Speculation: Ryzen 4000 series/Zen 3

Page 83 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

cytg111

Lifer
Mar 17, 2008
23,049
12,717
136
You guys not afraid that AMD is gonna bulldoze it with “a completely new arch” ??
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
You guys not afraid that AMD is gonna bulldoze it with “a completely new arch” ??
The "Zen Team" is hiring => https://www.linkedin.com/jobs/view/physical-design-engineer-76895-at-amd-1569427291

So, it seems unlikely that they'll depart away from Zen. Specifically, for a radical new architecture or clustered multithreading returning in top-end high-performance.

Zen -> Zen2 -> Zen3 -> Zen4 -> Zen5, are all set names. Nothing is going to push Zen naming out.
 
Last edited:
  • Like
Reactions: cytg111 and amd6502

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
You guys not afraid that AMD is gonna bulldoze it with “a completely new arch” ??
So, it seems unlikely that they'll depart away from Zen. Specifically, for a radical new architecture or clustered multithreading returning in top-end high-performance.

Zen -> Zen2 -> Zen3 -> Zen4 -> Zen5, are all set names. Nothing is going to push Zen naming out.
Not sure what you want to imply there. The worst they can realistically do is stagnate, of which there is no indication.

I think he meant if AMD is going to do Bulldozer Arch 2.0 with features like CMT. :D
 
  • Like
  • Love
Reactions: cytg111 and amd6502

amd6502

Senior member
Apr 21, 2017
971
360
136
I think he meant if AMD is going to do Bulldozer Arch 2.0 with features like CMT. :D

xen.jpg

I am hoping on BD CMT tricks as well. (Nosta probably too.) If they will also mirror about the horizontal axis (long axis indicated in white) so then they can cluster (share) the two FPUs of the adjacent cores together. But in Zen2 (above) they only mirror about the vertical (short) axis.

Maybe in Zen2 they did not mirror long axis to avoid hot spots, as you would get if you were to say run a pure FPU bench like cinebench multithread.

Heat could be an even bigger issue with denser 7nm+ (unless strategically sacrificing area).
 
Last edited:

amd6502

Senior member
Apr 21, 2017
971
360
136
NOooooooo....

Moore did nothing wrong.

If he had a working crystal ball he should've gone wider; maybe 3+2 on the integer side with room to grow to 4+2. Even so with the 2+2 integer core limitations, the frequencies might've made up for it if GF had gotten FDX to work many years earlier.

Clustering added to SMT could have nice benefits.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
Is there any chance of GF 12LP+ happening with Zen3 for IOD?
Any density improvements from 12LP+ on top of the gain from 7nm+ could come in handy if AMD want to use that silicon real estate for going wider or adding more cache.
 

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
I'm hoping that there's a premium SKU for 8C/16T, ideally monolithic in full. I want lower latency, max clock/IPC, without the better dies being binned entirely up to behemoth multi chiplet SKUs.

From experiencing my TR2920x vs 3700X vs 2700X, along with a stupid iMac Pro Xeon, I've come to the conclusion that I really don't need more cores. 8 is even overkill a lot of the time, though 6 feels bare minimum now. But I'd choose 5Ghz all-core 8C almost every time vs 4.5Ghz of even 64 cores or more.

I'm not saying that I wouldn't mind seeing 24C/32C Zen3 AM4+ SKUs, should that come to pass. Just want better 8C (or higher monolithic options). Something good could replace my 3700X and 9900KS into a one box solution. 2920X is being gifted off, it is a bad fit for my uses.
 

Thibsie

Senior member
Apr 25, 2017
727
752
136
Well, in theory monolithic might actually prevent you from getting better performance IMO.
I dot see the point at all unless you chose an APU ans discard the GPU (since they are monolithic at least for now) .
 

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
Well, in theory monolithic might actually prevent you from getting better performance IMO.

I dot see the point at all unless you chose an APU ans discard the GPU (since they are monolithic at least for now) .


If the APU has the bin and performance for it, I'd be all for it.


The multi-chiplet design has the higher latency cache, which along with the clock speed ceiling is the biggest things I'd like to see improved again with the 4000s. But no matter how good IF+Chiplet gets, it's necessarily slower than a theoretical identical design with everything on a unifed die.

 
  • Like
Reactions: Richie Rich

Hulk

Diamond Member
Oct 9, 1999
4,191
1,975
136
I'm hoping that there's a premium SKU for 8C/16T, ideally monolithic in full. I want lower latency, max clock/IPC, without the better dies being binned entirely up to behemoth multi chiplet SKUs.

From experiencing my TR2920x vs 3700X vs 2700X, along with a stupid iMac Pro Xeon, I've come to the conclusion that I really don't need more cores. 8 is even overkill a lot of the time, though 6 feels bare minimum now. But I'd choose 5Ghz all-core 8C almost every time vs 4.5Ghz of even 64 cores or more.

I'm not saying that I wouldn't mind seeing 24C/32C Zen3 AM4+ SKUs, should that come to pass. Just want better 8C (or higher monolithic options). Something good could replace my 3700X and 9900KS into a one box solution. 2920X is being gifted off, it is a bad fit for my uses.

While this is of course current reality, I am hoping that developers realize that clockspeed and IPC are topping out and better coding for multicore is the path to make your application run faster than your competitor's. 16 (or more) cores humming along at a nice efficient 4GHz in my compute heavy apps would be so nice.
 
  • Like
Reactions: guachi

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,685
136
But no matter how good IF+Chiplet gets, it's necessarily slower than a theoretical identical design with everything on a unified die.
You can't have your cake and eat it too.

Chiplets allow for more die area to be used for the same dollar spent. In a normalized cost scenario that means smaller cores on the unified die (or smaller cache, or both). At the end of the day you might as well end up paying more for mostly the same performance.

Look at Renoir and the sacrifices it had to make to keep costs in check.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Is there any chance of GF 12LP+ happening with Zen3 for IOD?
Any density improvements from 12LP+ on top of the gain from 7nm+ could come in handy if AMD want to use that silicon real estate for going wider or adding more cache.
12LP+ doesn't have a density improvement. However, if GloFo is a requirement then 12FDX does.

125 mm2 / 14LPP/12LP/12LP+ => ~100 mm2 / 12FDX

12FDX for I/O => Poly-biasing(any wider channel width supported), Continuous length(any channel length min to max), Body-biasing(active isolation), better Analog support(analog circuits produce less heat and operate faster).
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
Anybody want a Renoir review in an actual device?


Looks to be a 15W device, LPDDR4-3200, battery life looks splendid compared to Picasso (12 hours in PCMark10!) with a 4471mAh battery. Lots to talk about here, and it's a rather impressive first showing if you ask me.
 

exquisitechar

Senior member
Apr 18, 2017
655
862
136
I'm hoping that there's a premium SKU for 8C/16T, ideally monolithic in full. I want lower latency, max clock/IPC, without the better dies being binned entirely up to behemoth multi chiplet SKUs.

From experiencing my TR2920x vs 3700X vs 2700X, along with a stupid iMac Pro Xeon, I've come to the conclusion that I really don't need more cores. 8 is even overkill a lot of the time, though 6 feels bare minimum now. But I'd choose 5Ghz all-core 8C almost every time vs 4.5Ghz of even 64 cores or more.

I'm not saying that I wouldn't mind seeing 24C/32C Zen3 AM4+ SKUs, should that come to pass. Just want better 8C (or higher monolithic options). Something good could replace my 3700X and 9900KS into a one box solution. 2920X is being gifted off, it is a bad fit for my uses.
Not worth it for AMD to make something like that. Still, 4700x/4800x(?) will make every current 8 core CPU obsolete performance wise.
Anybody want a Renoir review in an actual device?


Looks to be a 15W device, LPDDR4-3200, battery life looks splendid compared to Picasso (12 hours in PCMark10!) with a 4471mAh battery. Lots to talk about here, and it's a rather impressive first showing if you ask me.
That looks good. Both performance and battery life, that is. Renoir just slaughters everything at 15W.
 
  • Like
Reactions: uzzi38

amrnuke

Golden Member
Apr 24, 2019
1,181
1,772
136
Zen2 jump was +15% so:
1.15 x 1.22 = 1.40 (+40%)
1.15 x 1.176 = 1.35 (+35%)... I guess you calculate also with Zen1+ jump.
From Zen to Zen+ was +3.5%, Zen+ to Zen2 was +15%.

Zen = 1.0 x. If we then add 3.5% for the Zen->Zen+ bump, it becomes 1.035 x Zen. If we then add onto that the 15% Zen+->Zen2 bump, that is 1.15 x 1.035 = 1.19025 x Zen. To get to 40% improvement from Zen to Zen3, one need only increase about 17.6% between Zen2 and Zen3 (1.19025 x 1.176 = 1.399 x Zen).

It's off topic here, but you are welcome to discuss A13 performance in its thread here. At page 17 and 18 there are SPEC bench and also Geek Bench 5.1 comparison. For example A13 in Povray benchmark is +25% faster which means 2.2x higher IPC (+121%) so Blender might be very similar to it. It's kind of unfair because A13 core has double transistors, 4x more L1$ and 16x more L2$ than Zen2 core. Different core for different market.
Not quite off topic, because AMD's Zen3 performance targets are directly pertinent to Apple's potential Intel replacement, in that they will be competing in the laptop marketplace. My question to you is - on that Povray benchmark where A13 is +25% faster, I'd like to have a source for that and any other benchmarks where A13 beats Zen2, for my own review (of course I've already seen coverage of the SPEC benchmarks - the Povray ones interest me).
 

RetroZombie

Senior member
Nov 5, 2019
464
386
96
But no matter how good IF+Chiplet gets, it's necessarily slower than a theoretical identical design with everything on a unifed die.
All pretty until you get into 1000mm2 die sides and get into the realm of the unmanufacturable, next to impossible to cool down, internal connectivity nightmare, ...
And I bet performance increases with chiplets by better binning of the best yields.

The intel gaming stronghold is still guaranteed by clock speeds with no TDP wall and not by the fact of being monolithic, if that was the case where is the intel HEDT gaming monster?
 
Last edited:
  • Like
Reactions: spursindonesia

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146
it has been taken down.

Archive link - you'll have to translate this one.

 

yeshua

Member
Aug 7, 2019
166
134
86
I'm surprised with the battery life results, because more cores more idle power usage and more cores more load power usage.

AFAIK both Intel and AMD have long found a way to shut down individual unused cores almost completely.
 

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
I'm surprised with the battery life results, because more cores more idle power usage
This hasn't been true with any Zen chip actually. The Zen cores themselves all have been very good at power gating when not in use. It always has been the uncore that ate pretty much all the power at idle, and Renoir looks to be the Zen uncore most optimized for mobile usage yet.