Speculation: Ryzen 3000 series

Page 8 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

What will Ryzen 3000 for AM4 look like?


  • Total voters
    230

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
AMD is going to use a huge amount of cream-of-the-crop Epyc chips for this bad boy alone and, if we apply the "rule" of only the best 5% of chiplets go to Epyc and only the best of those goes to the top chip, than that's a heck of allot "not good enough" 8c chiplets after the binning process.
I'm pretty sure initial yields are much less ... shall we say ... favorable ... then AMD would have us believe, simply because it's a new process. This will ofc improve with time and then it's quite possible AMD will end up sacrificing perfectly good chips for lower solutions, depending on the needs.

I expect it to happen right away personally. But it will probably be tiered. There will always be the issue of cheaper processors selling in more volume. I think it's a while before AMD disables an 8c chiplet down to 2 or even 4 cores. But I expect most of the chips on a 4c Ryzen to be made up of 2 probably 4+ core capable dies, 6c CPU's with again 4c+ dies, 8 for 6c+ capbable dies 10 and 12 with 8c capable dies.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
Maybe. The Xbox One and PS4 both had 8 cores. I believe 1 core was set for the OS, so the games had 6-7 cores to access. We still saw very little progress for multi core gaming.
Not true. I just think people are expecting some kind of instantaneous migration. There are a lot more games out there now that utilize more Mt than people realize. Ones that on PC even see significant performance upgrades with more cores. But it's a slow switch, it wasn't till late 2015 early 2016 that games started to hit the market that were designed from the start with the PS4/Xbone as primary development platforms. One of the reasons we don't see significant performance increases with more cores on a lot of new games is due to how simplified the coding is for CPU work because of the Cat cores they are using. All that changes if they are using Zen cores. Now the primary development platform will be on the "enthusiast" side of feature sets (even if its a bit midgrade for us). 8 Medium clocked Zen cores will go a long way to pushing what should be much simpler ports, for PC's, into driving high core count processing requirements for games in the future.
 
  • Like
Reactions: lightmanek

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I expect it to happen right away personally. But it will probably be tiered. There will always be the issue of cheaper processors selling in more volume. I think it's a while before AMD disables an 8c chiplet down to 2 or even 4 cores. But I expect most of the chips on a 4c Ryzen to be made up of 2 probably 4+ core capable dies, 6c CPU's with again 4c+ dies, 8 for 6c+ capbable dies 10 and 12 with 8c capable dies.

You don't have to use 2 dies for Ryzen, since the memory controller is on the IO die.
You could do something like this and it would work just fine:

R9: 2 6-core dies
R7: 2 4-core dies
R5: One 6 core die
R3: One 4 core die

I guess AMD could do something like have the R3 be a 6 core Zen 1, maybe for WSA reasons.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,229
9,990
126
If there is a $100 6-core part, it would almost certainly have a locked multiplier
That's a possibility, I suppose, and it would go along with the idea that those CPUs were for OEMs. (Although, the lack of an iGPU would seem to me to put off OEMs from integrating it, if they had to provide an additional dGPU just to get things to function.)
 

DarthKyrie

Golden Member
Jul 11, 2016
1,531
1,279
146
Maybe. The Xbox One and PS4 both had 8 cores. I believe 1 core was set for the OS, so the games had 6-7 cores to access. We still saw very little progress for multi core gaming.

The cores in PS4 and Xbone are weak Jaguar cores, the cores that will be in PS5 and NeXtbox should be 8 Zen2 cores, imagine what developers will be able to do with that much power at hand.
 

realibrad

Lifer
Oct 18, 2013
12,337
898
126
Not true. I just think people are expecting some kind of instantaneous migration. There are a lot more games out there now that utilize more Mt than people realize. Ones that on PC even see significant performance upgrades with more cores. But it's a slow switch, it wasn't till late 2015 early 2016 that games started to hit the market that were designed from the start with the PS4/Xbone as primary development platforms. One of the reasons we don't see significant performance increases with more cores on a lot of new games is due to how simplified the coding is for CPU work because of the Cat cores they are using. All that changes if they are using Zen cores. Now the primary development platform will be on the "enthusiast" side of feature sets (even if its a bit midgrade for us). 8 Medium clocked Zen cores will go a long way to pushing what should be much simpler ports, for PC's, into driving high core count processing requirements for games in the future.

The point was that we have had 6-7 cores on consoles for a while. We have had multi core desktops with 4 and high end systems with more than 4 cores. Its possible that we see more usage, but, I think its also possible that we don't see much progress.
 

realibrad

Lifer
Oct 18, 2013
12,337
898
126
The cores in PS4 and Xbone are weak Jaguar cores, the cores that will be in PS5 and NeXtbox should be 8 Zen2 cores, imagine what developers will be able to do with that much power at hand.

Oh for sure the cores of Zen blow Jaguar cores away. But, the incentive is actually not so clear here. More slower cores should mean that to get out extra perf you have to multi-thread your games. You suddenly present cores that are drastically faster, and the incentive there is to max out a few faster cores.

Game companies are lazy and when you give them the option to take a shortcut, they often do. Once games max out the much faster cores, then they will push their code to use more cores.

So because these cores are so much faster, we may not see the growth we want until much later in the cycle. That I think is a reasonable possibility.
 

DarthKyrie

Golden Member
Jul 11, 2016
1,531
1,279
146
The point was that we have had 6-7 cores on consoles for a while. We have had multi core desktops with 4 and high end systems with more than 4 cores. Its possible that we see more usage, but, I think its also possible that we don't see much progress.

Intel kept max cores on the desktop at 4 cores for so long that most developers have only really ever coded for a max of 4 cores. AMD had to release an 8c/16t desktop part to advance everything forward itself because Intel wasn't going to do it willingly. Why should developers pour time and money into taking advantage of more cores than are available to the PC segment?
 

realibrad

Lifer
Oct 18, 2013
12,337
898
126
Intel kept max cores on the desktop at 4 cores for so long that most developers have only really ever coded for a max of 4 cores. AMD had to release an 8c/16t desktop part to advance everything forward itself because Intel wasn't going to do it willingly. Why should developers pour time and money into taking advantage of more cores than are available to the PC segment?

I hope they do use more than 4 cores now that we have Zen pushing intel. My worry is that the big boost in console cores will create a lag. I think eventually it will have the effect of pushing more cores but, that big jump will make a lot of companies take shortcuts in the beginning.

But, it could be the case that PCs having more cores could push coding to be more multi threaded for consoles. It will be interesting.

Last gen consoles were as close to PC as they have ever been. This next gen should bring them even closer.
 

DarthKyrie

Golden Member
Jul 11, 2016
1,531
1,279
146
I hope they do use more than 4 cores now that we have Zen pushing intel. My worry is that the big boost in console cores will create a lag. I think eventually it will have the effect of pushing more cores but, that big jump will make a lot of companies take shortcuts in the beginning.

But, it could be the case that PCs having more cores could push coding to be more multi threaded for consoles. It will be interesting.

Last gen consoles were as close to PC as they have ever been. This next gen should bring them even closer.

This is AMD's plan, they are providing the ecosystem to bring gaming together and forcing the issue of supporting more than 4 x86-64 cores in game engines.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
I guess AMD could do something like have the R3 be a 6 core Zen 1, maybe for WSA reasons.

They probably don't need to. If you look at the massive size of the one for Epyc, even a smaller one for Ryzen/TR is going to take up plenty of wafers. The chiplets are small enough that AMD can get hundreds of functional parts per wafer. Between that an the Polaris GPUs, I don't think AMD has too many issues meeting their requirements.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
You don't have to use 2 dies for Ryzen, since the memory controller is on the IO die.
You could do something like this and it would work just fine:

R9: 2 6-core dies
R7: 2 4-core dies
R5: One 6 core die
R3: One 4 core die

I guess AMD could do something like have the R3 be a 6 core Zen 1, maybe for WSA reasons.
It's possible but I am discounting it for now. On TR and Epyc because of the Rome design I suspect they could use 4 instead of 8 chips as doing so won't unbalance the cooling. But My personal guess is that all AM4 Zen2 based CPU's will be 3 chip affairs if 16c Ryzen is going to exist. If for only because I get the feeling not dedicating manufacturing lines to single chip versions will be cheaper and not wanting to unsettle the cooling balance where 2 chips would skew the heat on the heatplate to one side.
 

inf64

Diamond Member
Mar 11, 2011
3,686
3,958
136
My guess for 3700X CB R15 scores.
Assumption: 3700X is 12C/24T SKU running at 4.2Ghz base, 4.4Ghz all core turbo and 4.7Ghz ST Turbo. Zen2 has around 15% higher average IPC (just rough number to get it in the perf. estimate) For the record I expect it to have this much if not higher IPC vs Zen1.

CB R15 MT : ~1.9x the score of 2700x~=3300 pts
CB R15 ST: ~1.24x the ST score of 2700X ~=220 pts
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
That might be possible with HBM3 or HBM4 but not HBM or HBM2.

Are there even any solid release schedules for HBM3? I'll admit I haven't been paying much attention to it, but anything I've seen says 2019 at the earliest, with 2020 being much more likely. It might even be later than that if the past is any history as HBM has always fallen a little short of what the performance should have been and it's always been a pain to get the kinds of yields that make it economical outside of the top-end.

If AMD really wanted to, couldn't they just use an IO die and some of the extra extra memory channels paired with some plain-Jane DDR4 in order to alleviate the memory bottleneck? It's probably not quite that simple, but the general approach seems easier.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
The point was that we have had 6-7 cores on consoles for a while. We have had multi core desktops with 4 and high end systems with more than 4 cores. Its possible that we see more usage, but, I think its also possible that we don't see much progress.
My Point is we do see games spawning more the threads but for now outside about 5-8 games the upper end is still capped by the overall compute ability of the CPU''s in the console and any post port optimization has been for 4c8t CPU's. Drive up the compute power on the consoles and in 3 years base compute power of the Consoles and to a degree their thread count become a bare min on PC.
 
  • Like
Reactions: OTG and DarthKyrie

Zapetu

Member
Nov 6, 2018
94
165
66
1. The chips that EPYC will like , low leakage , low clocking are the polar opposite of the chips we want, high leakage , high clocking chips

That's exactly right. EPYCs (low leakage, high perf/W) come mostly from different bin than TRs (high leakage, high clockspeeds) . There are some (low volume) exceptions and likely Rome will have similar SKUs (16C or 32C). Still it's important to be aware that the top 5% bin of chips for ThreadRippers is using a different metric than the top X% (*) bin of chips for (higher end) EPYCs. Still as we know, most of everything (including chips) tends to be avarage.

2. A chiplet is smaller then the A12 and it has been in production for 1/2 a year right now, so 1 full year next june

I looks like 7nm volume is not going to be a limiting factor for a while at least:
AMD’s Zen 2 and Navi supply won’t be limited by TSMC’s 7nm chip manufacturing

So if AMD is confident with their product line, they should utilize the free capacity if necessary, even if it takes for a while to ramp up the production.

3. much like the 6 core ryzen i think there will be plenty of "sacrificing" of fully functional chips for lower end product right off the bat.

True, but that's how it always goes. Still when yields are lower, there's more of those salvageable partially working dies. That's why later on (when yields get better) it would be very beneficial for AMD and Sony/MS to share the benefits of the binning from the same much larger pool of 8C chiplets.

Edit: (*) Fixed 5% to X% because as H T C suggested later on, we're just speculating on percentages here.
 
Last edited:

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
True, but that's how it always goes. Still when yields are lower, there's more of those salvageable partially working dies. That's why later on (when yields get better) it would be very beneficial for AMD and Sony/MS to share the benefits of the binning from the same much larger pool of 8C chiplets.

It's one of the things I really liked about the Adored video. If he is right that the IO chip is as large as it is because it's mostly a dozen different memory controllers. Then AMD didn't find a way to minimize yield issues. They found a way to erase them. The only chips AMD wouldn't be able to use are ones that on the outer edge that aren't fully manufactured.
 

H T C

Senior member
Nov 7, 2018
549
395
136
That's exactly right. EPYCs (low leakage, high perf/W) come mostly from different bin than TRs (high leakage, high clockspeeds) . There are some (low volume) exceptions and likely Rome will have similar SKUs (16C or 32C). Still it's important to be aware that the top 5% bin of chips for ThreadRippers is using a different metric than the top 5% bin of chips for (higher end) EPYCs. Still as we know, most of everything (including chips) tends to be avarage.

That's not what i was trying to get across: if 5% of total chiplets are "Epyc 64c / 128t chip worthy" and the super-computer i mentioned earlier uses 80K of those, that's 1.6M chiplets right there (total pool from which the 80K were selected from). But i seriously doubt 5% of all chiplets will make the cut for 64c / 128t Epyc chip: doubt it will reach 1%, even, and that means the total chiplet pool could be as high as 8M chiplets or more. 5% might make the cut for the lowest of Epyc chips but the highest? Seriously doubt it.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
AMD sold 2 Million Ryzen's in first month of its launch last year. This was when AMD was at the lowest market share they ever were at and the only available CPUs were the three most expensive.

Even your extremely harsh 1% was true, it's not like AMD will drop off 80k of CPUs at once. This will multi monthly or maybe even a year of AMD dropping off thousands at a time. But at the pool of 8 million chiplets. AMD will have over a dozen non server configurations to use these dies. Guessing a min of 2 dies per CPU (so looking at 4 million chiplets if this was launched in March of last year). I don't see how that 8 million seems that big of a stretch. That might be less than one month of production to make all those chiplets.
 
  • Like
Reactions: xjfv

Zapetu

Member
Nov 6, 2018
94
165
66
That's not what i was trying to get across: if 5% of total chiplets are "Epyc 64c / 128t chip worthy" and the super-computer i mentioned earlier uses 80K of those, that's 1.6M chiplets right there (total pool from which the 80K were selected from).

While AMD has stated that ThreadRippers use chips from the top binned 5% , we don't know what the portion of the top binned chips for EPYCs is. So while your calculations are otherwise correct, we don't know if that 5% stands for EPYCs also. I'll admit that I should not have put that 5% there again, as beginner99 earlier pointed out, even if none of us have the exact right figures and I was only referring to 'high end EPYCs'.

But i seriously doubt 5% of all chiplets will make the cut for 64c / 128t Epyc chip: doubt it will reach 1%, even, and that means the total chiplet pool could be as high as 8M chiplets or more. 5% might make the cut for the lowest of Epyc chips but the highest? Seriously doubt it.

That supercomputer may use lower clocking 64C core EPYCs that come from a larger bin than 5% or at least larger than 1%. Sure you can't put those high leaking chiplets there but we should not assume too much on the percentages here. I see still your point and adding consoles to the chiplet mix would help a lot as AdoredTV also speculated.