[AMD] K12 will be on 28nm

[Fjodor2001]

The fact that it is ARM does not mean it is more or less efficient for a specific workload.

Edit: And add to the fact it will be competing with other ARM options on better nodes and x86 on better nodes.

The question is: Will the lower power consumption and higher performance be enough to justify 14 nm instead of 28 nm for server CPUs?

The last few years we've seen that later nodes does no longer bring much higher performance. The power consumption will likely be a bit lower, but is it enough to justify the much higher cost? I'm not so sure. I bet it's decisions like that which makes AMD provide such an evasive answer.

 

[mrmt]

The question is: Will the lower power consumption and higher performance be enough to justify 14 nm instead of 28 nm for server CPUs?

Yes, it will. That's why IBM, Intel and Oracle are focusing in shrinking their processors to the latest nodes.

The last few years we've seen that later nodes does no longer bring much higher performance. The power consumption will likely be a bit lower, but is it enough to justify the much higher cost? I'm not so sure. I bet it's decisions like that which makes AMD provide such an evasive answer.

That's not correct for the server market. Core count went through the roof in the last few years, and there's a lot of performance/watt improvements. Just compare 8C SNB-EP servers with 18C-HSW-EP. Just because desktops are stagnant doesn't mean servers will follow the same suit.

 

[Fjodor2001]

Does ARM have some kind of magical fairy dust?

If ARM ISA making more power efficient uArch designs possible, then yes.

 

[RampantAndroid]

The question is: Will the lower power consumption and higher performance be enough to justify 14 nm instead of 28 nm for server CPUs?

The last few years we've seen that later nodes does no longer bring much higher performance. The power consumption will likely be a bit lower, but is it enough to justify the much higher cost? I'm not so sure. I bet it's decisions like that which makes AMD provide such an evasive answer.

...in server farms, I think any amount of power savings (which in turn means less heat and less cooling needed) would be a VERY good thing...

 

[Fjodor2001]

Yes, it will. That's why IBM, Intel and Oracle are focusing in shrinking their processors to the latest nodes.

Maybe they are targeting a different segment of the server market.

 

[Fjodor2001]

...in server farms, I think any amount of power savings (which in turn means less heat and less cooling needed) would be a VERY good thing...

Of course, but at what cost. Are you willing to pay 2x as much for 30% power savings? 4x as much... ?

 

[mrmt]

Maybe they are targeting a different segment of the server market.

Not Intel. Intel 14nm Atom server chips will eat AMD 28nm chips for breakfast in any metric you can think off.

 

[RampantAndroid]

Of course, but at what cost. Are you willing to pay 2x as much for 30% power savings? 4x as much... ?

Isn't the power savings going from Intel's 22nm to 14nm, from published documents, something like 30% right there?

I've been around labs that had a fire alarm go off. You hear all of the AC shut off and have a split second to cover your ears before the alarms sound. Best fireman response time, your lab is down for 15-30 minutes for a false alarm. In that time, the lab temps have risen well over 10C. In one such instance, the systems overheated. The servers themselves shut down fine; the hard drives seized. Now, if you tell me that I can make the lab heat 30% slower in that scenario...yeah, the cost probably would be worth it.

 

[Fjodor2001]

Not Intel. Intel 14nm Atom server chips will eat AMD 28nm chips for breakfast in any metric you can think off.

Including perf/cost?

 

[mrmt]

Including perf/cost?

Nobody buys a server based on perf/costs, but based on TCO for a given scenario. Given that the costs of powering and cooling a processor in datacenter are bigger than the costs of acquisition of a server processors, even when considering the huge price premium over desktop counterparts, perf/watt is the absolute metric for servers. This is why Opteron was wiped out on servers but not on desktops.

 

[Phynaz]

They never stated it but that doesnt mean that these are better than at 22nm,

Node shrinks are about economics - number of die per wafer. Do you have any evidence that Intel's 14nm process is any worse than their 22nm at the same point in it's lifetime?

Do you have any evidence that somehow AMD can compete with Intel when AMD is saddled with a 400% larger die for the same transistor count?

Factual evidence please, not your opinion. Please provide links. Otherwise you're just trolling.

 

[Abwx]

Node shrinks are about economics - number of die per wafer. Do you have any evidence that Intel's 14nm process is any worse than their 22nm at the same point in it's lifetime?

I dont talk of economics, that s another matter, but of electrical caracteristics.

Do you have any evidence that somehow AMD can compete with Intel when AMD is saddled with a 400% larger die for the same transistor count?

I did say that Intel s current 14nm is of no threat, did you notice the exact terms or should i bold them..?.

Factual evidence please, not your opinion. Please provide links. Otherwise you're just trolling.

Evidence is the significantly higher voltage than BTrail, as much voltage at 2.0-2.4 than a 4770K at 3.5, some people here can do easily the maths with the few infos available on this subject, these are real evidences, you know that i m a fan of thoses and not of urban legends...

 

[krumme]

Kumar is not a good orator, his discourse is way above confused (albeit I never saw him give wrong information in any of AMD's Q&A). If you check the transcript you'll see that he spoke "I think" 21 times.

So much energy you use on that. Give yourself some rest.

 

[mrmt]

So much energy you use on that. Give yourself some rest.

Ctrl +f isn't much energy spending.

 

[Phynaz]

I dont talk of economics, that s another matter, but of electrical caracteristics.



I did say that Intel s current 14nm is of no threat, did you notice the exact terms or should i bold them..?.



Evidence is the significantly higher voltage than BTrail, as much voltage at 2.0-2.4 than a 4770K at 3.5, some people here can do easily the maths with the few infos available on this subject, these are real evidences, you know that i m a fan of thoses and not of urban legends...

Ok, no facts. Got it.

 

[PPB]

Ok, no facts. Got it.

So then if a product in a smaller node needs higher voltage to operate at half the frequencies than the other one on a bigger node, when historically the trend shows it is pretty much the other way (the smaller node needing less Vcore to operate at the same frequencies, thus giving you perf/watt gains) is no proof for you to at least think something strange is happening there, then what can I say?

 

[john5220]

can we expect a K12 AMD to beat a 2500K i5 sandy bridge?

if AMD does this its a win. 2500K will remain a beast for the next 5 years.

 

[chrisjames61]

Isn't the power savings going from Intel's 22nm to 14nm, from published documents, something like 30% right there?

I've been around labs that had a fire alarm go off. You hear all of the AC shut off and have a split second to cover your ears before the alarms sound. Best fireman response time, your lab is down for 15-30 minutes for a false alarm. In that time, the lab temps have risen well over 10C. In one such instance, the systems overheated. The servers themselves shut down fine; the hard drives seized. Now, if you tell me that I can make the lab heat 30% slower in that scenario...yeah, the cost probably would be worth it.

There is a lot more making heat than just the cpu's when you have a dense server farm. A 30% cpu power savings isn't going to be linear with a decline in heat generated either.

 

[tential]

can we expect a K12 AMD to beat a 2500K i5 sandy bridge?

if AMD does this its a win. 2500K will remain a beast for the next 5 years.

K12 is believed to be low cost/scope constrained project. Says so right in the OP...
So why would you think to compare it to a 2500k?

 

[ShintaiDK]

can we expect a K12 AMD to beat a 2500K i5 sandy bridge?

if AMD does this its a win. 2500K will remain a beast for the next 5 years.

No.

 

[ShintaiDK]

So then if a product in a smaller node needs higher voltage to operate at half the frequencies than the other one on a bigger node, when historically the trend shows it is pretty much the other way (the smaller node needing less Vcore to operate at the same frequencies, thus giving you perf/watt gains) is no proof for you to at least think something strange is happening there, then what can I say?

Lets stay at facts and not some jibberish from Abwx. Its about the same thing as listening to Seronx.

Also dont compare 2 different designs with completely different design goals.

 

[RampantAndroid]

There is a lot more making heat than just the cpu's when you have a dense server farm. A 30% cpu power savings isn't going to be linear with a decline in heat generated either.

I absolutely get that and was trying to make a point - that if you're able to reduce power consumption, it's more important than a slightly higher cost on the CPU itself more often than not (the point someone else made better than I did.)

 

[john5220]

K12 is believed to be low cost/scope constrained project. Says so right in the OP...
So why would you think to compare it to a 2500k?

I was hopeful AMD would once again be able to compete with intel.

Maybe their plans can change along the way, I guess by 2016 and 2017 we will know how good their K12 really is.

Reaching 2500K performance is a giant leap for any company. I would think by 2016/2017 AMD would atleast be able to pull off those performance considering a 2500K will be REALLY OLD by then.

 

[tential]

I was hopeful AMD would once again be able to compete with intel.

Maybe their plans can change along the way, I guess by 2016 and 2017 we will know how good their K12 really is.

Reaching 2500K performance is a giant leap for any company. I would think by 2016/2017 AMD would atleast be able to pull off those performance considering a 2500K will be REALLY OLD by then.

Why would they chase the 2500k? You need to ask yourself why that's relevant before even asking it. In 2016, Sandybridge is 5 years old.

Sandybridge, Ivybridge, Haswell, Devil's Canyon, Broadwell, Skylake. That's 6 generations of CPUs that people will have. Why would catching Sandybridge mean anything to AMD? That would mean there are 6 years of products all better than their NEW product. It's a marketing disaster. Why would anyone even purchase a chip that's as fast as a chip from 5 years ago? That doesn't even make sense...

Again, I ask you to read the threads rather than just posing questions. You'll see that LowCost/LowPower solutions are what AMD is focusing on moving forward. Tablets, Smartphones, Ultrabooks, 2in1s, that's the future. That's where money is. AMD is moving into ARM designs hoping to use it's GPU to make a viable chip in Zen.

The High End CPU market doesn't fit into AMD's portfolio at this time.

This isn't a game. It's real life, and people don't care about winning some arbitrary prize if it means losing out on PROFITS. So AMD is trying to reposition their efforts and become more focused on a few markets. Intel still has to hit every single market segment, AMD is hoping to specialize in a few of those and maybe come out ahead (or rather just keep it's head afloat).

Sure someone else can clarify this post better.

 

[Phynaz]

So then if a product in a smaller node needs higher voltage to operate at half the frequencies than the other one on a bigger node, when historically the trend shows it is pretty much the other way (the smaller node needing less Vcore to operate at the same frequencies, thus giving you perf/watt gains) is no proof for you to at least think something strange is happening there, then what can I say?

First you can prove that Intel's 14nm process requires a higher voltage for the same design. Your buddy hasn't proven a thing. Operating voltage is a design parameter and a higher operating voltage does not indicate inferiority.

Your logical fallacy is that perf/watt gains can only occur by reducing voltage. what if I designed a circuit that was more efficient but yet required a higher voltage to operate than a less efficient design?