AMD AT CES

[_mockingbird]

8 × 2 = 16, not 12. I don't think amd is going to release a gimped smt part like that. Otoh, I don't think the typo is that big a deal. Point it out and then stfu.

I just saw that.

I had 8C/16T in a different post and was looking at that.

 

[FrgMstr]

Please take the personal bickering elsewhere.

 

[Digital Viper-X-]

I really doubt that AMD is just going to charge just $100 more while adding an entire chiplet.

Why not?

8 Core / 16T = 1 good 8 core die
12 core / 24T = 2 partially defective 8 core dies .

 

[_mockingbird]

Why not?

8 Core / 16T = 1 good 8 core die
12 core / 24T = 2 partially defective 8 core dies .

There aren't that many defective dies.

For example, most Ryzen 5 2600 are chopped from perfectly working dies.

That's not to mention the additional packing cost of adding another die.

 

[Digital Viper-X-]

There aren't that many defective dies.

For example, most Ryzen 5 2600 are chopped from perfectly working dies.

That's not to mention the additional packing cost of adding another die.

How do you know if they're good or bad ?

either way, new process, plus, if a 6c/12t chip is $150-$200, it's entirely possible to get 12c/24t for 300-400 range

 

[_mockingbird]

How do you know if they're good or bad ?

either way, new process, plus, if a 6c/12t chip is $150-$200, it's entirely possible to get 12c/24t for 300-400 range

Ryzen 5 outsells Ryzen 7, 3:1 and I am pretty sure the yield rate isn't 25%

Industrial standard is > 90%

Even 80% is considered really bad.

Also, there was a month when AMD forgot disabled 2 cores and there ended up being a lot of 8 cores Ryzen 5 on the market

 

[_mockingbird]

How do you know if they're good or bad ?

either way, new process, plus, if a 6c/12t chip is $150-$200, it's entirely possible to get 12c/24t for 300-400 range

And here's another reason it won't happen:

For most people 6-8 cores is enough.

What that means is that AMD's average selling price (for processors) is going drop by $100 and that really hurts AMD's bottom line.

$500 to $600 12C and $700 to $800 16C would compete with Intel's Core i9 line

Threadripper has more PCIe lanes so it targets a different segment of customers.

 

[Pieter3dnow]

And here's another reason it won't happen:

For most people 6-8 cores is enough.

What that means is that AMD's average selling price (for processors) is going drop by $100 and that really hurts AMD's bottom line.

$500 to $600 12C and $700 to $800 16C would compete with Intel's Core i9 line

Threadripper has more PCIe lanes so it targets a different segment of customers.

And were back to the I3 baseless argument. I3 was the gaming cpu and you need nothing more. That worked for Intel up to the point where API became no longer bound to single core and with newer API and newer game engines there was progress beyond what Intel told us. Intel screwed everyone on this just consumers but also developers for too long.
Telling people what they need or can use is such an abstract thing because it strangles to ability to move forward. In the end there was no case to be made by the people back when I3 was the gaming cpu because the API was so limited that it could survive well up to the point where DX11 was released.
Suddenly it died out , wonder why?
AMD is not Intel your speculation is different from others but that does not mean it is true. The zen 2 leaks thread has the video from AdoredTV explaining why the prices are not too good to be true.

I would like to remind you where the original Threadripper prices ended up compared to what Intel had going for it until then......

 

[juanrga]

I wonder about the cost differences involved in separating the uncore I/O from the cores. Will this mitigate some of the cost increases involved with 7nm, or exacerbate them?

One thing that is interesting is it may allow for more salvaging. Dies with bad I/O had to be thrown out in most cases before (though how many is that, really?) even if the cores were good. I think the multi-chip model allows for more die salvaging, thusly. But packaging costs have surely risen as a result of this. And 7nm isn't cheap...

COST:

[ 7nm single die ] > [ 7nm chiplet + 14nm chiplet ] > [ 14nm single die ]

 

[c3k]

Sigh.

"X > Y" means X is "greater than" Y.

"A < B" means A is "lesser than" B.

 

[DuronBurgerMan]

There aren't that many defective dies.

For example, most Ryzen 5 2600 are chopped from perfectly working dies.

That's not to mention the additional packing cost of adding another die.

This may change with 7nm. New process, more defects. Agree about packaging cost.

 

[DuronBurgerMan]

COST:

[ 7nm single die ] > [ 7nm chiplet + 14nm chiplet ] > [ 14nm single die ]

7nm single die definitely the most expensive. Agree that chiplets are probably more expensive than 14nm single die, also. This gets a little tricky with regards to yield, though. AMD demoed a single chiplet with 8 cores. Can we have a dual chiplet with 2x 4 cores that might have otherwise either been chucked, or used for very low end parts? Also, how great is the cost difference for single chiplet 8c vs., say, a 2700X? AMD will likely have to match 2700X launch prices with their 8 core product, at least. I agree with Mockingbird that the 12c and 16c products will almost certainly have to be more expensive for there to be any decent margin on them.

 

[Digital Viper-X-]

I understand that 7nm is new, but why are we saying that a 7nm chiplet is going to cost more than a 14nm single die?
Assuming yield rates are the same, there should be more chips per wafer on 7nm than on 14nm, given that 7nm should cost more per wafer, what is the actual cost? People are just throwing statements out, without adding anything to support them. Give me some numbers.

 

[c3k]

I understand that 7nm is new, but why are we saying that a 7nm chiplet is going to cost more than a 14nm single die?
Assuming yield rates are the same, there should be more chips per wafer on 7nm than on 14nm, given that 7nm should cost more per wafer, what is the actual cost? People are just throwing statements out, without adding anything to support them. Give me some numbers.

This.

The chiplet has FEWER components than a full die. That's why it's called a "chiplet".

I don't know what the cost premium is for going from 14nm to 7nm. I would think it obvious that a 14nm chiplet would be cheaper to produce (and have fewer errors) than a 14nm full cpu. If a chiplet is cheaper, then the issue of cost comes down to whether a 7nm process erases that savings...or not.

The move to chiplets has two benefits, as I understand it: the ability to "absorb" more errors in the wafer on a per item basis (far more chiplets per wafer than full cpu dies per wafer); the ability to "tune" a cpu based on the chiplet selected while utilizing known/mature 14nm processes for the IO portion.

(The reason for using 7nm is the increased performance and lowered power req't...that goes without saying.)

 

[_mockingbird]

And were back to the I3 baseless argument. I3 was the gaming cpu and you need nothing more. That worked for Intel up to the point where API became no longer bound to single core and with newer API and newer game engines there was progress beyond what Intel told us. Intel screwed everyone on this just consumers but also developers for too long.
Telling people what they need or can use is such an abstract thing because it strangles to ability to move forward. In the end there was no case to be made by the people back when I3 was the gaming cpu because the API was so limited that it could survive well up to the point where DX11 was released.
Suddenly it died out , wonder why?
AMD is not Intel your speculation is different from others but that does not mean it is true. The zen 2 leaks thread has the video from AdoredTV explaining why the prices are not too good to be true.

How many games out there can benefit from having more than 8-cores?

I would like to remind you where the original Threadripper prices ended up compared to what Intel had going for it until then......

Intel dominated the market so AMD can disrupt the market.

The difference is that AMD already dominate the ~$200 and under (processor) market.

It doesn't make sense for AMD to drop the prices further to compete with itself.

It's the $300+ (processor) market that Intel dominate with products such as the Core, i7-8700K, Core i7-9700K, Core i9-9900K

 

[DuronBurgerMan]

The thing that excites me about competition lately is that it drives new tech, not necessarily that it drops pricing brackets (which are pretty stable over time in various market segments).

Zen wasn't exciting because it was cheap in an absolute pricing sense. It was exciting because it radically increased core count for the dollar without crippling problems like Bulldozer had.

 

[Pieter3dnow]

How many games out there can benefit from having more than 8-cores?

That same argument was used by people for the I3 , you are just recycling it. Move forward not back.

 

[Gideon]

I understand that 7nm is new, but why are we saying that a 7nm chiplet is going to cost more than a 14nm single die?
Assuming yield rates are the same, there should be more chips per wafer on 7nm than on 14nm, given that 7nm should cost more per wafer, what is the actual cost? People are just throwing statements out, without adding anything to support them. Give me some numbers.

The fabs charge quite a bit more for a 7nm wafer then a 14nm one, they have to recover R&D plus cost of the new fab. This has a good info and has cost analysis in it, as for exact costs your never going to know unless a company is willing to tell you. https://www.icknowledge.com/news/Technology and Cost Trends at Advanced Nodes - Revised.pdf

 

[_mockingbird]

That same argument was used by people for the I3 , you are just recycling it. Move forward not back.

Except that the argument wasn't true and having more than 2 cores definitely helps with gaming performance.

 

[_mockingbird]

The fabs charge quite a bit more for a 7nm wafer then a 14nm one, they have to recover R&D plus cost of the new fab. This has a good info and has cost analysis in it, as for exact costs your never going to know unless a company is willing to tell you. https://www.icknowledge.com/news/Technology and Cost Trends at Advanced Nodes - Revised.pdf

 

[Digital Viper-X-]

View attachment 135442

OK, so lets break that down

Going from 45nm -> 20nm , (a bit less than half the size, meaning nearly double the dies per wafer) is double in cost,
32nm -> 16/14nm (about the same)
14->7nm (about the same)

so why is 7nm so fundamentally different?

The only difference I see, is that there was no half node jumps like 45->32->28->20, it's like going directly from 45->20nm if we are talking about percentages.

 

[schmide]

COST:

[ 7nm single die ] > [ 7nm chiplet + 14nm chiplet ] > [ 14nm single die ]

This is one thing AdoredTV put in my head that really makes sense.

AMD is still on the hook for a ton of chips from Global Foundries. They can satisfy part of that with a bunch of 14nm chiplets.

I addition, based on yield the combinations of partially defective chips and chiplets, AMD can recover a lot of product.

 

[_mockingbird]

OK, so lets break that down

Going from 45nm -> 20nm , (a bit less than half the size, meaning nearly double the dies per wafer) is double in cost,
32nm -> 16/14nm (about the same)
14->7nm (about the same)

so why is 7nm so fundamentally different?

The only difference I see, is that there was no half node jumps like 45->32->28->20, it's like going directly from 45->20nm if we are talking about percentages.

The difference is that, lately, node size has become a marketing term rather than an actual measurement.

 

[Digital Viper-X-]

The difference is that, lately, node size has become a marketing term than an actual measurement.

Ok, well look at it at by %

7nm is 50% of 14, and 16nm is 50% of 32, and so forth etc etc...

 

[Gideon]

The difference is that, lately, node size has become a marketing term than an actual measurement.

Doesnt matter in the grand scheme of things, 7nm marketing is better then the 12nm marketing label. As long as it's better no one really cares anyway on the retail side.

 

[_mockingbird]

Doesnt matter in the grand scheme of things, 7nm marketing is better then the 12nm marketing label. As long as it's better no one really cares anyway on the retail side.

I agree with that.

 

[juanrga]

OK, so lets break that down

Going from 45nm -> 20nm , (a bit less than half the size, meaning nearly double the dies per wafer) is double in cost,
32nm -> 16/14nm (about the same)
14->7nm (about the same)

so why is 7nm so fundamentally different?

The only difference I see, is that there was no half node jumps like 45->32->28->20, it's like going directly from 45->20nm if we are talking about percentages.

You are mixing marketing node names with real node names. Node naming was correct up to 20nm, then foundries started all this marketing nonsense. 16nm and 14nm are 'hybrid' nodes and 7nm is pure marketing label for a node has zero resemblance to a 7nm node.

From 45nm to 20nm density increased by 5x whereas cost increased by 2x.

From 14nm to 7nm density increased only by 2x whereas cost increased by 2x.

Ok, well look at it at by %

7nm is 50% of 14

So a 7nm node would give four times higher density than 14nm node... and that is not the case with Glofo/TSMC because 7nm is a marketing label.

 

[Digital Viper-X-]

You are mixing marketing node names with real node names. Node naming was correct up to 20nm, then foundries started all this marketing nonsense. 16nm and 14nm are 'hybrid' nodes and 7nm is pure marketing label for a node has zero resemblance to a 7nm node.

From 45nm to 20nm density increased by 5x whereas cost increased by 2x.

From 14nm to 7nm density increased only by 2x whereas cost increased by 2x.



So a 7nm node would give four times higher density than 14nm node... and that is not the case with Glofo/TSMC because 7nm is a marketing label.

• 7nm (9.2nm standard node) (based on the doc above)
14/16 is actually 18nm @ TSMC and 17nm @ GF

So either way you slice it, the 7nm (9.2) is still roughly 50% of the previous 14nm(17/18)

 

[juanrga]

• 7nm (9.2nm standard node) (based on the doc above)
14/16 is actually 18nm @ TSMC and 17nm @ GF

So either way you slice it, the 7nm (9.2) is still roughly 50% of the previous 14nm(17/18)

No. A true 9.2nm node would be 3.4x more dense than a true 17nm node. The TSMC node is only 2.4x more dense. Your 9.2nm is in reality a 11nm node.

(17nm / 11nm )² = 2.4

 

[Digital Viper-X-]

No. A true 9.2nm node would be 3.4x more dense than a true 17nm node. The TSMC node is only 2.4x more dense. Your 9.2nm is in reality a 11nm node.

(17nm / 11nm )² = 2.4

I'm using the document for reference, I appreciate that you're going by the densities, but where is your source for that?

 

[juanrga]

I'm using the document for reference, I appreciate that you're going by the densities, but where is your source for that?

I am using the official info disclosed by the foundries, such as the SRAM cell densities:

Glofo 14LPP HD SRAM = 0.064 um²
TSMC 7HPC HD SRAM = 0.027 um²

You can find this kind of info in many places dealing with foundry tech. A good resume can be found in wikichip.

TSMC 7nm SRAM cells are only 2.4x more dense than Glofo 14nm. However, a true 7nm node would be 4x more dense. TSMC 7nm is a 11nm or 12nm node relabeled as "7nm" for marketing purposes.

Moreover I am using SRAM cell densities, which gives advantage to non-Intel foundries. The situation is worse when one considers logic densities. Intel 10nm is more dense than TSMC 7nm

 

[Digital Viper-X-]

I am using the official info disclosed by the foundries, such as the SRAM cell densities:

Glofo 14LPP HD SRAM = 0.064 um²
TSMC 7HPC HD SRAM = 0.027 um²

You can find this kind of info in many places dealing with foundry tech. A good resume can be found in wikichip.

TSMC 7nm SRAM cells are only 2.4x more dense than Glofo 14nm. However, a true 7nm node would be 4x more dense. TSMC 7nm is a 11nm or 12nm node relabeled as "7nm" for marketing purposes.

Moreover I am using SRAM cell densities, which gives advantage to non-Intel foundries. The situation is worse when one considers logic densities. Intel 10nm is more dense than TSMC 7nm

View attachment 135494

can you post some links please ?

 

[Gideon]

can you post some links please ?

It's a Intel marketing slide.

 

[DuronBurgerMan]

Intel 10nm is a pubic hair better than TSMC 7nm.

Intel 10nm MTx/mm2 (millions of transistors per mm squared) density of 106.1
TSMC 7nm MTx/mm2 density of 96.49

This compared to GloFo's 14LPP MTx/mm2 density of 32.5 and Intel's 14nm MTx/mm2 density of 43.5

The difference is very academic, though. For all intents and purposes, they are roughly equivalent.

 

[Digital Viper-X-]

Intel 10nm is a pubic hair better than TSMC 7nm.

Intel 10nm MTx/mm2 (millions of transistors per mm squared) density of 106.1
TSMC 7nm MTx/mm2 density of 96.49

This compared to GloFo's 14LPP MTx/mm2 density of 32.5 and Intel's 14nm MTx/mm2 density of 43.5

The difference is very academic, though. For all intents and purposes, they are roughly equivalent.

So based on this, TSCMs 7nm is 3x the density of GF 14nm

 

[juanrga]

can you post some links please ?

https://en.wikichip.org/wiki/7_nm_lithography_process

Use the menu on the right to see 14nm and 10nm.

It's a Intel marketing slide.

No. It is a figure prepared by David Schor using the official values for each node.

Intel 10nm is a pubic hair better than TSMC 7nm.

Intel 10nm MTx/mm2 (millions of transistors per mm squared) density of 106.1
TSMC 7nm MTx/mm2 density of 96.49

This compared to GloFo's 14LPP MTx/mm2 density of 32.5 and Intel's 14nm MTx/mm2 density of 43.5

The difference is very academic, though. For all intents and purposes, they are roughly equivalent.

100.76 (not 106.1) is the density of HD libraries of Intel 10nm. HP and UHP libraries have smaller density: 80.61 and 67.18 repsectively.

96.49 is the density of the TSMC 7FF node aimed at mobile applications. AMD is using 7HPC node for Zen and GPUs. The density of the 7HPC node is ~66 MTr/mm².

 

[bobzdar]

So based on this, TSCMs 7nm is 3x the density of GF 14nm

3x density at 2x the cost means 1/3 lower price per transistor. Depends on yields of course, so probably equivalent at the start and then getting cheaper as yields improve. Even at the start the cost difference for an 8 core will be driven more by packaging 2 die together (which they already have a ton of experience with) than silicon cost.

 

[DuronBurgerMan]

100.76 (not 106.1) is the density of HD libraries of Intel 10nm.

I know where you got the 100.76 value. Intel reported it some time ago. Semiwiki, however, has it at 103 - saying that Intel is underestimating density. Many other sources cite the 106.1 value I am using (feel free to look it up, it's everywhere). I used the one that would paint Intel in the best light. I will accept any of these three values, however.

96.49 is the density of the TSMC 7FF node aimed at mobile applications. AMD is using 7HPC node for Zen and GPUs. The density of the 7HPC node is ~66 MTr/mm².

I've been reading a lot of material on this. I could not find any official numbers for 7HPC - even at the usual sites - vs 7FF. If you are right, then it's more like a 10-12nm process, in which case Intel's 14++ is almost as good. But I did not get the impression there are a great density difference between the two.

If you have official material on this, link me.

 

[Pieter3dnow]

The "official" CES at AMD video from the bring up crew

 

[tangoseal]

I understand that 7nm is new, but why are we saying that a 7nm chiplet is going to cost more than a 14nm single die?
Assuming yield rates are the same, there should be more chips per wafer on 7nm than on 14nm, given that 7nm should cost more per wafer, what is the actual cost? People are just throwing statements out, without adding anything to support them. Give me some numbers.

Yeah like double the yield theoretically. Were talking a 50% reduction in size.