Speculation: Ryzen 4000 series/Zen 3

Page 75 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
Granted 7nm and chiplets is such gigantic step i havnt seen similar the last 30 years

What do you mean by this?

To me the IO die is basically what the old school Northbridge used to be back before A64 and Infinity Fabric is a fancy FSB. It seems like the Zen2 chiplets are essentially what CPUs used to be before they were packaged with additional IO.

It is on old idea updated and is very useful for maximising the advantage of newer nodes without having as large a yield penalty but I don't see it as a gigantic step.
 
  • Like
Reactions: yeshua

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
What do you mean by this?

To me the IO die is basically what the old school Northbridge used to be back before A64 and Infinity Fabric is a fancy FSB. It seems like the Zen2 chiplets are essentially what CPUs used to be before they were packaged with additional IO.

It is on old idea updated and is very useful for maximising the advantage of newer nodes without having as large a yield penalty but I don't see it as a gigantic step.
We weren't talking about the moon landing, which was a gigantic step for mankind, but about deciding to go with chiplets, which was a really gigantic step for AMD in a lot of ways.
 
  • Like
Reactions: lightmanek

kapulek

Member
Oct 16, 2010
56
33
91
Wow, that stepping number sounds like they taped out the Zen3 APU before they even announced Renoir - excellent roadmap execution if so.

Now all we need is details on the iGPU, RDNA1 or 2?

Given how early the news about Vega in Renoir turned up, I doubt it will be very long before we know the score.
We know that Van Gogh has RDNA2 based iGPU. It could be that AMD will skip RDNA1 and go for better (more power efficient?) RDNA2 iGPUs going forward (Van Gogh, Cezanne, Rembrandt).

 
  • Like
Reactions: soresu

stockolicious

Member
Jun 5, 2017
80
59
61
If Apple ordered from AMD a custom Renoir APU AKA Van Gogh with more powerfull integrated GPU then this could be a reason for Renoir having just 8 CU. I wouldn't surprised using HBM as more power efficient configuration.

Interesting is why AMD didn't released info about Zen 3. Could it be due to Zen 3 is delayed to Q4 20?
They are still pushing Zen2 stuff and don't want buyers to old off. Zen 3 is supposed to be a lot better
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
We know that Van Gogh has RDNA2 based iGPU. It could be that AMD will skip RDNA1 and go for better (more power efficient?) RDNA2 iGPUs going forward (Van Gogh, Cezanne, Rembrandt).

I had heard Rembrandt was going to be 5nm, surely that would be RDNA3 then?

Though Vega having 3 separate gens (16/14nm,12nm,7nm) was a surprise, I guess we can't take it for granted that there will be a new RDNA gen for every new node.
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
They are still pushing Zen2 stuff and don't want buyers to old off. Zen 3 is supposed to be a lot better
Quite, they only just officially rounded out the TR3 SKU announcements with 3980WX and 3990WX, and the Zen2 APU Renoir.

I'd expect at least 2-3 months before first official Zen3 info drops (not counting the unintended and removed unified L3 cache info), then release 6-8 months later maybe (September-December release timeframe).
 
  • Like
Reactions: stockolicious

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
What is the purpose of dumping here this imaginary nonsense of 'your guy', when AMD already said it publicly, officially and multiple times in the past month, that Zen 3 is a completely new uArch? Sure, they just suddenly dropped it last week. Totally realistic.
Seriously though, I'd be shocked if Zen3 was really a 'totally new' uArch. Surely, it's an evolution of the Zen2 design - even if it is more substantially changed.
Zen, was a new uArch compared to the construction cores. There is some semantic game afoot.
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
Seriously though, I'd be shocked if Zen3 was really a 'totally new' uArch. Surely, it's an evolution of the Zen2 design - even if it is more substantially changed.
Zen, was a new uArch compared to the construction cores. There is some semantic game afoot.
That has been my opinion on it - much like K10 was 'new', but definitely derivative of K8 and its predecessors.

If it were really a ground up redesign as Zen was over Excavator, I would at least expect a new name rather than Zen(n), which implies an iteration on the existing Zen designs.
 
  • Like
Reactions: Ajay

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
I missed half the conversation, but I will point out they mentioned those 8 'radeon cores' boost up to 1,770MHz depending on the part. I am dying for a desktop variant for my HTPC.

I am still wondering if my (offhanded) prediction that consoles will use Renoir + dGPU will come true. Turns out it appears to be true on PC.
 
  • Like
Reactions: lightmanek

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
I missed half the conversation, but I will point out they mentioned those 8 'radeon cores' boost up to 1,770MHz depending on the part. I am dying for a desktop variant for my HTPC.

I am still wondering if my (offhanded) prediction that consoles will use Renoir + dGPU will come true. Turns out it appears to be true on PC.

Yeah good luck with that as the 8CU will only be avalible for the 8C APU at around $350. Since the max CU number on Renoir is 8, they are going to segment the hell out of Vega.

Vega 8 went from being a $99 SKU to a $350 one.

And to be honest 1770mhz is not that much, i managed to OC my 3200G to 1700mhz, i had seen 3400G with Vega 11 at 1700mhz as well, is hard to stabilize it as VRMs becomes the problem. Thats the other problem for Renoir 7mm OC, some think they are going to be able to OC past 2GHz, but VRMs will be already on fire, AM4 motherboards arent designed for very high loads on chipset/gpu power.
 
Last edited:

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
I am still wondering if my (offhanded) prediction that consoles will use Renoir + dGPU will come true.
We already know that Xbox Series X (XSX?) gets a ~405mm² monolithic semi custom die.

dO7wSpcd_400x400.jpg


And PS5 is semi custom as well.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
I don't think that getting 8CUs on 7nm to 2Ghz is going to be out of reach for most. It should also be able to provide a decent performance uplift over Vega11 on the 3400g, so long as it gets decent RAM. Vega 11 was quite memory bandwidth starved, and clocking it higher was bringing very little extra performance in return. If Renoir can consistently get DRAM speeds over 4000, those 8CUs should be well utilized and perform quite a bit better. Again, we're grasping at straws here, and it's unlikely to hit 1050 levels, and will probably not be much past DDR5 1030 levels in most cases.
 

blckgrffn

Diamond Member
May 1, 2003
9,110
3,028
136
www.teamjuchems.com
We already know that Xbox Series X (XSX?) gets a ~405mm² monolithic semi custom die.

dO7wSpcd_400x400.jpg


And PS5 is semi custom as well.

So if Navi is ~250mm and the CPU portion of Renoir is ??? and the non CPU, non GPU portion is ??? how much more GPU is the Series X really going to have? Assuming it's an 8 CPU part and all that.

I guess I am wondering if these are really going to be the equivalent of 3700 CPU and a 5700 tied to a pool of fast memory. What is the power budget on this thing? (reading online, sounds like 300W)

I guess the GPU would be ~2x of the current X and the CPU massively faster.... still seems like 4k gaming is going to be a stretch.
 

Shivansps

Diamond Member
Sep 11, 2013
3,835
1,514
136
I don't think that getting 8CUs on 7nm to 2Ghz is going to be out of reach for most. It should also be able to provide a decent performance uplift over Vega11 on the 3400g, so long as it gets decent RAM. Vega 11 was quite memory bandwidth starved, and clocking it higher was bringing very little extra performance in return. If Renoir can consistently get DRAM speeds over 4000, those 8CUs should be well utilized and perform quite a bit better. Again, we're grasping at straws here, and it's unlikely to hit 1050 levels, and will probably not be much past DDR5 1030 levels in most cases.

We have a 8CU gpu that is not Vega, but it is Polaris, called RX550 that comes with a 128bit GDDDR5 memory.
TR;DR is barely better than a 3400G Vega 11 with 3200 rams stock vs stock (meaning the Vega 11 is 217mhz faster).

Desktop Vega needed more bandwidth, but there is no point now, who cares how well Vega 8 runs on a $350 APU? And anything below that will be more than enoght with just 3000-3200 rams.
 
Last edited:

HurleyBird

Platinum Member
Apr 22, 2003
2,670
1,250
136
That has been my opinion on it - much like K10 was 'new', but definitely derivative of K8 and its predecessors.

If it were really a ground up redesign as Zen was over Excavator, I would at least expect a new name rather than Zen(n), which implies an iteration on the existing Zen designs.

The "new" architecture thing is certainly ambiguous and open to interpretation. And does the interconnect and CCX stuff even count, or just the actual core? I'd say the later, but it probably depends who you ask. For the record, I'm more or less expecting Zen 3 to be to Zen 2 what Sandy Bridge was to Haswell.
 

RetroZombie

Senior member
Nov 5, 2019
464
386
96
On amd roadmap we know RDNA2 is on 7nm+.
Lisa Su already said we will see RDNA with zen 2 soon, so or it's just for custom guys, microsoft and/or apple or zen 3 is very ahead in it's design/testing phase.
 

RetroZombie

Senior member
Nov 5, 2019
464
386
96
What do you mean by this?

To me the IO die is basically what the old school Northbridge used to be back before A64 and Infinity Fabric is a fancy FSB. It seems like the Zen2 chiplets are essentially what CPUs used to be before they were packaged with additional IO.

It is on old idea updated and is very useful for maximising the advantage of newer nodes without having as large a yield penalty but I don't see it as a gigantic step.

Gigantic step because nobody else is doing that no more, they sure have done a regression, but without any penalty doing so which is the amazing part of all this.

Infinity Fabric fancy FSB? Who have done it of chip and on chip like them? With no limits, it's 8 chiplets on epic could be more.
 

DiogoDX

Senior member
Oct 11, 2012
746
277
136
So if Navi is ~250mm and the CPU portion of Renoir is ??? and the non CPU, non GPU portion is ??? how much more GPU is the Series X really going to have? Assuming it's an 8 CPU part and all that.

I guess I am wondering if these are really going to be the equivalent of 3700 CPU and a 5700 tied to a pool of fast memory. What is the power budget on this thing? (reading online, sounds like 300W)

I guess the GPU would be ~2x of the current X and the CPU massively faster.... still seems like 4k gaming is going to be a stretch.
I will especulate 360mm (60CU cut to 56 for yield) for the gpu and 50mm (cut L3) for the cpu.
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
The "new" architecture thing is certainly ambiguous and open to interpretation. And does the interconnect and CCX stuff even count, or just the actual core? I'd say the later, but it probably depends who you ask. For the record, I'm more or less expecting Zen 3 to be to Zen 2 what Sandy Bridge was to Haswell.
Probably both - it was in an interview with Mark Papermaster I think that they would revise the IF spec/version with every new generation, so it's doubtful that they could just use the CCD's interchangeably.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Seriously though, I'd be shocked if Zen3 was really a 'totally new' uArch. Surely, it's an evolution of the Zen2 design - even if it is more substantially changed.
Zen, was a new uArch compared to the construction cores. There is some semantic game afoot.
He said it wasa refresh. Core 9nd gen is a refresh.
 

Veradun

Senior member
Jul 29, 2016
564
780
136
I will especulate 360mm (60CU cut to 56 for yield) for the gpu and 50mm (cut L3) for the cpu.
I second this and further add that I think the CPU part will be 8c/8t (no SMT) to have better efficiency and also a strict parallelism with jaguar for perfect portability of whatever code one wants to port from the one x
 

maddie

Diamond Member
Jul 18, 2010
4,722
4,627
136
I second this and further add that I think the CPU part will be 8c/8t (no SMT) to have better efficiency and also a strict parallelism with jaguar for perfect portability of whatever code one wants to port from the one x
Is this still a belief?
 

thor23

Member
Jul 13, 2019
80
22
81
No smt has worse efficiency on my 2600, it only has reduced max tdp. SMT allows cores to enter c states faster.