Haswell & Steamroller Power Consumption

[ctsoth]

My Question is this:

Even if Steamroller manages to be competitive with Haswell, as far as raw performance goes, is there any way it can be as power efficient? These days it's all about performance per watt, and looking at all of Haswell's power optimizations, I am left wondering if Steamroller can compete... I know there is a similar thread, but I feel that thread is performance oriented, my main concern is performace -> power usage.

 

[nehalem256]

Given that Steamroller will be built on GF 28nm process and Haswell will be on a 22nm process I would say it is decidedly unlikely.

 

[ShintaiDK]

It wont be competitive to say it mildly, it will be far from as power efficient due to the ondie VRM as well and SOix states.

I assume you talk about FM2 Steamroller.

For ultrabooks, Haswell comes in a single package design. Unlike the 2 chip design by LGA1156/1155/1150/FM1/FM2.

 

[CHADBOGA]

My Question is this:

Even if Steamroller manages to be competitive with Haswell, as far as raw performance goes, is there any way it can be as power efficient? These days it's all about performance per watt, and looking at all of Haswell's power optimizations, I am left wondering if Steamroller can compete... I know there is a similar thread, but I feel that thread is performance oriented, my main concern is performace -> power usage.

Whilst the discrepancy of energy efficiency will be very large, I doubt very much there is going to be anything like performance competitiveness.

 

[Idontcare]

...is there any way it can be as power efficient?

Very unlikely.

28nm won't be SOI, so it won't even have that going for it.

I will be surprised if the power efficiency of GloFo's 28nm bulk-Si is any better than their 32nm SOI process to be honest.

If 28nm bulk is more power efficient than 32nm SOI then it really casts into doubt the merits of pursuing SOI for any future nodes.

AMD's move to 28nm bulk is likely to net AMD a benefit of lower cost (smaller die) and maybe a benefit of slightly faster clockspeeds if the Idrives are higher (should be).

But power usage is not likely to benefit in this "half-step forward, half-step back" dance that AMD is doing by going from 32nm SOI to 28nm bulk-Si.

 

[IntelUser2000]

If 28nm bulk is more power efficient than 32nm SOI then it really casts into doubt the merits of pursuing SOI for any future nodes.

It could well be lower power than their 32nm process though. The reason I suggest this idea is because the focus is likely changing from pure current drive focused transistors to more power optimized, but somewhat lower clock speed transistors. Process may be one reason why Steamroller is said to reduce emphasis on clock speeds(and may actually end up less).

That also seems to be the approach for Intel with their 22nm process. Based on their presentations, it doesn't seem to be all out performance, but performance/watt.

 

[Idontcare]

It could well be lower power than their 32nm process though. The reason I suggest this idea is because the focus is likely changing from pure current drive focused transistors to more power optimized, but somewhat lower clock speed transistors. Process may be one reason why Steamroller is said to reduce emphasis on clock speeds(and may actually end up less).

That also seems to be the approach for Intel with their 22nm process. Based on their presentations, it doesn't seem to be all out performance, but performance/watt.

Not disagreeing, but if power reduction was the focus of 28nm then why drop SOI when they developed it?

You've got to admit that if SOI provided so little benefit to 32nm that going back to bulk for 28nm resulted in substantial power savings over 32nm SOI then it begs the question of why bother going back to SOI for 22/20nm?

 

[NostaSeronx]

Given that Steamroller will be built on GF 28nm process and Haswell will be on a 22nm process I would say it is decidedly unlikely.

Globalfoundries' & Samsung's process is ahead of ITRS numbers so GloFo's/Sam's 28-nm will be closer to 22-nm than Intel's.

 

[RU482]

Globalfoundries' & Samsung's process is ahead of ITRS numbers so GloFo's/Sam's 28-nm will be closer to 22-nm than Intel's.

O'RLY

 

[NostaSeronx]

O'RLY

Give or take a significant amount of picometers...

 

[ShintaiDK]

Globalfoundries' & Samsung's process is ahead of ITRS numbers so GloFo's/Sam's 28-nm will be closer to 22-nm than Intel's.

Its a joke post...right?

 

[NostaSeronx]

Its a joke post...right?

Nope. 28-nm from GloFo and Samsung has a denser node, denser metal layers, and denser wires.

The best guesses I have: (some are obvious)
AMD Jaguar => TSMC 28-nm HPM(2nd gen libraries)
AMD Steamroller => Samsung & GlobalFoundries 28-nm LPH(2nd-3rd gen libraries)/(It can be HPP as well which uses 2nd gen libraries but that appears to be GloFo only, 28-nm SLP/28-nm LPH are Samsung/Glofo sync'ed and why waive out of 28-nm exclusivity then go on a foundry exclusive process. ext1)
Intel Haswell => 2nd gen 22-nm

ext1: 28-nm LPH has a lot more options than 28-nm HPP by the way especially the fact than you can have high speed SRAM with LPH where you can't with HPP. Which makes LPH more solid as a choice.

 

[LoneNinja]

Nope. 28-nm from GloFo and Samsung has a denser node, denser metal layers, and denser wires.

If you've got a link explaining this I'd like to see it

 

[Idontcare]

Nope. 28-nm from GloFo and Samsung has a denser node, denser metal layers, and denser wires.

Good, so they'll have no excuses for not beating Intel then. What's not to like?

 

[NostaSeronx]

Good, so they'll have no excuses for not beating Intel then. What's not to like?

Two real issues can happen:
Maturity of the process.
Steamroller's architecture takes a turn for the worse.

The more information coming out for SR points it is going for double the specified clock from Bulldozer. Bulldozer was planned to have nominal performance around 3.5 GHz. Steamroller is looking at 7 GHz to have nominal performance.

If you've got a link explaining this I'd like to see it

22-nm/28-nm no longer specifies the actual gate length and there is various reasons why this occurs. Intel's Lfin is ~26-nm where GloFo/Samsung is going for ~25-nm Lgate. Samsung wanted 28-nm to be a full node drop from 32-nm, they didn't quite make it but they got pretty close to the 22-nm gate length. So, where 22-nm from Intel is like a half-node leap GloFo/Samsung 28-nm is like a three-quarter-node jump from 32-nm.

 

[mikk]

It means AMD has (or will have) the superior process technology over Intel. Nice. There is no excuse anymore for the CPU itself. Good news for AMD, the better process should help them.

 

[NTMBK]

Steamroller is looking at 7 GHz to have nominal performance.

7GHz?! As in, ship processors that operate at 7GHz?

 

[CHADBOGA]

It means AMD has (or will have) the superior process technology over Intel. Nice. There is no excuse anymore for the CPU itself. Good news for AMD, the better process should help them.

LOL. Someone's having a lend of you and you are falling for it.

 

[Abwx]

28nm won't be SOI,

Why not..??...

Globalfoundries is to manufacture FD-SOI chips for STMicroelectronics and ST-Ericsson at both 28nm and 20nm process generations which will be made avaialble to anyone wanting to use them.

http://www.electronicsweekly.com/Ar...up-28nm-20nm-fd-soi-process-to-all-comers.htm

Geneva, June 11, 2012 - STMicroelectronics (NYSE: STM), a global semiconductor leader serving customers across the spectrum of electronics applications, announces that GLOBALFOUNDRIES, a worldwide leader in the technology roadmap race, has agreed to manufacture devices for ST using ST’s proprietary Fully Depleted Silicon-on-Insulator (FD-SOI) technology in both the 28nm and 20nm nodes

http://www.st.com/internet/com/press_release/c2680.jsp

 

[Denithor]

7GHz?! As in, ship processors that operate at 7GHz?

Of course.

And they'll be happy to charge you $1000+ for an LN2 cooler to mount on it to keep things "Cool 'n Quiet" as their tagline reads. Refills are extra. (The new AMD business model - continuously charge customers for the gear they've already bought. Oh, wait - they already do that with power bills on Bulldozer.)

 

[Idontcare]

Why not..??...



http://www.electronicsweekly.com/Ar...up-28nm-20nm-fd-soi-process-to-all-comers.htm



http://www.st.com/internet/com/press_release/c2680.jsp

Don't be obtuse, I'm obviously referring to the 28nm bulk-Si process that AMD is using at GloFo for production of Steamroller.

But yes, the question is why did GLoFo and AMD drop SOI for 28nm if it is so crucial to delivering low-power? My thoughts on that have already been expressed above.

Just because ST announed plans to do something someday with SOI on something they will call 28nm doesn't mean anything in this convo, completely unrelated to the topic I was discussing.

 

[Abwx]

Don't be obtuse, I'm obviously referring to the 28nm bulk-Si process that AMD is using at GloFo for production of Steamroller.

But yes, the question is why did GLoFo and AMD drop SOI for 28nm if it is so crucial to delivering low-power? My thoughts on that have already been expressed above.

Just because ST announed plans to do something someday with SOI on something they will call 28nm doesn't mean anything in this convo, completely unrelated to the topic I was discussing.

True that early information point to use of bulk silicon for next CPUs ,
but why would GF use a more expensive process for lower value ICs.?..

According to the infos , ST FD-SOI is better than TSMS s bulk
so why should GloFo s own bulk be better than the former..?.

 

[meloz]

These days it's all about performance per watt, and looking at all of Haswell's power optimizations, I am left wondering if Steamroller can compete...

I had a somewhat similar question and the general conclusion was no, AMD cannot compete anymore.

AMD might still have some niches Intel will leave them where AMD can survive by being better than Intel at performance / $, but as far as performance per watt is concerned the gap betwen Intel and AMD is only going to increase with each new product cycle.

In many ways with Haswell Intel are merely laying the foundation for what they are going to do to improve performance/watt in future. Broadwell should be just incredible.

 

[Idontcare]

True that early information point to use of bulk silicon for next CPUs ,
but why would GF use a more expensive process for lower value ICs.?..

According to the infos , ST FD-SOI is better than TSMS s bulk
so why should GloFo s own bulk be better than the former..?.

You lost me there.

This isn't a question of whether SOI reduces power usage. Presumably it does.

The question is if you take power usage for 32nm with SOI versus 28nm without SOI.

How much benefit did SOI give 32nm in terms of reducing the power of 32nm? Because that benefit, however big or small, won't be there on the 28nm bulk-Si that AMD's steamroller chips will be manufactured with.

So if SOI was immensely beneficial in reducing power usage on 32nm, then the lack of SOI at 28nm will probably mean that 28nm bulk-Si will have comparable leakage to that of 32nm w/SOI. Meaning steamroller will likely have comparable leakage to that of bulldozer, and that will answer the question of that is the thread title.

However, if SOI provided little benefit in reducing power usage of 32nm, then the absence of SOI from AMD's 28nm CPU's won't be detrimental and AMD's 28nm CPU's can be expected to have a significant reduction in power usage.

They either need it or they don't. If they need it then it is going to be a problem that it isn't there on 28nm. If they don't need it on 28nm then they probably didn't need it on 32nm and likewise won't really need it for 22/20nm either.

(Steamroller is not a low-value IC)

Personally I believe SOI provided AMD great benefit for lowering bulldozer's power usage, and I think steamroller is going to see no improvement in power usage despite the half-node shrink because it loses the benefit that SOI provides.

Steamroller will be cheaper to manufacture, shrinking it and dropping SOI are all cost-reduction moves, but they are not performance enhancing moves when you are TDP and clockspeed limited to start with.

 

[NostaSeronx]

SOI doesn't lower power consumption it lowers leakage.

AMD isn't going on SLP.