Speculation: Ryzen 3000 series

[maddie]

Concerning the Ryzen 3000 series, I'm thinking that the fastest 8C CPUs could have 2 chiplets. Remember the chiplet binning process to enable a wider range of clocks, low power, etc in the assembled CPU? This also works within a chiplet.

If you disable the 4 slowest cores and only enable the fastest 4 cores, the possibility exists for some ultra fast CPUs relative to a 1 chiplet 8C. A 50th anniversary special edition?

 

[jpiniero]

If you disable the 4 slowest cores and only enable the fastest 4 cores, the possibility exists for some ultra fast CPUs relative to a 1 chiplet 8C. A 50th anniversary special edition?

Have to think the 1% dies will be reserved for Epyc and/or Threadripper.

 

[PotatoWithEarsOnSide]

I'm not sure where @PeterScott bought his logic from, but I think that he was conned.
Every additional AMD sale removes Intel's margin on that sale from 40% to 0%. Their own margin is immaterial just so long as the bottom line still expands healthy, which it will with 25% margins in place of 30% margins. We're not talking huge swings in market share being needed to produce more profits overall. Meanwhile, Intel makes a 100% margin on no sale.

 

[krumme]

I'm not sure where @PeterScott bought his logic from, but I think that he was conned.
Every additional AMD sale removes Intel's margin on that sale from 40% to 0%. Their own margin is immaterial just so long as the bottom line still expands healthy, which it will with 25% margins in place of 30% margins. We're not talking huge swings in market share being needed to produce more profits overall. Meanwhile, Intel makes a 100% margin on no sale.

Absolutely. The fixation on margins is from the stone age.
Look at Amazon value and profit while old style retail with higher margins is a train wreck.

 

[maddie]

Have to think the 1% dies will be reserved for Epyc and/or Threadripper.

What I mean is that within each chiplet, you will have a range of performance values for each core. The possibility exists for a chiplet to score poorly due to the average being depressed by a few bad cores. The chiplet will NOT be a 1% full 8C. but can be an extreme 4C if the slower 4 cores are fused off. These die will probably never be used for threadripper in the first place.

Think of superimposed bell curves. First is with full chiplets as in the AMD research paper, and a second 4core distribution within the chiplet itself.

 

[Topweasel]

Concerning the Ryzen 3000 series, I'm thinking that the fastest 8C CPUs could have 2 chiplets. Remember the chiplet binning process to enable a wider range of clocks, low power, etc in the assembled CPU? This also works within a chiplet.

If you disable the 4 slowest cores and only enable the fastest 4 cores, the possibility exists for some ultra fast CPUs relative to a 1 chiplet 8C. A 50th anniversary special edition?

No it wouldn't work this way if anything they would do something like a a 8c 3600x Black Edition and that one would be the one chiplet version with max clocks.

You see it a little bit in the 1500x and 1600x AMD was able to eek out an extra 100mhz but really power wise by cutting of the cores the savings aren't that much. We see the same thing with the SL-X well. Part of this is because compared to a ring bus arch, the mesh's take a lot of the power requirement for a CPU and that doesn't go down with less cores. So using 2 dies will hurt the power usage and because of that the top clocks. AMD would be better of trying to push for a little extra for a single die version as the superior version and advertise it as the 9900k killer.

 

[Mockingbird]

I'll address each point in turn.

With their 1 CPU design, Zen1 and now Zen2 for the most, by far expensive part, I tend to believe that AMD might have the advantage here in contrast to Intel with a host of separate designs. Inventory costs alone should be substantial. The ability to be extremely agile in filling orders is another. AMD can produce whichever CPU they want with much lower lead time than Intel. A big, often neglected cost saving. A design cost amortized over their entire pure CPU lineup (vs APU) is another big cost saving. I cannot accept your 1st statement as fact.

I hardly think that the operational cost of a Fab is low. I'm open to being corrected but any extreme hi-tech manufacturing operation is not trivial. I think you're underrating this cost.

The fact that Intel has been unable to get their production of 3rd party designs successful probably have many factors but one reason is they might not be that much of a lower cost than TSMC to justify the move. Just because you produce inhouse does not guarantee being cheaper. The outsourcing in many industries testify to this. TSMC is a specialized production powerhouse, arguably second to none at present.

All of this provides doubt to your claim of "So, the cost for Intel to produce each unit is much lower than the cost for AMD to produce each unit." Not a given.

And this "As a result, if Intel is selling each unit at cost, AMD wouldn't be able to match the price unless AMD is selling each unit at a loss". Not a given

Well, everyone around here is making assumptions about things that are not publicly known, so I am entitled to do the same.

For example, people keep saying that 3rd gen Ryzen processors are cheaper to make than 2nd generation Ryzen processors, yet there is no public information about how much 2nd or 3rd generation Ryzen processors cost to make.

 

[PeterScott]

I'm not sure where @PeterScott bought his logic from, but I think that he was conned.
Every additional AMD sale removes Intel's margin on that sale from 40% to 0%. Their own margin is immaterial just so long as the bottom line still expands healthy, which it will with 25% margins in place of 30% margins. We're not talking huge swings in market share being needed to produce more profits overall. Meanwhile, Intel makes a 100% margin on no sale.

Lets wait and see who turns out to be right: The realists, or the dreamers.

You only have to look at GPU pricing to see AMD chases margins over market share, once it actually has enough market share to have a vested interest in protecting it's margins.

 

[maddie]

No it wouldn't work this way if anything they would do something like a a 8c 3600x Black Edition and that one would be the one chiplet version with max clocks.

You see it a little bit in the 1500x and 1600x AMD was able to eek out an extra 100mhz but really power wise by cutting of the cores the savings aren't that much. We see the same thing with the SL-X well. Part of this is because compared to a ring bus arch, the mesh's take a lot of the power requirement for a CPU and that doesn't go down with less cores. So using 2 dies will hurt the power usage and because of that the top clocks. AMD would be better of trying to push for a little extra for a single die version as the superior version and advertise it as the 9900k killer.

Are you certain of this in Zen2? Their 25/20 has them some way to go.

In any case, I'm not arguing for lower power per se, but ability to clock an extreme 8C higher than a 1 chiplet version.

A 1chiplet 8C cannot clock an all core boost higher than the slowest core.
A 2chiplet 8C version can. Almost by definition, each core is different and so you can always select 4 cores of an 8 core chiplet to be above the average of any chiplet, even a top 1% one. Two of these chiplets will enable a faster overall CPU.

One possibility. You take the top clocking 1% die and further test the individual cores to select the top 4C combinations with the overall 1% grouping. These are the die you use for the 50th anniversary edition. An extra 10% in clocks? Not possible? I think it is. How many would want to buy this even with more power consumed?

 

[jpiniero]

but can be an extreme 4C if the slower 4 cores are fused off. These die will probably never be used for threadripper in the first place.

If AMD has the clocks and IPC, they could try to go after the HFT market (see the 9900XE).

 

[krumme]

Lets wait and see who turns out to be right: The realists, or the dreamers.

You only have to look at GPU pricing to see AMD chases margins over market share, once it actually has enough market share to have a vested interest in protecting it's margins.

Well it becomes a tad theoretical this discussion. Once we know the latency numbers for handling expensive server loads and the consistency in that latency across a variety of loads we will know what is left for consumer market. I think that more than anything dictates price in practice.
Zen lacked predictability for many server loads and it hurt prices and uptake. Not because of some arma 3 dx 9 bm.

 

[Mockingbird]

I'm not sure where @PeterScott bought his logic from, but I think that he was conned.
Every additional AMD sale removes Intel's margin on that sale from 40% to 0%. Their own margin is immaterial just so long as the bottom line still expands healthy, which it will with 25% margins in place of 30% margins. We're not talking huge swings in market share being needed to produce more profits overall. Meanwhile, Intel makes a 100% margin on no sale.

AMD can’t take market share from Intel if AMD already dominated the market and AMD already dominated the ~$200 and under market.

It’s the $300+ market that Intel dominates and that’s where AMD can take marketshare from Intel.

Products such as the $329 8C/16T Ryzen 7 3700X (with similar performance to the Core i9-9900K) and $499 12C/24T Ryzen 9 3800 (with 50% more cores/threads than Core i9-9900K) would allow AMD to do that.

 

[Topweasel]

Are you certain of this in Zen2? Their 25/20 has them some way to go.

In any case, I'm not arguing for lower power per se, but ability to clock an extreme 8C higher than a 1 chiplet version.

A 1chiplet 8C cannot clock an all core boost higher than the slowest core.
A 2chiplet 8C version can. Almost by definition, each core is different and so you can always select 4 cores of an 8 core chiplet to be above the average of any chiplet, even a top 1% one. Two of these chiplets will enable a faster overall CPU.

One possibility. You take the top clocking 1% die and further test the individual cores to select the top 4C combinations with the overall 1% grouping. These are the die you use for the 50th anniversary edition. An extra 10% in clocks? Not possible? I think it is. How many would want to buy this even with more power consumed?

What you are talking about is an issue for Intel to worry about on SL-X. This isn't a worry for the chiplets because AMD will get so many of them per wafer. I mean look at it now. Some of AMD's best clocks come from dies with all cores enabled. Hell probably the only reason they stepped it down a bit on the 1800x had more to do with staying within TDP and not because it couldn't hit the clocks.

But on the chiplets the dies are so small I doubt there is much clock limit deviation between cores on more than a handful of chips for every two chips they can find where only some cores can hit a higher clock I am guessing they will be able to find a dozen chips that get nearly as high. But you can't ignore power levels.

So lets say AMD has a 2 Chips that can get 4.8GHz but only on about 2 cores per CCX. Each chiplet at this point might be running at 65w per. Then lets add a really gentle 10w for the IO die. That's a 140w CPU.

Now AMD has a single die maybe you are right, it can only get to 4.7GHz. But with all 8 cores. It does this at 90w for the chiplet, 10w for the IO die making it a 100w CPU (probably marketed as a 105w or 110w (forget which they are likely to use)). Which one is AMD more likely to use. Specially since two dies can then be used in non-X 4/6/8/12/16 CPU's

 

[Topweasel]

Lets wait and see who turns out to be right: The realists, or the dreamers.

You only have to look at GPU pricing to see AMD chases margins over market share, once it actually has enough market share to have a vested interest in protecting it's margins.

Vega VII and even Vega itself are bad examples of it. While Polaris and other GDDR cards of it's ilk has been holding AMD back they have been well priced for a while. MSRP wise Vega was actually pretty decently priced and that's knowing that most would still end up being up sold for inflated prices to miners. But Vega VII is a small volume card with drastically increase production costs of a rebranded multi-thousand dollar enterprise card. I doubt they could go much cheaper and if they did it wouldn't matter. AMD is selling them specifically because they can at least sell some but I think are indifferent to how it affects the market because they never intended to sell it in retail in the first place. The market is even worse there. The dGPU market wants AMD to instigate a price war specifically to be able to buy Nvidia hardware cheaper. It's really really weird there.

But none of this is to say AMD won't chase margin. They have and they will. They will in AM4 as well. But I don't think that means a $750-$900 16c AM4 chip somewhere around $500 and $600 where they will actually sell these CPU's is the point. AMD already has a low volume high margin enthusiast CPU platform. That isn't AM4. I can understand AMD holding back 16c chips till later. But I doubt AMD who will want to actually sell it's CPU's will launch a CPU priced so far out of it's market to make it a low volume chip.

 

[maddie]

Lets wait and see who turns out to be right: The realists, or the dreamers.

You only have to look at GPU pricing to see AMD chases margins over market share, once it actually has enough market share to have a vested interest in protecting it's margins.

I think you might be arguing from a mistaken position of the opposing side.

I for one, and most of the ones arguing for prioritizing marketshare, have never denied the belief that if AMD was dominant, they would persue a similar high margin strategy as Intel.

Of course AMD will try to increase margins if they have a healthy or dominant marketshare. We're saying that for now and the next couple years their best strategy is to chase marketshare and settle for a lesser margin than they could achieve otherwise. You said it yourself.

"once it actually has enough market share to have a vested interest in protecting it's margins"

This is for the future and no one argued otherwise. You're ascribing a viewpoint that no one here actually forwarded. Strawman?

 

[Mockingbird]

I think you might be arguing from a mistaken position of the opposing side.

I for one, and most of the ones arguing for prioritizing marketshare, have never denied the belief that if AMD was dominant, they would persue a similar high margin strategy as Intel.

Of course AMD will try to increase margins if they have a healthy or dominant marketshare. We're saying that for now and the next couple years their best strategy is to chase marketshare and settle for a lesser margin than they could achieve otherwise. You said it yourself.

"once it actually has enough market share to have a vested interest in protecting it's margins"

This is for the future and no one argued otherwise. You're ascribing a viewpoint that no one here actually forwarded. Strawman?

AMD can’t take market share from Intel if AMD already dominated the market and AMD already dominated the ~$200 and under market.

It’s the $300+ market that Intel dominates and that’s where AMD can take marketshare from Intel.

Products such as the $329 8C/16T Ryzen 7 3700X (with similar performance to the Core i9-9900K) and $499 12C/24T Ryzen 9 3800 (with 50% more cores/threads than Core i9-9900K) would allow AMD to do that.

 

[maddie]

What you are talking about is an issue for Intel to worry about on SL-X. This isn't a worry for the chiplets because AMD will get so many of them per wafer. I mean look at it now. Some of AMD's best clocks come from dies with all cores enabled. Hell probably the only reason they stepped it down a bit on the 1800x had more to do with staying within TDP and not because it couldn't hit the clocks.

But on the chiplets the dies are so small I doubt there is much clock limit deviation between cores on more than a handful of chips for every two chips they can find where only some cores can hit a higher clock I am guessing they will be able to find a dozen chips that get nearly as high. But you can't ignore power levels.

So lets say AMD has a 2 Chips that can get 4.8GHz but only on about 2 cores per CCX. Each chiplet at this point might be running at 65w per. Then lets add a really gentle 10w for the IO die. That's a 140w CPU.

Now AMD has a single die maybe you are right, it can only get to 4.7GHz. But with all 8 cores. It does this at 90w for the chiplet, 10w for the IO die making it a 100w CPU (probably marketed as a 105w or 110w (forget which they are likely to use)). Which one is AMD more likely to use. Specially since two dies can then be used in non-X 4/6/8/12/16 CPU's

I agree it all depends on the details, and I'm not talking about a regular SKU but a special edition one, maybe for the 50th anniversary. The ultimate performing 8C desktop CPU. Even at a much higher price, I think many here would consider buying it.

You always get a variation in cores. The diffusion process, the etching, so many others over so many mask stages, pretty much guarantee it. Also I'm expecting some improvement in idle power used by the fabrics. I'm assuming that Zen1 is not the ultimate possible.

 

[OTG]

I for one, and most of the ones arguing for prioritizing marketshare, have never denied the belief that if AMD was dominant, they would persue a similar high margin strategy as Intel.

Thanks for bringing this up. This distinction is fairly significant; Zen 2 is AMD's biggest chance, maybe ever, to get lots of market share, really quickly.
Say what you want about Intel's business practices, they know how to make a CPU, and they'll happily spend what it takes to do so. This next generation or two will be pretty rough for them, but it would be foolish to think they won't catch up.
When they do, marketshare gains (for AMD) will be much slower, and margins for both companies will drop from competition.
If they go for margins now, let's say they get to 35% marketshare in five years. Then say they disregard margins, price Zen 2 to move, and get to 40%.
They may give up millions, or tens of millions in sales to get that 5%, but gain billions in the long run.

 

[maddie]

Well, everyone around here is making assumptions about things that are not publicly known, so I am entitled to do the same.

For example, people keep saying that 3rd gen Ryzen processors are cheaper to make than 2nd generation Ryzen processors, yet there is no public information about how much 2nd or 3rd generation Ryzen processors cost to make.

Not really.

People are saying that it's quite possible that Ryzen is cheaper to make. You on the other hand are saying outright, that Intel can make cheaper processors.

 

[OTG]

I agree it all depends on the details, and I'm not talking about a regular SKU but a special edition one, maybe for the 50th anniversary. The ultimate performing 8C desktop CPU. Even at a much higher price, I think many here would consider buying it.

I wonder if a 16-core might not be their 50th anniversary special.
It would give them a totally dominant CPU for mainstream (hedt?), plus they could crank up the TDP, and still be able to say that Ryzen TDPs are the same as the first couple generations.
Call it the 50th Anniversary Turbo Black Suck It Intel edition or whatever, make it only compatible with x570 boards that meet power delivery requirements, charge whatever they want and spend the extra margins on a swimming pool full of champagne for the company party.

 

[maddie]

AMD can’t take market share from Intel if AMD already dominated the market and AMD already dominated the ~$200 and under market.

It’s the $300+ market that Intel dominates and that’s where AMD can take marketshare from Intel.

Products such as the $329 8C/16T Ryzen 7 3700X (with similar performance to the Core i9-9900K) and $499 12C/24T Ryzen 9 3800 (with 50% more cores/threads than Core i9-9900K) would allow AMD to do that.

I really can't see how you keep stating this as fact. When you say AMD dominates the $200 and lower market, what do you mean? Do you really believe that AMD sells more CPUs than intel in that range? Seeing that most CPU sales are ones less that $200, then are they already outselling Intel?

 

[maddie]

Thanks for bringing this up. This distinction is fairly significant; Zen 2 is AMD's biggest chance, maybe ever, to get lots of market share, really quickly.
Say what you want about Intel's business practices, they know how to make a CPU, and they'll happily spend what it takes to do so. This next generation or two will be pretty rough for them, but it would be foolish to think they won't catch up.
When they do, marketshare gains (for AMD) will be much slower, and margins for both companies will drop from competition.
If they go for margins now, let's say they get to 35% marketshare in five years. Then say they disregard margins, price Zen 2 to move, and get to 40%.
They may give up millions, or tens of millions in sales to get that 5%, but gain billions in the long run.

One thing that appears relevant.

This new AMD board and management looks to be playing a very long game with no 2nd class role accepted. Catmerc has mentioned that Zen3 appears to be another radical design move. They appear to want people to upgrade as in the old days, generation to generation. No more sitting on a CPU because it's good enough.

I suspect that those buying 6C/12T CPUs with the intention of it lasting 5,6,7 years as has been the recent experience will be surprised, especially if the new consoles can start to usefully utilize a full 16 threads. This is from where the new gaming experiences will be advanced.

 

[maddie]

I wonder if a 16-core might not be their 50th anniversary special.
It would give them a totally dominant CPU for mainstream (hedt?), plus they could crank up the TDP, and still be able to say that Ryzen TDPs are the same as the first couple generations.
Call it the 50th Anniversary Turbo Black Suck It Intel edition or whatever, make it only compatible with x570 boards that meet power delivery requirements, charge whatever they want and spend the extra margins on a swimming pool full of champagne for the company party.

My problem with this, is that cores are ubiquitous with AMD right now. Outside of servers you can get from 2 to 32 cores between AM4 and TR4 and possibly rising to 64C with TR2. More cores don't really stand out.

A 5.5 GHz 8C, if possible, makes a bigger statement in my view. Fastest CPU in the world.

 

[scannall]

My problem with this, is that cores are ubiquitous with AMD right now. Outside of servers you can get from 2 to 32 cores between AM4 and TR4 and possibly rising to 64C with TR2. More cores don't really stand out.

A 5.5 GHz 8C, if possible, makes a bigger statement in my view. Fastest CPU in the world.

While IBM has pushed a 5.5 Ghz CPU out the door at stock, it was a pretty exotic affair. zEC12

We're kinda at the wall on clock speeds with current materials. Without exotic cooling, able to warranty, not catch the house on fire etc. If they can come in at 4.8 to 5.0 Ghz stock, with included cooler that will be more than good enough.

 

[OTG]

[QUOTE="maddie, post: 39709781, member: 278074"

A 5.5 GHz 8C, if possible, makes a bigger statement in my view. Fastest CPU in the world.[/QUOTE]

How long before they can manage a 5ghz 16C?
This gen is probably too soon, but someday...