• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Discussion AMD Cezanne/Zen 3 APU Speculation and Discussion

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Hitman928

Diamond Member
Apr 15, 2012
3,400
3,554
136
While the Zephrus was impressively thin, the Flow X13 is even more impressively thin.
I'm not disagreeing that thermals are playing an issue in many of the results, but in my post (I should have been more clear) I was referring to the 5980HS in 15W configuration being unusually low (compared to 35W configuration) in certain results. I'm sure it's a platform/configuration issue though as it is only in a limited number of results and is unrealistic when looking at the 15W vs 35W Renoir results.
 

Mopetar

Diamond Member
Jan 31, 2011
5,400
1,982
136
A bit of a disappointment, and looks like Vega really reached it's limit.
I'm surprised they managed to get as much out of Vega as they did. It's pretty clear that whatever they did to the architecture has put it leaps and bounds above where it was when it started and some of that makes me wonder if or how much Vega is just a name that they kept.

Ultimately though any mobile APU is going to hit a wall due to memory bandwidth limits since a discrete GPU typically has an order of magnitude more bandwidth available, which is why we've seen the CU counts decrease over the generations as clock speeds have risen.

Maybe when they move over to RDNA2-based designs, the massive uplift that infinity cache provides will alleviate a lot of that, but they might be working on a tighter integration with the CPU cores for that cache as well since it's obvious that Zen 3 has itself benefited from having more cache available.
 

moinmoin

Platinum Member
Jun 1, 2017
2,286
2,742
106
I said some time ago that MT score would probably be disappointing, because when you compare the voltage curve between Matisse and Vermeer, the voltage difference at the same frequency decreases the lower you are in the curve.



That's some great info. By The Stilt perchance? I did a quick sloppy overlay to show more clearly how much the improvement is between the two.

 

Panino Manino

Senior member
Jan 28, 2017
350
400
136
A key part of Zen is the chiplet strategy, allowing AMD to fab cheaper.
The change from Vega to RDNA1/2 means a significant an increase in the monolithic die even if it comes without IF, making it more expensive, lowering the yelds, at a time where fabs (TSMC) are overloaded. Looks like AMD wants to delay this change as much and possible, instead focusing on selling notebooks with stronger dedicated GPUs.

I wouldn't be surprised if when a mobile APU with RDNA comes it'll be very small with little gains over Vega, mostly for the improved media block.
 
  • Like
Reactions: Tlh97

gdansk

Senior member
Feb 8, 2011
590
252
136
I wouldn't be surprised if when a mobile APU with RDNA comes it'll be very small with little gains over Vega, mostly for the improved media block.
I would be very surprised. There remains strong demand for better mobile GPUs without extra cooling solutions. Intel is improving rapidly again. And Samsung is using RDNA2 so I expect some very good downscaling to have enticed them away from Mali. Regardless, even a measly 2TF 8 CU design would have better game performance than its equivalent Vega++. Even without any inf cache it has better efficiency from wave32+simd32 and making better use of available bandwidth with more compression.
 
  • Like
Reactions: french toast
Mar 11, 2004
20,846
3,009
126
If I had to guess, the reason they've stuck with Vega is more about compatibility than performance (which is good enough and they're getting small gains here and there) at this point. Making it easier for themselves and OEMs to keep compatibility across laptops (to make transitioning to newer chips, like from Zen 2 to 3, easier/faster). From what I gather there's been quite a lot of change in the software backend with RDNA2 and so they might be holding out til that is further streamlined (think one of the AT writers even talked about how GPU drivers are huge and complex undertakings these days).

It might also hint that a lot of the benefits post Vega were actually in software and not hardware. They were doing a significant overhaul of that aspect with Vega and that was definitely an issue when they delayed then deprecated some of the software benefits that were supposed to happen. Although I'd guess RDNAx improvements have been a mix, of them eschewing much of the excessive compute of Vega enabling higher CU counts and other improvements within CUs while improving the software graphics throughput, bringing the big gains we've seen. And the iGPU Vega likely was already somewhat streamlined with regards to compute compared to the large dGPU versions. That Intel chip (and the previous consoles) might also give some hints about how Vega is a bit of a meaningless nomenclature.

A key part of Zen is the chiplet strategy, allowing AMD to fab cheaper.
The change from Vega to RDNA1/2 means a significant an increase in the monolithic die even if it comes without IF, making it more expensive, lowering the yelds, at a time where fabs (TSMC) are overloaded. Looks like AMD wants to delay this change as much and possible, instead focusing on selling notebooks with stronger dedicated GPUs.

I wouldn't be surprised if when a mobile APU with RDNA comes it'll be very small with little gains over Vega, mostly for the improved media block.
While that is fairly true, I think that can be mitigated via a shared cache in an APU vs the two larger caches of CPU and GPU chiplets.

I think they're more waiting for other improvements like memory bandwidth (could see Zen 3+ or whatever the expected refresh is called will support DDR5) before they upsize the GPU.

I feel like that new ASUS laptop that Anandtech used for the 5000 series laptop testing might be a bit of a harbinger of what many PC laptops will look like, but where the internal dGPU is instead a chiplet on package with the CPU. I'm not expecting a console like GPU heavy APU (although I'd love one), but I think we'll see APUs get stronger GPUs with the option for gamers to add an external one (with their being a mix of options, smaller transportable ones like the ASUS is going for, but also the larger stationary boxes we'd been seeing prior).
 

Panino Manino

Senior member
Jan 28, 2017
350
400
136
Actually... when I pay attention to the 15W results, it's much more impressive, right?
There's little difference in performance and I imagine the difference in battery time must be huge.
 

Bam360

Member
Jan 10, 2019
30
58
61
Remember not only MT score leaks from Geekbench 5 where Cezanne showed lower than expected MT gains, but also AdoredTV leaks about Milan, where he said ST gains much higher than MT gains in the same power envelope, it all makes sense, Zen3 consumes more power when frequency and voltage is the same as Zen2. So when you go lower in the F/V curve, the lower the difference will be vs Zen2 at same power envelope.
Desktop Zen3 is more impresive in MT workloads because the power limit is much higher and it sits very high in the F/V curve, where there is more difference in the voltage required vs Zen2.
 

LightningZ71

Senior member
Mar 10, 2017
658
583
106
Those extra transistors don't get used for free. Zen3 is a larger core than Zen 2 from a transistor count perspective. Those extra transistors aren't there as window dressing, but to be used. Zen3, as it exists on Cezanne, is still using essentially the same process node as Renoir did, so, the only advantage that it would have with respect to power draw would be general layout efficiency improvements in the Zen3 core, which we don't know really exist, and improvements in the power management and voltage distribution design of the processor itself. However, when compared to Lucienne, according to AMD, it also has those same power management improvements, which means that that advantage won't even be there.

I'm actually quite interested in the comparison between Lucienne and Cezanne in the U series lineup. The most telling comparison will be between the 5700u (essentially an improved 4800u) and the 5800u. I have a feeling that the 5700u will be just as fast, or even faster, in a small set of MT tasks, especially ones that scale well and don't have a lot of communication between the various threads and aren't highly L3 cache bound. It'll likely also be just about as fast for games that are lightly threaded, especially when using the iGPU.
 

Hitman928

Diamond Member
Apr 15, 2012
3,400
3,554
136
Actually... when I pay attention to the 15W results, it's much more impressive, right?
There's little difference in performance and I imagine the difference in battery time must be huge.
Yes, I noticed the same thing. It would be interesting to see the same comparison in a laptop with a more robust cooling solution to see how much the 35W configuration is being limited by thermals.
 

Zepp

Member
May 18, 2019
54
56
51
I'm surprised they managed to get as much out of Vega as they did. It's pretty clear that whatever they did to the architecture has put it leaps and bounds above where it was when it started and some of that makes me wonder if or how much Vega is just a name that they kept.
Yes, I'm surprised they never came up with a refresh name for it or called it something like Vega+.
 

uzzi38

Golden Member
Oct 16, 2019
1,363
2,529
96
Remember not only MT score leaks from Geekbench 5 where Cezanne showed lower than expected MT gains, but also AdoredTV leaks about Milan, where he said ST gains much higher than MT gains in the same power envelope, it all makes sense, Zen3 consumes more power when frequency and voltage is the same as Zen2. So when you go lower in the F/V curve, the lower the difference will be vs Zen2 at same power envelope.
Desktop Zen3 is more impresive in MT workloads because the power limit is much higher and it sits very high in the F/V curve, where there is more difference in the voltage required vs Zen2.
Desktop Zen 3 is more impressive because the power limit is much higher huh.

Is now a good time to mention the Scar 15 and 17 both support running the 5900HX at 80W, and the Zephyrus Duo 15 supports it at 90W (CPU loads only, CPU+GPU load it drops to 54W).

The best thing about Cezanne isn't the chip itself. Renoir was already fantastic, Cezanne is a minor improvement. What makes Cezanne so good is the laptop designs it comes with.

Also, don't expect people to believe AdoredTV when they say dumb things like Milan has basically no IPC uplift in FP workloads.
 

Mopetar

Diamond Member
Jan 31, 2011
5,400
1,982
136
Yes, I'm surprised they never came up with a refresh name for it or called it something like Vega+.
I think that even though AMD used it as the name for an actual product that most people weren't familiar with it and the only people who associated it with disappointment are people on tech forums that really won't care what you call it if it's good.

To the average consumer it's an easy to remember piece of branding that sounds kind of cool and maybe reminds them of the masked claw fighter from Street Fighter who was also pretty damn cool as well. Even if you don't know a thing about technology you can still brag about your Vega graphics which will sound impressive to anyone else who doesn't know a thing about technology. It's like having blast processing but without all the extra syllables.
 
  • Haha
Reactions: rUmX

jpiniero

Diamond Member
Oct 1, 2010
8,851
1,617
126
Reviews of the 5980HS and 5900H :
You can see why OEMs wanted to wait for Cezanne before doing AMD gaming notebooks, even if Intel did juice the pot a little bit. Comet Lake is pretty close although these benchmarks are a little strange due to comparing a 3060 at various power levels versus 3070 and 3080.
 

zir_blazer

Senior member
Jun 6, 2013
988
186
106
Giving the law of diminishing returns, I'm actually extremely anxious of seeing how a 8C Cezanne compares to a 8C Vermeer. Sort of comparing Renoir vs Matisse to see how going back to a monolithic die with the IMC on-die compares to an off-die IMC with 4x the Cache L3. Just now that the difference in Cache L3 is reduced to just 2x.

Renoir: 2 CCX w/4C and 4 MiB Cache L3 each with on-die IMC
Matisse (Single Die): 2 CCX w/4C and 16 MiB Cache L3 each with off-die IMC

Cezanne: 1 CCX w/8C and 16 MiB Cache L3 with on-die IMC
Vermeer (Single Die): 1 CCX w/8C and 32 MiB Cache L3 with off-die IMC


Before Renoir benchmarks were released I was interesed in how close Renoir 4700G and Matisse 3700X were to being equivalent, depending on price. If I recall correctly Cache L3 made a sizeable impact on gaming. If wasn't because desktop Renoir is close to unavailable I would consider it a great alternative.
As Cezanne has effectively four times the Cache L3 for a Single Thread than Renoir, Cezanne should be much closer to Vermeer than Renoir was to Matisse, so I'm wondering about whenever it could finally cannibalize all 8C and lower Vermeers with no IGP. IGPs are usually overlooked but they are a massive bonus when you need them, and I would lean Cezanne assuming than similarly specced Cezannes and Vermeers were at the same price.
 

Asterox

Senior member
May 15, 2012
454
623
136
Giving the law of diminishing returns, I'm actually extremely anxious of seeing how a 8C Cezanne compares to a 8C Vermeer. Sort of comparing Renoir vs Matisse to see how going back to a monolithic die with the IMC on-die compares to an off-die IMC with 4x the Cache L3. Just now that the difference in Cache L3 is reduced to just 2x.

Renoir: 2 CCX w/4C and 4 MiB Cache L3 each with on-die IMC
Matisse (Single Die): 2 CCX w/4C and 16 MiB Cache L3 each with off-die IMC

Cezanne: 1 CCX w/8C and 16 MiB Cache L3 with on-die IMC
Vermeer (Single Die): 1 CCX w/8C and 32 MiB Cache L3 with off-die IMC


Before Renoir benchmarks were released I was interesed in how close Renoir 4700G and Matisse 3700X were to being equivalent, depending on price. If I recall correctly Cache L3 made a sizeable impact on gaming. If wasn't because desktop Renoir is close to unavailable I would consider it a great alternative.
As Cezanne has effectively four times the Cache L3 for a Single Thread than Renoir, Cezanne should be much closer to Vermeer than Renoir was to Matisse, so I'm wondering about whenever it could finally cannibalize all 8C and lower Vermeers with no IGP. IGPs are usually overlooked but they are a massive bonus when you need them, and I would lean Cezanne assuming than similarly specced Cezannes and Vermeers were at the same price.

In today GPU market+bitcoin mining madness, Renoir/Cezanne iGPU performance can not be overlooked.Renoir/Cezanne Price=CPU+iGPU performance, who cares about lower CPU gaming performanse with dedicated GPU.Most people don't care for PCI-express 4.0 either, or PCIe Express 3.0 is more the enough.Here is example, current prices in my country Renoir vs Ryzen 3000 Desktop.

Ryzen 5 3600

- local currency in dolars 270$ box version, tray version 260$



Ryzen 5 Pro 4650G

- local currency in dolars 240$

- iGPU performanse similar to GT 1030

- Renoir/Cezanne it is easier to cool and consumes less energy



Dedicated GT 1030 as new, lowest local retail price is 95$.

 

misuspita

Member
Jul 15, 2006
190
186
116
What would be more appropriate for a 6 core cezanne?

3200 cl14 or 4400 cl19? Would the mem controller go to 4400 speeds or it is useless?
 

LightningZ71

Senior member
Mar 10, 2017
658
583
106
If the IMC in Renoir is any indication, it should reach fairly high clock rates on dram. The question is, will the reduced CU count on the lower end SKUs be able to use higher memory throughput? Each rung down the ladder is 10-15% less gpu performance.
 

Mopetar

Diamond Member
Jan 31, 2011
5,400
1,982
136
I don't think it will be too much of an issue considering that the clock speed has been increased significantly. Compare this against the lower-end Picasso parts which had up to 8 CUs at 1.2 GHz. Without any IPC improvements, 6 CU at 1.6 GHz has the same theoretical performance. The bottom Renoir chips had only 5 CU at a lower clock speed, so this isn't likely to perform worse when you stack it up against past products.

Maybe it's a bit unusual to pair the lowest end APU with such fast RAM, but historically they've always been pretty starved for bandwidth, so perhaps it won't be quite as pointless as imagined.
 

ASK THE COMMUNITY