New Zen microarchitecture details

[Atari2600]

Yes they are. IC design is 2x or more (~3x in your article). Per transistor cost is higher.

But that is not what you said.

You were specifically talking about production.

 

[AtenRa]

No, your first post about it didn't include today in relation to the FX4350

really ?? RYZEN is launching next 4-8 weeks, do i have to include TODAY in every sentence ??? you are searching for straws

 

[ShintaiDK]

Those are only DIY volume from Newegg, its not surprising that enthusiasts will want 6Core High-End Desktop CPUs. Its why 6700K was selling 4x more than second Core i5 CPU in that list.

You are contradicting yourself. There is a 6 core on the list. People buy 6700K for a different reason.

Best Selling CPUs 2016—First Half
Sorted by Sales Volume (# sold)

1. Intel Core i7-6700K

2. Intel Core i5-4460

3. Intel Core i7-4790K

4. AMD FX-6300

5. Intel Pentium G3258

6. Intel Core i5-6500

7. Intel Core i7 5820K

8. AMD FX-8350

9. Intel Core i7-4790

10. AMD FX-4350

 

[ShintaiDK]

But that is not what you said.

You were specifically talking about production.

And how is it different? IC design is part of production cost so is transistor cost.

 

[ShintaiDK]

really ?? RYZEN is launching next 4-8 weeks, do i have to include TODAY in every sentence ??? you are searching for straws

Its like claiming every APU since Llano have been a sales success, despite selling less and less.

7700K also outsells 6700K today but its a meaningless fact.

 

[Atari2600]

And how is it different? IC design is part of production cost so is transistor cost.

Because design is amortised over the production run.

We've no idea what the production run is going to be right now in either rate or longevity.

The apples-to-apples comparison is the cost of making dies from wafers --- which is no doubt what you intended given the way you wrote the sentence I quoted --- and now your desperately backtracking.

Unfortunately, I've come to expect little else except FUD in your posts.

 

[itsmydamnation]

And how is it different? IC design is part of production cost so is transistor cost.

under which accounting standard?

 

[ShintaiDK]

Because design is amortised over the production run.

We've no idea what the production run is going to be right now in either rate or longevity.

The apples-to-apples comparison is the cost of making dies from wafers --- which is no doubt what you intended given the way you wrote the sentence I quoted --- and now your desperately backtracking.

Unfortunately, I've come to expect little else except FUD in your posts.

You ran out of arguments that quick?

What is the 28nm and 14FF wafer cost and what is the yield numbers for FX CPUs and Ryzen CPUs? And IC design still needs to be paid.

Trolling is not allowed
Markfw
Anandtech Moderator

 

[zinfamous]

So now you add today

If the 4C/8T Ryzen had an IGP it would be different. But lets be honest, a lot less desktops and a lot less graphics cards are sold now. And Ryzen better outsell FX4350 today. Even a 64$ Pentium wipes the FX line and it comes with an IGP too.

Yes and the world is quite different today. Feel free to link something else if you think its wrong.

wut? I think these 2 points contradict each other, assuming I get what you are saying:
--Few people really need the IGP on these chips because this market is almost always buying a dGPU. the IGP is wasted space
--PC gaming desktops AND GPU sales for mid to high end is actually up significantly over the last year, so this is a weird statement.

 

[ShintaiDK]

wut? I think these 2 points contradict each other, assuming I get what you are saying:
--Few people really need the IGP on these chips because this market is almost always buying a dGPU. the IGP is wasted space
--PC gaming desktops AND GPU sales for mid to high end is actually up significantly over the last year, so this is a weird statement.

No, what you miss is the target segment. Even if you could match a Ryzen CPU with every single discrete GPU sold you end up with what, 10 million units per quarter or less? When FX was released this number was almost twice as high.

These chips wouldn't exist without servers.

 

[Atari2600]

You ran out of arguments that quick?

This isn't an "argument", its a one sided correction of inaccurate rubbish.

What is the 28nm and 14FF wafer cost

I've already supplied estimates. Read over the posts again.

and what is the yield numbers for FX CPUs and Ryzen CPUs?

There was some data on expected flaw rates kicking around, obviously exact figures won't be in the public domain. I can't remember the expected numbers off the top of my head. But given 14nm has been around for a while now, any reasonable expectations are that it will be pretty good, indeed the respins to correct bugs should further improve this as they can stamp out production issues with each update.

 

[ShintaiDK]

This isn't an "argument", its a one sided correction of inaccurate rubbish.

Only your single point of data is right?

I've already supplied estimates. Read over the posts again.

So did I. But you didn't like it.

There was some data on expected flaw rates kicking around, obviously exact figures won't be in the public domain. I can't remember the expected numbers off the top of my head. But given 14nm has been around for a while now, any reasonable expectations are that it will be pretty good, indeed the respins to correct bugs should further improve this as they can stamp out production issues with each update.

So you just make a guess. May I point you to the first quote in this post?

 

[cytg111]

May I remind these gentlemen that winning on the internet is like winning .. ... ....... .......

 

[Atari2600]

Only your single point of data is right?
So did I. But you didn't like it.

No, I didn't like it. Particularly when it appears IBS is at odds with itself:

http://www.eetimes.com/author.asp?doc_id=1329887

How can they say 28nm is at £1.30 /100M gates in one article, then $0.92 /100M gates in another.

So you just make a guess. May I point you to the first quote in this post?

Yep. An educated guess. Usually not too far wrong.

Unless you can provide sources for flaw rates that would mean Ryzen costs 2x XV?

 

[jpiniero]

Unless you can provide sources for flaw rates that would mean Ryzen costs 2x XV?

The cost is probably somewhere inbetween really what people are arguing. But they really need to make money off of Ryzen, so going cheap doesn't make sense initially esp when there's at least some amount of pent up demand.

I'm assuming the die size is smaller than Broadwell-E is (240 mm2) but not by much.

 

[Glo.]

ZEN dies at 200mm2 TODAY could be close to 1.8x more expensive than 300mm2 Bulldozer dies at 28nm BUT,

If cheapest ZEN die price is $199, then its 2.5x times higher than cheapest Bulldozer FX4350 at $80.

4C 8T ZEN at $199 will outsell Bulldozer 2M 4T FX4350 by more than 2x
Add higher profit per die and you still make more money than selling Bulldozer dies at $80

Fottemberg have said that cost of Zen die is between 18 and 20$.

And they will sell them for more than 100$.

 

[Atari2600]

Fottemberg have said that cost of Zen die is between 18 and 20$.

And they will sell them for more than 100$.

$20 /die x 200 dies/wafer = $4000 /wafer

Which sounds ballpark.

 

[krumme]

Yeaa. Who cares if it cost 20 or 40usd to make a 8c die? 2 pages of irrelevant discussion as i can tell. This is just sidetracking to get into a barfight.

 

[leoneazzurro]

$20 /die x 200 dies/wafer = $4000 /wafer

Which sounds ballpark.

Based on this and an estimated area of 200 mm^2 then the yields should be between 60% and 70%, roughly.
BTW, die is not the only cost, we have packaging, testing, packing and delivery, and also the R&D and other indirect costs must be covered.
I don't think the final cost will be 100$ or more, more inside the 30-50$ window.

 

[Dresdenboy]

You cant as Intel either. A reality check for people, look at steam. There isn't even half a million 8 core CPUs after several years with 8 core CPU offerings. FX8xxx, FX9xxx, X series HEDT etc.

Even if you include 6 cores, then you are at 2½-3 million combined 6 and 8 cores installed over the last 5-6 years or so.

Its a very tiny market to begin with that wouldn't exist without servers.

First you should make sure that Steam doesn't count 8C and 6C FX models as 4 and 3 cores, as many versions and SPs of Windows show them like being half the physical cores, but with 2x that as logical cores. The difference between 6C/8C in Windows systems (1.8%) and Linux systems (5.5%) might be a hint. Also the Newegg CPU charts show, that still some 8C FX are being sold. With 5 years old and older CPUs still populating the Steam and other HW surveys (Mozilla, Unity), the picture of the current situation and demand might be softened quite a bit.

Where did you get absolute numbers BTW? Did you base that on ~200M users?

And there are also 45% dual cores and about as many quad cores. Based on that let's predict, that the future will be bright for 2C and 4C processors.

Yes they are. IC design is 2x or more (~3x in your article). Per transistor cost is higher.

The numbers given by you don't include the 32SHP process (with HKMG), which uses SOI wafers. If we'd know the actual costs for them instead of some cheap foundry 28nm bulk process (how many layers? HKMG?), we might do the calc. At ~57% the 8C PD's die size, the cost disadvantage per wafer will be alleviated a bit. And why do we compare these 2 or 3 processes anyway, if the older ones are out of question for a current gen product? Intel doesn't sell 32nm or 28nm BDW-E, to my knowledge.

Depends if you compare a 2M Excavator to a 4C zen part. if that is the case, it can easily be larger.
there is L3 in zen which can increase the die by 60-80% easily, although increased L3 size will probably be countered by a lack of iGPU.

another thing to consider is that 28nm has been around for "ages" and most definitely enjoying peak yields. 14nm. Compared to 28 nm, 14 nm is still in its infancy and can be argued that it will not have as high a yield. Time will however negate this factor.

The L3 in Zen is about 15 sqmm per CCX.

For the yields, it might be interesting to try a yield calculator due to the interesting effects at different die sizes. And peak would never reach 100%. So even with 14LPP, but a ~25% smaller die, yields might be 70% vs. 80%.

 

[AtenRa]

I'm not so certain about that.

I see a lot of potential for the 8C/16T in Workstations and the small server market will love it. Depending on how great XFR works, it could also be an interesting alternative for pre-build gaming rigs (ex Alienware).

Let's hope we get enough CPUs to even build our own PCs ...

I was specifically talking about the Desktop SKUs, the ones we are talking the last few page. Those SKUs are specifically for the retail market. Except of a few OEM Gaming Builders the traditional OEM systems dont use CPUs without iGPUs. ZEN Server SKUs will be used for Servers and Workstations by OEMs, but the vast majority of the Desktop RYZEN SKUs will be sold in re/e-tail from places like NewEgg, Amazon, MicroCenter etc etc.

 

[AtenRa]

So even with 14LPP, but a ~25% smaller die, yields might be 70% vs. 80%.

Bulldozer die is 315mm2, 25% smaller is 236mm2. Do you believe the 8C ZEN die will be that large ??

 

[jpiniero]

Bulldozer die is 315mm2, 25% smaller is 236mm2. Do you believe the 8C ZEN die will be that large ??

Somewhere between 200 and 240 I think.

 

[AtenRa]

Fottemberg have said that cost of Zen die is between 18 and 20$.

And they will sell them for more than 100$.

my calculations in that post was for $21 per die and 229 dies per wafer.

 

[AtenRa]

The L3 in Zen is about 15 sqmm per CCX.

If the red rectangle is 15mm2 then the die is smaller than 200mm2