png

5mCUumz.png
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
Well, 1CU and 100mhz is not really something to expect any big gains from anyway, but its certanly faster.
It is more or less in line with my initial expectations, you are hightly memory limited at DDR4-3200 to DDR4-3600(2R), my previous experiences with 2200G/2400G/3200G and 3400G taught me that DDR4-3200/DDR4-3466 IS an issue for Vega 8 and 11 once you hit north of 1500mhz IGP clock in a big way. The 5600G at stock starts with more or less 3400G stock IGP performance, so that was not a big suprise to me there.

At 2R DDR4-3866/DDR4-4000 you stop(mostly) being memory limited and iGPU becomes the larger problem again. But you gain something out of faster IGP clocks at every case, even DDR4-3200.

What i did not expect to see is:
-Such small gains, i certanly expected more out of combining IGP OC and higher memory bandwidth vs 2R DDR4-3200.
-The 2R vs 4R results, the bad about 4R is that you are forced into 32GB unless someone has a very old 2R 16GB kit laying around(Or 4x4GB? i would not recomend that). This makes me want to try for a 8R 64GB setup.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Isn't Ian IGP tests nonsensical?
Very low resolutions, even 360p all lowest quality. Who plays like that? Won't "real people" not use a mix of settings to get maximum quality possible at playable framerates? Aren't these new IGPs enough to 720p-900p medium?

Ouch this is indeed nonsensical, nobody plays with 360p or 480p - 720p is the bare minimum.

No one actually plays at these frame rates. I suspect the inclusion of them is because the tests are done that way with desktop CPUs as well and this is merely just for the sake of having a full set of data across all of the CPUs reviewed.

While it may seem pointless to do this, there is an actual reason. If you were to look at 4K results for any particular CPU you'd most likely notice that every single CPU in the list is tightly bunched because the game is heavily GPU-bound and practically any CPU, sometimes even those like a 2-core Celeron or Athlon can achieve the same frame rate as a the top-end offerings from either AMD or Intel. Even moving down to 1440p or 1080p isn't always enough to get to a point where the GPU isn't the limiting factor.

By setting the resolution as low as possible, even though no one will actually play at that resolution, it shifts the bottleneck off of the GPU and actually makes it possible to determine if, or to what extent, there's any difference between CPUs. This really doesn't matter much in the here and now because no one plays at those resolutions and even if they did there aren't any monitors that are capable of outputting that many frames per second. However, some believe that this offers a good prediction of future performance as games do become more CPU intensive or as future GPUs become less bottlenecked by current requirements.

There's some contention over whether or not this matters, but testing of older CPUs with modern GPUs on recent games has shown that older CPUs (or even newer CPUs with reduced core counts) can actually result in a bottleneck. Assuming that sites that do that kind of testing continue to do so into the future we can probably revisit this in a few years to see if it actually matters, or to what extent low resolution benchmarks are a predictor of future performance.
 
  • Like
Reactions: Tlh97 and moinmoin

Panino Manino

Senior member
Jan 28, 2017
813
1,010
136
It appears that, functionally, the extra CU for the 5700 makes very little difference. It's all very DRAM bandwidth limited in a big way.

It's exactly the same as with the previous generations. The cheaper APU is the best one, same the money on better memory and overclock.
But I think this will change with RDNA.

No one actually plays at these frame rates. I suspect the inclusion of them is because the tests are done that way with desktop CPUs as well and this is merely just for the sake of having a full set of data across all of the CPUs reviewed.

While it may seem pointless to do this, there is an actual reason.

It's useful by being useless, I understand.
But still useless except for academic curiosity.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
Well, 1CU and 100mhz is not really something to expect any big gains from anyway, but its certanly faster.
It is more or less in line with my initial expectations, you are hightly memory limited at DDR4-3200 to DDR4-3600(2R), my previous experiences with 2200G/2400G/3200G and 3400G taught me that DDR4-3200/DDR4-3466 IS an issue for Vega 8 and 11 once you hit north of 1500mhz IGP clock in a big way. The 5600G at stock starts with more or less 3400G stock IGP performance, so that was not a big suprise to me there.

At 2R DDR4-3866/DDR4-4000 you stop(mostly) being memory limited and iGPU becomes the larger problem again. But you gain something out of faster IGP clocks at every case, even DDR4-3200.

What i did not expect to see is:
-Such small gains, i certanly expected more out of combining IGP OC and higher memory bandwidth vs 2R DDR4-3200.
-The 2R vs 4R results, the bad about 4R is that you are forced into 32GB unless someone has a very old 2R 16GB kit laying around(Or 4x4GB? i would not recomend that). This makes me want to try for a 8R 64GB setup.
Measurably faster and noticeably faster are two different things unfortunately. 1CU +100 mhz, or more with OC is 12.5% more compute resources plus 5-15% more frequency, yet, for configurations that have similar memory configurations, the net performance gain is usually no more than 2-3 fps, or, usually under 10%. Whereas something as simple as increasing rank interleaving show outsized gains to performance over raw bandwidth increases.

VEGA is, in a word, not that great for memory bandwidth efficiency (as compared to more modern or competing architectures). It does well with gargantuan memory bandwidth (maybe it was designed for HBM?) But seems to suffer with regular dram. It IS an improvement over Polaris in many ways, but, not much in the memory department.