Fudzilla: New AMD Zen APU boasts up to 16 cores (plus Greenland GPU with HBM)

[erunion]

If this got posted, I didn't see it.
I had expected 8 cores and 4 cores, but not 2 cores. Interesting. AMD is rebooting its PC lineup and intends to continue competing with Intel.


Desktop:

Mobile:

 

[DrMrLordX]

I hadn't seen (or simply didn't notice) the mobility roadmap. I think I saw the desktop one over on XS.

What's interesting to me is that they have Amur on there, but not Nolan. Hmmmm!

Also, they've got Beema in their desktop lineup now, 'bout time AMD! Please save us from the crappy AiOs with slightly less-crappy AiOs, thank you very much. Though one wishes for higher clocks for the "desktop" version of Beema (essentially, Carrizo-L) than what could be found in the few laptops that used it (or tablets like the BungBungame Photon 2).

 

[ShintaiDK]

But this is not a constant. Newer process will get better yields and get cheaper with time. Samsung/GF process is very expensive now, but will be cheaper next year.

Its not about yields. But about double patterning etc. This is why everyone talks about EUV, since it will again lower the cost without massive design costs.

 

[ShintaiDK]

+1 as usual
Nothing AMD does or have in its pipeline is ever good for some

AMDs constant dropping marketshare tells the same story, despite the constant promises of the next turnaround product.

 

[AtenRa]

AMDs constant dropping marketshare tells the same story, despite the constant promises of the next turnaround product.

Q1 2015 Desktop decrease sales were due to market conditions and both AMD and Intel lost Revenue and profit.

AMD lost less than 70M Desktop CPU Revenue Q to Q while Intel lost 1.4B Desktop CPU Revenue Q to Q.

Edit: Less than 70M loss from Desktop Revenue

 

[Gikaseixas]

AMDs constant dropping marketshare tells the same story, despite the constant promises of the next turnaround product.

Not surprising when the competition gives chips for free and AMD's marketing is as good as it is. Newbies mostly know Intel and Nvidia.
AMD actually makes good products and it's one of the main tech innovators.

 

[CHADBOGA]

Q1 2015 Desktop decrease sales were due to market conditions and both AMD and Intel lost Revenue and profit.

AMD lost ~150M Desktop CPU Revenue while Intel lost 1.4B Desktop CPU Revenue.

None of that changes the fact that AMD lost more desktop share to Intel.

 

[AtenRa]

None of that changes the fact that AMD lost more desktop share to Intel.

Where did you see that ???

edit: We are talking about Q1 2015

 

[Fjodor2001]

Its not about yields. But about double patterning etc. This is why everyone talks about EUV, since it will again lower the cost without massive design costs.

So you don't think Samsung/GF 14 nm will be cheaper per mm^2 die area next year!?

Also, how is EUV relevant to this? Will that stop Samsung/GF 14 nm from becoming cheaper in the next few years?

Regardless, you see EUV coming in actual high volume products when... 2020?

 

[carop]

But this is not a constant. Newer process will get better yields and get cheaper with time. Samsung/GF process is very expensive now, but will be cheaper next year.

The costs are going up because of multiple patterning. Lars Liebmann at IBM has a cost chart for the various techniques:

The foundries are using LELE (Litho-Etch-Litho-Etch) whereas Intel is using SADP (Self Aligned Double Patterning). Compared to SE (Single Exposure), LELE is 2.5x and SADP is 3x more expensive. So, the costs are going up for Intel as well. It would, however, be foolish to expect Intel to admit that their costs are going up.

At its 10nm node Intel will most probably have at least one critical layer which will need SAQP (Self Aligned Quadruple Patterning). According to the cost chart, SAQP is 4.5x more expensive than SE (Single Exposure). Mark Bohr at Intel spins this follows:

At 10nm [we’re] running 50% faster in steps per day through the fab, increasing the rate of wafer movement. I think that will keep 10nm on track…[We’re] getting very good area scaling and cost per transistor reduction in 10nm. In our development fab we’ve sped up the move of wafers to offset the increased number of mask steps.

http://www.eetimes.com/document.asp?doc_id=1325734&page_number=2

Good luck explaining to the Intel crowd that throwing more equipment at the problem does not shorten the cycle time in a wafer process that is serial. Yes, serial. It will still take (# of steps) x (days per step) to get a wafer through. If there are 50% more steps for node 10nm then the cycle time will be 50% longer than node 14nm and the cost per wafer will be 50% higher because of this.

 

[Dresdenboy]

AMDs constant dropping marketshare tells the same story, despite the constant promises of the next turnaround product.

Did they promise a turnaround product? And are you also talking about embedded and console SoCs? Biz is biz. As long as there are humans involved, there won't be any determinism due to some specs.

 

[Arachnotronic]

The foundries are using LELE (Litho-Etch-Litho-Etch) whereas Intel is using SADP (Self Aligned Double Patterning). Compared to SE (Single Exposure), LELE is 2.5x and SADP is 3x more expensive. So, the costs are going up for Intel as well. It would, however, be foolish to expect Intel to admit that their costs are going up.

Cost per wafer might be going up, but the idea is that density improvements offset that cost per wafer to drive a lower cost per transistor.

Intel has never said that wafer costs aren't going up; they have talked about this, and their expectations of cost/transistor improvement from density scaling, quite a bit.

 

[JDG1980]

1. No solid turbo tech means low freq is optimal

AMD has had decent turbo core tech since the original Bulldozer. See the Anandtech review for details. In fact, this is one of the few things it did right with BD.

 

[ShintaiDK]

Q1 2015 Desktop decrease sales were due to market conditions and both AMD and Intel lost Revenue and profit.

AMD lost less than 70M Desktop CPU Revenue Q to Q while Intel lost 1.4B Desktop CPU Revenue Q to Q.

Edit: Less than 70M loss from Desktop Revenue

How was the revenue YoY?

861M$ to 532M$?

Bad excuses is bad excuses.

 

[ShintaiDK]

So you don't think Samsung/GF 14 nm will be cheaper per mm^2 die area next year!?

Also, how is EUV relevant to this? Will that stop Samsung/GF 14 nm from becoming cheaper in the next few years?

Regardless, you see EUV coming in actual high volume products when... 2020?

Compared to 28nm? Absolutely not. 14nm will cost over twice of 28nm per mm2. Even Samsung says so. Even a 1B transistor 14nm design goes up in manufactoring cost against a 1B transistor 28nm design. Plus the design cost is 4x more.

While yield may improve and electricals. The cost per each individual chip, working or not, goes up.

 

[AtenRa]

How was the revenue YoY?

861M$ to 532M$?

Bad excuses is bad excuses.

The Revenue you quoting is for both CPUs and GPUs. But yes Desktop sales and volumes were decreased.

Lets see what happen to Intel,


Intel Q1 2014 Client Computing Group = 8097M
Intel Q1 2015 Client Computing Group = 7420M

677M decrease of Revenue primarily from Desktop decreased sales.

Notebook platform volumes increased 3% from Q1 2014 to Q1 2015
Desktop platform volumes decreased 16% from Q1 2014 to Q1 2015

So again, in Q1 2015 both AMD and Intel lost a lot of Desktop revenue but they both Increased Mobile (laptops) sales and Revenue.

 

[ShintaiDK]

The Revenue you quoting is for both CPUs and GPUs. But yes Desktop sales and volumes were decreased.

Lets see what happen to Intel,


Intel Q1 2014 Client Computing Group = 8097M
Intel Q1 2015 Client Computing Group = 7420M

677M decrease of Revenue primarily from Desktop decreased sales.

Notebook platform volumes increased 3% from Q1 2014 to Q1 2015
Desktop platform volumes decreased 16% from Q1 2014 to Q1 2015

So again, in Q1 2015 both AMD and Intel lost a lot of Desktop revenue but they both Increased Mobile (laptops) sales and Revenue.

AMDs number contains servers, discrete GPUs, mobile and desktop.

AMDs PC sales are now only 61.7% of what it was in Q1 2014.
Intels desktop+mobile sales are now 91.7% of what it was in Q1 2014.
Intels server sales are now 119.2% of what it was on in Q1 2014.

And if you combine Intels 2, you get a flat one.

As a complete company, Intels revenue is flat vs Q1 2014.
As a complete company, AMD went from 1.4B to 1.03B. Or only 73.5% of what it had in Q1 2014.

You still dont see the issue?

 

[Gikaseixas]

The cost per each individual chip, working or not, goes up.

So what? They can always charge more for a better, faster, newer product. A price bump should be ok to offset manufacturing costs as long as it provides decent jump over theexisting line up

 

[ShintaiDK]

So what? They can always charge more for a better, faster, newer product. A price bump should be ok to offset manufacturing costs as long as it provides decent jump over theexisting line up

Absolutely. If they significant improve their product vs the competition they can raise the price to offset the increase cost. If the competition does the same, then its just status quo with higher cost structure.

 

[Erenhardt]

What any of this stock holder discussion have anything to do with the topic is beyond me.
I hoped for some post with the people with an actual knowledge, yet here I come again to read a bunch of bs from a industry messiah with his nebulous crystal ball.

 

[Gikaseixas]

Absolutely. If they significant improve their product vs the competition they can raise the price to offset the increase cost. If the competition does the same, then its just status quo with higher cost structure.

You shouldn't care so much with AMD's financials since you hate them to the point that you set up camp in every AMD related thread just to spread your failed forecasts and future visions.

Zen could bring back the good old battles between Intel and AMD, people just have to accept that it is a possibility.

 

[imported_ats]

At its 10nm node Intel will most probably have at least one critical layer which will need SAQP (Self Aligned Quadruple Patterning). According to the cost chart, SAQP is 4.5x more expensive than SE (Single Exposure). Mark Bohr at Intel spins this follows:

If Intel is going to require SAQP, then so will everyone else. Also the charts numbers for LELE vs SADP are open to vigorous debate. For instance, some would argue that the significantly reduced defect density and layout complexity of SADP makes it significantly more attractive than LELE. In fact, the majority of cost focused processes bypassed LELE and went directly to SADP.

Good luck explaining to the Intel crowd that throwing more equipment at the problem does not shorten the cycle time in a wafer process that is serial. Yes, serial. It will still take (# of steps) x (days per step) to get a wafer through. If there are 50% more steps for node 10nm then the cycle time will be 50% longer than node 14nm and the cost per wafer will be 50% higher because of this.

For the LITHOETCH sequence there are a number of ways to skin the cat. You have the number of steps, you have the time per step, and you have number of steppers(literally). In addition, you have all the steps outside of the actual LITHOETCH sequence that are also variable in number of steps and time per step. While Intel is likely using more steppers, its also likely that they are time and sequence optimizing everything in order to reduce total steps required and avg time per step.

Just to restate, there could be 100% more steps for 10nm but its latency could still be reduced and its throughput increased because there are a lot more variable besides just # of steps.

 

[Insert_Nickname]

Zen could bring back the good old battles between Intel and AMD, people just have to accept that it is a possibility.

I'd really like to see AMD make a comeback in the performance segment. More competition = better products for everyone. How anyone could think otherwise is beyond me.

I go were my $ buy me the most features/performance/efficiency.

 

[carop]

Cost per wafer might be going up, but the idea is that density improvements offset that cost per wafer to drive a lower cost per transistor.

Intel has never said that wafer costs aren't going up; they have talked about this, and their expectations of cost/transistor improvement from density scaling, quite a bit.

So, if they save money by running 10nm wafers through 50% faster, why not do this for 14nm as well? Oh, because then it would become apparent that 10nm is more expensive.

The cost of multiple patterning techniques according to ASML:

 

[Dresdenboy]

The cost of multiple patterning techniques according to ASML:

I didn't see any explicit fixed cost/invest in the slide. Is it hidden in "cost"?