Analysis: Haswell "Refresh" Is Not Broadwell

[Pandamonia]

Nobody cares about what your views are. We care about what's true. You can continue to believe false things, but it is a fact that Intel uses is a different process for high performance and low power designs. Unless you can show me a CPU that is faster than Haswell and Ivy Bridge, then I think they probably qualify as "Super High Performance."

Show me the technical details of the different 22nm processes from intel.

There is clear documentation of the 28nm processes TSMC uses and Glofo also.

22nm 3D transistor is clearly low power design since its used in ultrabooks and desktops with just binning being the only difference between the chips. Low power performance has increased significantly while temps and max clock speed have gotten worse from 32nm.

They are the same CPU's in either Dual or Quad core. There is no separate technology for their ULV chips and their 4770k or even server chips. They are just binned differently.

When 22nm Atom starts it will just be an even further tweaked 22nm process the same as with Ivy/Haswell.

Intel does not have a 22nm Process for high performance. Temps and clock speed in haswell clearly shows this.

 

[Blandge]

Show me the technical details of the different 22nm processes from intel.

There is clear documentation of the 28nm processes TSMC uses and Glofo also.

22nm 3D transistor is clearly low power design since its used in ultrabooks and desktops with just binning being the only difference between the chips.

They are the same CPU's in either Dual or Quad core. There is no separate technology for their ULV chips and their 4770k or even server chips. They are just binned differently.

When 22nm Atom starts it will just be an even further tweaked 22nm process the same as with Ivy/Haswell.

Intel does not have a 22nm Process for high performance. Temps and clock speed in haswell clearly shows this.

Read this presentation, specifically slide 31.

http://www.intel.com/content/www/us...licon-technology-leadership-presentation.html

 

[Pandamonia]

Read this presentation, specifically slide 31.

http://www.intel.com/content/www/us...licon-technology-leadership-presentation.html

Clearly shows ALL Desktop, Notebook, Server AND ULV Haswell chips are made on the same process. So the previous point is clearly invalid.

Just as i said that Atom will be further tweaked for low power SOC's and thats what that slide says.

22nm has clearly been designed for low power use compared to the previous 32nm process. This is haswells big selling point. It matches Ivybridge performance at lower system power consumption.

The evidence speaks louder than any intel marketing material.

 

[jpiniero]

If Intel are being greedy / cheap and just pushing back Broadwell to 2015 (or cancelling it altogether), why would they bother with the expensive and time consuming task of verifying all new chipsets for existing CPUs? Is SATA Express that important?

Because OEMs want new products.

 

[Blandge]

Clearly shows ALL Desktop, Notebook, Server AND ULV Haswell chips are made on the same process. So the previous point is clearly invalid.

Just as i said that Atom will be further tweaked for low power SOC's and thats what that slide says.

22nm has clearly been designed for low power use compared to the previous 32nm process. This is haswells big selling point. It matches Ivybridge performance at lower system power consumption.

The evidence speaks louder than any intel marketing material.

I never said that Intel manufactures Haswell ULV on its low power process.

And yes Intel's 22nm high performance CPU process is tweaked until it becomes their 22nm low power SoC process. That's two distinct processes. They tweaked the 486 for 20 years until it became Ivy Bridge too.

Yes 22nm was designed, in part, to extend Intel's product range to fit a lower power profile than 32nm. If that's the only point you are trying to make then yes, you are correct in your assessment that water is wet, but they still scale to the high end of Intel's product line approaching 4 GHz.

 

[Sweepr]

Just as i said that Atom will be further tweaked for low power SOC's and thats what that slide says.

The evidence speaks louder than any intel marketing material.

There are two processes:
Current Ivy Bridge and Haswell use the 22nm P1270. Silvermont Atom uses low-power 22nm P1271.

 

[meloz]

Because OEMs want new products.

This had not even crossed my mind.

This kind of BS is normal in GPU land, I am hoping against hope it does not come to CPUs.

 

[Pandamonia]

I never said that Intel manufactures Haswell ULV on its low power process.

And yes Intel's 22nm high performance CPU process is tweaked until it becomes their 22nm low power SoC process. That's two distinct processes. They tweaked the 486 for 20 years until it became Ivy Bridge too.

Yes 22nm was designed, in part, to extend Intel's product range to fit a lower power profile than 32nm. If that's the only point you are trying to make then yes, you are correct in your assessment that water is wet, but they still scale to the high end of Intel's product line approaching 4 GHz.

My point was that intels 22nm haswell, ivy and uses a process clearly designed for low power leakage. Broadwell 14nm will continue that process even further and this is why they may not launch desktop 14nm chips until the new arch gives some IPC gain to offset the even poorer overclocking.

Even intel cant launch a desktop chip worse than the previous year.

Intel does not have another process which can produce high performance at the expense of power usage which is the opposite of their current technology. My point was that GloFo and TSMC have processes which do both currently. Intel seems focused on only one and thats low power usage. People are saying 20nm TSMC is going middle of the road for both performance and power usage.

 

[RU482]

would it be safe to say that the "haswell refresh" will be the switchover point from DDR3 to DDR4?

 

[ShintaiDK]

would it be safe to say that the "haswell refresh" will be the switchover point from DDR3 to DDR4?

Thats not happening until Skylake I bet for LGA11xx.

Servers/workstations get DDR4 in 2014. Desktops in 2015.

 

[willomz]

Do we need DDR4 when high speed DDR3 has such little effect on overall performance?

 

[ShintaiDK]

Do we need DDR4 when high speed DDR3 has such little effect on overall performance?

DDR4 is about 3 things. IGP speed, lower power consumption and higher density.

 

[Idontcare]

Do we need DDR4 when high speed DDR3 has such little effect on overall performance?

That is a tricky question to answer because what we can't determine, as outsiders, is how much dependence on the memory has been "engineered out" of the existing crop of CPUs by way of larger and larger on-die caches combined with better and better prefetchers and prediction algos.

You might be compelled to respond along the lines of "well if the cache and prefetchers are doing such a bang-up job nowadays in insulating the CPU from the horrid performance of DDR3, then why not continue to depend on CPU designers to keep giving us more of the same there too? Making DDR4 as irrelevant as DDR3?"

The pitfall of this line of thinking is that we, as outsiders, don't know what trade-offs the design engineers are making in their decision of allocating x-amount of diespace for a larger cache, or spending more engineering resources crafting beefier prefetchers instead of allocating that same die-space and engineering resources towards developing every better intrinsically higher IPC circuits and so forth.

Today's processors don't need DDR4 because the engineers spent time, effort, and money (at a loss to spending it on other things while developing the CPU) to better insulate the processor from the deleterious performance robbing aspects of DDR3.

DDR4 may provide those same engineers, who are currently developing tomorrow's processors, with a lowered imperative for making design tradeoffs at the expense of developing better cores (more IPC, etc) instead of focusing those resources on developing better uncore circuits.

To throw in the obligatory automobile-inspired analogy, imagine if today's roads did not exist...that we all still drove on wagon-wheel rutted dirt and gravel roads. In such a world, today's auto engineers would find themselves spending an outsized percentage of their engineering resources in the development of tires, shocks, and suspension systems to insulate and isolate the rest of the car (and its occupants) from the realities of the road-surface.

Your jaguar wouldn't be a jag not just because of the road situation, but the jag engineers wouldn't have the resources to develop the rest of the car to be the high-performance machine that it is in our world.

Today's CPU engineers are sent into battle with one hand tied behind their back in that they are expected to spend a considerable amount of time and money developing support systems to buffer their high-performance cores from the dirty realities of what lies beyond the socket.

Better memory is like a better road, today's cars won't necessarily take immediate advantage of a smoother road but tomorrow's cars can be optimized and tuned for it, freeing up development resources that can then be used to improve other aspects of the processor.

 

[BallaTheFeared]

You might be compelled to respond along the lines of "well if the cache and prefetchers are doing such a bang-up job nowadays in insulating the CPU from the horrid performance of DDR3, then why not continue to depend on CPU designers to keep giving us more of the same there too? Making DDR4 as irrelevant as DDR3?"

Actually my thought was, how are you going to increase DDR bandwidth to the point where you wouldn't need to keep focusing on cache?

With Haswell Intel doubled the bandwidth, so obviously even with cache running away from DDR, they still needed more, twice as much in fact...

 

[dealcorn]

Intel introduced 22 nm with it's high power process and there was considerable delay in bringing the low power 22 nm process to market (Baytrail). At 14 nm, I believe Intel will lead with it's low power process. Have they also scheduled the high power process? Which process is used for Broadwell embedded desktop?

 

[IntelUser2000]

Intel introduced 22 nm with it's high power process and there was considerable delay in bringing the low power 22 nm process to market (Baytrail). At 14 nm, I believe Intel will lead with it's low power process. Have they also scheduled the high power process? Which process is used for Broadwell embedded desktop?

Bay Trail is end of this year/early next year, and Airmont is a year after that.

Broadwell is coming sometime in 2014, so at best, Atom will be coming at the same time.

 

[mikk]

In the Roadmap Broadwell-U and Broadwell-Y are placed within H1 2014. Although I think it is planned for around June, almost 2H 2014. I don't expect Airmont before Q4 2014. Broadwell should be available a couple of months earlier than Airmont. Of course the gap between the low process node and default process node narrows down.

 

[Ajay]

Intel has been about low power useage and has only low power low leakage 22nm 3D transistors to reduce leakage. As you push higher voltage the leakage increases massively and thats why clock speeds are worse than 32nm.

Well, based on Intel's own charts, xtor switching speed gains fall off quickly with increasing gate voltage - so there is a natural limiter built into finfet that didn't exist b/4.

Secondly, such small dice are naturally going to be more susceptible to changes in heat output since the heat flux will be much higher. Add to that Intel's current poor quality IHS/TIM combo and temps will, of course, rise quickly.

These two items alone could account for all of the difficulties overclocking 22nm CPUs. I haven't heard of any claims of massive increases in leakage for Ivy or Haswell.

 

[meloz]

In the Roadmap Broadwell-U and Broadwell-Y are placed within H1 2014. Although I think it is planned for around June, almost 2H 2014. I don't expect Airmont before Q4 2014. Broadwell should be available a couple of months earlier than Airmont.

Broadwell is 14nm. Airmont is 14nm. Both should be released about the same time, although it makes great sense to launch the lower TDP Airmont first since it can benefit from a manufacturing advanatge in the mobile segment. The smaller Atom die means it would also be easier to produce than Broadwell (relatively) and give better initial yields.

So Silvermont -which is yet to be released- will have a very short life indeed. 9-12 months. Airmont would be Silvermont shrunk to 14nm and that should put Intel in command position as far as perf/watt is concerned in very short order. Airmont's rivals designed by Apple and Qualcomm would still be ramping up* at TSMC 20nm, sans FinFET.


Back to desktop Broadwell, another crackpot theory:

Maybe they will release Broadwell for desktop but only in BGA form, while continuing to offer Haswell in LGA with Z97 and H97 chipsets?

By delaying the launch of socketed desktop Broadwell and by creating this artifical chasm, they can shepherd the market into buying BGA units from Intel while still doing lip-service to the socketed segment.




[*: I do not have any confidence in TSMC's ability to churn out 30,000+ wafers by the end of Q1 2014. They always experience a "wafer shortage" aka yield problems with a new node. They choked on 28 nm, 22nm was an abortion anyway and they will choke again on 20nm.

Maybe Intel choke, too, but they do not have external consumers to leak bad news and we never truly know when they start volume manufacturing just how good or bad their start yields are. I still remember reading that rumor/report about AMD having to throw away ~90% of the wafers when 28nm GPUs were first (paper)launched. No wonder the new generation cards were all so expensive back then. And no wonder TSMC had a "wafer shortage".]

 

[mikk]

Broadwell is 14nm. Airmont is 14nm.

Both 14nm but different process. Airmont is based on a 14nm ultra low power P1273 process while Broadwell should be based on a P1272 14nm process. Also different architecture. In the desktop space we know that Airmont is coming in 1H 2015. Realistically Q4 2014 for Tablet/Notebook.

 

[meloz]

Both 14nm but different process. Airmont is based on a 14nm ultra low power P1273 process while Broadwell should be based on a P1272 14nm process. Also different architecture. In the desktop space we know that Airmont is coming in 1H 2015. Realistically Q4 2014 for Tablet/Notebook.

Thanks to you, mikk, I stand better informed.

Since I did not know much about this, I searched for 'P1273 process' and found this document: Silicon Technology Leadership for the Mobility Era (Mark Bohr)

You have likely already read it, but others will find it interesting. And if this image is taken literally, it would definitely appear that the CPU (Broadwell) will come before SoC.

Also, some people have been channeling their disappointment with Haswell onto Intel's manufacturing and insisting that Intel's 22nm and future process are somehow unsuitable for higher performance CPU and such, take a look at this:

At 22nm, they have extended the curve to cover lower leakage without losing the option of high performance. And this is all SoC technology (P1271), not CPU. So they are not limited or crippled by their manufacturing process in anyway, but have even more options than before.

And note that the curve at 22nm is even more linear than at 32nm!

Okay, if we only judge progress by how much absolute transistor performance increases this might seem prosaic, but in the big picture this is huge. Mighty impressive. If this trend continues at 14nm Intel will own everyone even more so than now. Only their marketing department can stop them now.

EDIT: Blandge provided link earlier. I need to read better.

 

[Arachnotronic]

Yeah, Intel's process tech group is top notch; if only Intel had done a 32nm "tock" for Atom instead of warming over the Bonnell core, Intel may have been much more competitive at 32nm than it was...