Speculation: Ryzen 4000 series/Zen 3

Page 108 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Senior member
Oct 16, 2019
894
1,104
96
Just a little something I squeezed out of AdoredTV. Had a strong feeling it was the case with how late B550 is, but good to get confirmation.

This should also probably throw out any rumours that Zen 3 will not support AM4 or bresk backwards compatibility.


This is likely one of the things that will be discussed on Thursday.

EDIT: All I know of is that there is an embargo on Thursday to discuss something that could have easily been covered in the 3300X/3100 launch. I can't say I know specifics apart from one other thing - desktop Renoir is rumoured to launch alongside B550. Take that how you will.
 
Last edited:

YAYgee

Junior Member
May 4, 2020
12
5
41
In April Forbes said that AMD "announced an embargo of today for part of the news in a briefing I listened to this morning,", i.e. the Ryzen 3/B550 press release, "with the rest under wraps for a couple of weeks", this Thursday that part of the briefing will be made public. Regarding B550, Linus and others are also under NDA. It's made out to be a big deal. We'll see if that's indeed the case and if not I agree with @uzzi38 that could have easily been covered in the 3300X/3100 launch". Sadly, it sounds like you already know it isn't a big deal.

I doubt we'll get bad news on Thursday. Still, throwing this quote in here as I can't tell if there's any truth to it:

AMD has screwed up a bit and there's a hardware bug affecting all currently made B550 boards. It's not something that prevents regular operation, but it's a silly screwup that would be annoying, if those boards ship as they are. I can't share what it's is here though, as it might reveal where I got the info from. - from TPU's forum
 

YAYgee

Junior Member
May 4, 2020
12
5
41
@uzzi38 "desktop Renoir is rumoured to launch alongside B550" => that would be nice with 3400G's supply chain issues (or what else has caused stock to dry up) and considering the new Ryzen 3 CPUs donot address the same (iGPU) need; leaving e.g. 2200G and 3200G to compete with Intel. Yet how confident are you about this rumour when the new consoles must hit the shelves in large numbers and taking Renoir laptop delays into account?

PS according to GN AMD's internal roadmap shows that APUs will have DDR5 in 2022. First time hearing about Zen 3+, maybe it's 7nm IP reused on N6? Is it possible for this (AM5) processor to work on AM4+ motherboards and Thursday's "big deal" reveal = B550 with AM4+ (or x670 with AM4+), cfr. AM3 processors with DDR2 and DDR3 memory controllers? Sorry about these wild guesses, blame it on corona.
 

DisEnchantment

Senior member
Mar 3, 2017
578
1,077
106
Milan is shipping by late next Quarter for Cray it looks like. It will allow Cray enough time to assemble and ship it within Q4 for eventual deployment by year end.
Phase 1 of Perlmutter to be deployed in late 2020; GPU readiness efforts well underway
Phase 1 will consist of 1500 Milan CPUs and 6000 Volta-Next
Each of Phase 1's GPU-accelerated nodes will have 4 of the next generation (Volta-Next) NVIDIA GPUs, along with 256GB of memory for a total of over 6000 GPUs. In addition, the Phase 1 nodes will each have a single AMD Milan CPU.
Each of Phase 2's CPU nodes will have 2 AMD Milan CPUs with 512 GB of memory per node. The system will contain over 3000 CPU-only nodes.
Video showing deployment of HPE Apollo Hawk in Stuttgart. The Perlmutter is a lot more sophisticated system.


You can kind of make a reasonable guess about Ryzen 4K from all of this.

Thanks to @kokhua for the news on Twitter.
 

DisEnchantment

Senior member
Mar 3, 2017
578
1,077
106
Just a little something I squeezed out of AdoredTV. Had a strong feeling it was the case with how late B550 is, but good to get confirmation.

This should also probably throw out any rumours that Zen 3 will not support AM4 or bresk backwards compatibility.


This is likely one of the things that will be discussed on Thursday.

EDIT: All I know of is that there is an embargo on Thursday to discuss something that could have easily been covered in the 3300X/3100 launch. I can't say I know specifics apart from one other thing - desktop Renoir is rumoured to launch alongside B550. Take that how you will.
Funny, the guy from ADTV said Zen3 for 2020 is optimistic when AMD themselves said Zen3 in 2020 and that too in FAD no less just like you rightly said. Then next day, Perlmutter is planned to be deployed by late 2020.
And then in another tweet he asked Chia if the Perlmutter specs confirmed some of their leaks which we knew is from HPC Advisory council last year Oct 4 2019. They took the slides from that video removed the AMD watermark and posted it on their website as their "leak"
And another thing that is a common theme, that is so cringeworthy, they keep making up stuff like: yeah our sources ... yeah silicon is buggy, A0 no SMT, yeah power targets not met for Navi2X, B550 issues in HW, frequencies not decided.... An Engineer's day to day job is a leak.
By these standards, I have also "LEAKED" Zen3 last year Oct4 right here on AT Forums :laughing:
Video might get taken down


View attachment 11605


View attachment 11606




Zen 3 Milan highlights [AMD, Martin Hilgeman ]
- Unified L3 32+ MB per CCD
- Sampling already
- 7nm
- Same core count as Rome
- 2x SMT
- Planned for Q3 2020
- DDR4/SP3

What is Zen3's special sauce gonna be?
- Bigger cache most likely (32MB+)
- Improved IF
- ...
 

tamz_msc

Platinum Member
Jan 5, 2017
2,577
2,255
106
Milan is shipping by late next Quarter for Cray it looks like. It will allow Cray enough time to assemble and ship it within Q4 for eventual deployment by year end.


Phase 1 will consist of 1500 Milan CPUs and 6000 Volta-Next


Video showing deployment of HPE Apollo Hawk in Stuttgart. The Perlmutter is a lot more sophisticated system.


You can kind of make a reasonable guess about Ryzen 4K from all of this.

Thanks to @kokhua for the news on Twitter.
Looks like no AVX-512 for Zen 3.
Phase 1 of Perlmutter to be deployed in late 2020; GPU readiness efforts well underway
Phase 2 is when the CPU nodes arrive though, and that's mid-2021.
 
Last edited:

DisEnchantment

Senior member
Mar 3, 2017
578
1,077
106
Looks like no AVX-512 for Zen 3.

Phase 2 is when the CPU nodes arrive though, and that's mid-2021.
There will be 1500 Milan CPUs and 6000 Volta-Next GPUs for Phase1 GPU nodes. For every 4x Volta-Next there is 1x EPYC.
Each of Phase 1's GPU-accelerated nodes will have 4 of the next generation (Volta-Next) NVIDIA GPUs, along with 256GB of memory for a total of over 6000 GPUs. In addition, the Phase 1 nodes will each have a single AMD Milan CPU.
CPU only is for Phase 2. 3000 nodes, 2x EPYC per node.
 

senttoschool

Golden Member
Jan 30, 2010
1,412
153
106
It's not just cloud infrastructure. The PC market in general benefits from the situation as more people staying at home means the need for a PC or laptop increases (god even just saying that feels wrong, but it is what it is).
Maybe temporarily in March. By now, it should be declining due to all the layoffs and companies not wanting to spend cash to upgrade computers right now.
 

senttoschool

Golden Member
Jan 30, 2010
1,412
153
106
It's a brief increase in lappie sales. No idea how long that's gonna last. But cloud infrastructure has been hit pretty hard by extra Internet traffic worldwide, even from those users not buying new computers.
Internet traffic is up only for video traffic. That means Netflix, YouTube will need more processing power and everyone else will need less.

 

Valantar

Golden Member
Aug 26, 2014
1,792
508
136
Exactly. That is what I mean. It would not surprise me at all if they have decided to neglect desktop APUs in this generation due to increased demand from OEMs. It is the same die. Demand for (and sales of) their desktop APUs have been tepid since they released Raven Ridge. The sales are probably good by Kaveri/Godavari standards, but today . . . ? No. And that is probably why we haven't seen a 4400G or . . . whatever they might call AM4 Renoir.
Desktop APUs have always arrived significantly later than mobile APUs. 3000-series mobile APUs were launched at CES and started arriving in laptops in April-May or so. Desktop APUs hit the market in August or later. Given the CES launch of 4000-series mobile APUs and their arrival on the market in March, we're likely looking at a similar timeframe for 4000-series desktop APUs. Nothing at all unusual about that, and nothing indicating "neglect[ing] desktop APUs in this generation".
@Topweasel

Also, OEMs are not afraid to use laptop hardware in a desktop form factor if the price and performance are right. They don't need socket AM4 unless it offers some product advantage, like better cooling; lower prices; or better SKUs for their customer. Lots of desktop AiO solutions have mobile hardware in them. Well some of them do.

If AMD is able to sell all their Renoir dice in mobile form factor at attractive margins, they aren't going to want to repackage as many of them as 4000-series desktop APUs.
It's true that OEMs aren't afraid of that, but it's more expensive (having to solder the APU to the board runs the risk of scrapping whole boards if the APU is defective) so they avoid it unless it is for some application where they find it makes more sense (such as SFF systems where they want the direct-die cooling). Using socket AM4 allows them to re-use motherboard designs, which saves money, and allows for further scaling of the same product line should they want to.

Tl;dr: there is zero indication that there won't be a Renoir AM4 desktop APU series.
@Shivansps

It is inaccurate to assume that continuing to use Vega would be a decision borne of "no demand for powerful iGPUs". AMD is going to continue to update Vega as a part of CDNA. The big question is whether or not updated CDNA will find it in any display-capable hardware. RDNA is allegeldy not compute-friendly, so I would not expect to find it in workstation cards and other professional dGPUs intended for display duty (which do exist, and have for some time). We already know Renoir has Vega and AMD has decided on Vega for mobile Renoir, so logically desktop Renoir - if/when it appears - will also be a Vega product. Whatever happens on the desktop will be an extension of whatever it is AMD decides is important for mobile - not for desktop. Desktop APUs are also-rans.
I find it rather absurd to think CDNA - a datacenter compute centric spin-off of GCN - will find its way into APUs of all things. Even for a workstation APU the iGPU won't be big enough to deliver meaningful compute to the extent that a workstation needing GPU compute can skip the dGPU. And RDNA delivers perfectly acceptable compute performance, even if it isn't quite at GCN (or CDNA) levels per CU. And AMD has stated quite unequivocally that GCN is dead and that RDNA is their "GPU architecture for the next decade" - note the singular form of "architecture", indicating that no others will be carried forward and that CDNA is a compute accelerator architecture rather than a GPU architecture. Given that the main focus for APUs is integrated graphics for mobile and low-power applications (including Ryzen embedded and EPYC embedded) there is zero reason to think CDNA will be tapped for such a job - that's not what it's designed for. And GCN was warmed over for Renoir simply because the RDNA design wasn't ready for porting into the APU design when they needed it in the Renoir design cycle - 1 - 1 1/2 year (if not more) before launch, i.e. mid 2018 or earlier (a full year before the first RDNA dGPUs). Next gen APUs will 99% sure use RDNA, if for no other reasons than Intel finally becoming competitive on iGPU performance (but likely also to simplify future driver development).
 
  • Like
Reactions: Tlh97 and Gideon

Gideon

Senior member
Nov 27, 2007
831
1,259
136
I find it rather absurd to think CDNA - a datacenter compute centric spin-off of GCN - will find its way into APUs of all things.
I agree, we might see some 2.5D server chips down the line having CDNA chiplets (in fact this looks almost certain) but CDNA in a mobile or desktop APU is IMO also quite absurd. RDNA2 has enough compute perf and more importantly has all the required things for 3D rendering: TMUs, ROPs, Geometry/tessellation units possibly raytracing units (in APUs they migh skip these) ... all of these things will be removed for CDNA as they make no sense in datacenters. It would be quite strange if AMD decided to bolt them back on to a mainstream APU where 99% of the usecases won't benefit from the CDNA arch.
 

Gideon

Senior member
Nov 27, 2007
831
1,259
136
This is a small disappointment. I don't think anybody expected 512bit wide FP units (that would be quite the die-space waster considering they just doubled them up in Zen 2) but supporting a subset of AVX-512 instructions with 256bit units would still have been quite useful for some workloads. CLANG/GCC have already made 256bit wide AVX-512 the default as Intel throttles on 512bit wide instructions. It's still often useful and would have allowed more widespread use earlier for the devs.

I can't find the tweet I saw a while ago, but even with just 256 bit wide AVX-512 some devs have gotten 30%+ perf gains due to some nice new instructions missing in AVX2.

Of course it's totally fine for AMD decide to skip the additional decoder complexity that AVX-512 presents, it's far from a critical feature. Why I find it a bit of a dissapointment is just that:
1. Even VIA has added some AVX-512 support in it's latest Centaur processors (with also just 256 bit wide vector units and a fraction of development resources)
2. It looks like even Intel Atom cores will have it with Gracemont

So it's certainly doable. With AMD it looks we have to wait till 2022 (Zen 4) to get the first processors with any support (which means it will take years before this can be the default-path for most x86 processors in executables)

Hopefully AMD used the resources well to either improve AVX2 workloads significantly or added some variable extensions themselves.
 

inf64

Platinum Member
Mar 11, 2011
2,977
1,462
136
We already heard some rumors of ~50% better FP(SIMD) performance for Zen3 vs Zen2. Also, I think Zen4 will fully support 2x512bit datapaths , just like Zen2 did with AVX2.
 

DrMrLordX

Lifer
Apr 27, 2000
16,016
4,969
136
Desktop APUs have always arrived significantly later than mobile APUs.
Kaveri/Godavari wasn't laptop-first. Everything after that? Yeah. But that should tell you the difference between the AMD that takes desktop APUs seriously and the AMD that doesn't.

Tl;dr: there is zero indication that there won't be a Renoir AM4 desktop APU series.
. . . eventually.

I find it rather absurd to think CDNA - a datacenter compute centric spin-off of GCN - will find its way into APUs of all things.
We will see. I don't think anyone knows for sure what the CDNA/RDNA split is going to be like until we see CDNA products. If RDNA represents a compute performance regression vs. Vega20 on a per-CU basis, then some folks are not going to be happy.
 

Gideon

Senior member
Nov 27, 2007
831
1,259
136
We will see. I don't think anyone knows for sure what the CDNA/RDNA split is going to be like until we see CDNA products. If RDNA represents a compute performance regression vs. Vega20 on a per-CU basis, then some folks are not going to be happy.
I don't see any reason to keep any graphics related silicon on CDNA (e.g. think Frontier and so forth) Nvidia also doesn't do that. My bet is that CDNA is with HBM and strictly without any graphics bits. For everything that needs graphics there are RDNA2 derivatives.

The fact that CDNA is "professional/datacenter" doesn't mean there won't be Radeon PRO or W serires cards with RDNA. In fact there already are. Even datacenter (cloud-GPU) stuff will certainly be RDNA.

What are the usecases where the RDNA compute perf will be a critical shortcoming in desktop/mobile workloads? Any real HSA stuff has so little traction RDNA2 will be good enough for it for years (it isn't even 0.1% of the usage IMO)
 

soresu

Golden Member
Dec 19, 2014
1,324
520
136
I don't think anybody expected 512bit wide FP units
Yeah, the scaling benefits of N7+ are just not enough to justify that leap - Zen 4 on the other hand with N5P should have that at least, Genoa should be a beast for encoding once SVT xxx are more optimised for AVX 512 than they are already, not to mention likely offline path/ray tracers too.
 
  • Like
Reactions: Gideon

DisEnchantment

Senior member
Mar 3, 2017
578
1,077
106
If RDNA represents a compute performance regression vs. Vega20 on a per-CU basis, then some folks are not going to be happy
RDNA1, and specifically Navi12/GFX1011 in particular has more feature sets than Vega20/GFX906 except half rate DPFP.
Outside of DPFP, Navi12 should outperform Vega20 in all other aspects by virtue of higher clocks/CU and for the new mixed precision operations Navi12 will definitely beat it because of multiple ops getting executed at once.
So I think it should be fine in that perspective.

GFX908/Arcturus is the only GFX9 series that has way more feature set than Navi for pure vector math. It has MFMA with bfloat16 types. Even NVIDIA don't have that yet.
Dunno if AMD will add MFMA to their RDNA lineup. Then, other things like network workload distribution, networked coherency, XGMI and the like may not be that interesting for the Desktop.
 

soresu

Golden Member
Dec 19, 2014
1,324
520
136
If RDNA represents a compute performance regression vs. Vega20 on a per-CU basis, then some folks are not going to be happy.
I don't see why, RDNA2 will likely have an 80 CU maximum at least, and possibly higher clock speeds - at worst whatever losses will be offset by gains in clock and raw silicon.

Besides which, I seem to remember the whole spiel over the wave32 design of RDNA being about higher efficiency, less wasted utilisation in the waves/warps or whatever they are called, it seems unlikely that they would have said that if they expect a significant drop in compute.

I expect that there may be some tasks which are better in wave32 and some better in wave64, only time and a thorough review of RDNA2 and CDNA will reveal the outcome, albeit not with ML which obviously will benefit from the tensor units in CDNA.
 

soresu

Golden Member
Dec 19, 2014
1,324
520
136
RDNA1, and specifically Navi12/GFX1011 in particular has more feature sets than Vega20/GFX906 except half rate DPFP.
Not unexpected in general - I don't think that AMD have done many cards that even do half rate FP64.

I'm pretty sure Vega10 didn't, or Fiji - was there even one other card from AMD since Hawaii that could manage it?
It has MFMA with bfloat16 types.
Machine Learning FMA? Matrix FMA?
 

Timorous

Senior member
Oct 27, 2008
389
148
116
Yeah exactly it the exact same die developed for mobile but with a diferent package, they dont develop APUs for desktop.
Is not the same as for example, Matisse, it shares most of the development but is not the same server CPU packaged for AM4.

APUs had been always like this, and if im not mistaken, OEMs always got the new APUs first, it looks like everything still the same to me. Maybe they just understimated the Picasso demand in desktop.
To reinforce the point you are making.

Threadripper is to EPYC as desktop APUs are to mobile APUs.
 

Thicy1988

Junior Member
May 5, 2020
1
0
6
The usual questions spring to mind:
  1. Will it support currently existing motherboards (300/400/500 series chipsets)?
  2. What kind of IPC increase are we talking about?
  3. Will AMD manage to squeeze more frequencies?
  4. What node will it use?
  5. What will be its TDP?
  6. Will it support AVX512 instructions?
  7. When and if we can expect Ryzen 4000 CPUs with modern onboard graphics (e.g. Navi10/Navi20)?
It has supper graphic which is processing more productive playing game and other things will be easily move
 

soresu

Golden Member
Dec 19, 2014
1,324
520
136
On a per-CU basis, it's doing better than VegaFE with its 64 CU, but the 5700XT has a clockspeed advantage.
Precisely - at worst it's not doing much better per CU, but with an 80 CU model and perhaps even higher clock speeds it will be a step forward, and then RDNA3 will be better still.

Having said that if the reports of Hopper being chiplet based are true, I hope that RDNA3 is too.

Methinks further GPU conversation should be moved elsewhere.
 
  • Like
Reactions: Tlh97 and Thibsie

ASK THE COMMUNITY