Intel Skylake / Kaby Lake

[coercitiv]

you realize that Skylake is not offering any significant performance/watt improvement over its predecessor.

*snip*

One area where Intel have improved over the years is in reducing platform idle power consumption, and this is obviously important for battery operated toys. But my feel is that Intel have reached diminishing returns, and I for one would be surprised if upcoming Skylake notebooks offer tangible improvement over Broadwell.

I'm not pleasantly surprised by Skylake-S power usage, but let's not get overboard here: Intel is looking to improve perf/watt especially in the mobile sector, where frequencies are much lower, most chips hover near the 3Ghz mark. Moreover, most of their chips have base frequencies around 2Ghz, so their efficiency sweet spot must lie between these two values.

You do make a point though: if Skylake is to be a worthy successor for Broadwell in the mobile sector, it must improve load power consumption. (lower power at same performance, same power for more performance)

 

[witeken]

Okay, so here's a major plot twist for all you complainers: The problem isn't with the CPU, the problem is with the apps that have to feed the beast. They don't provide enough ILP, so there's little else Skylake can do. The only thing that matters is frequency, but we've seen how hot Skylake becomes.

In what way are ARM, NVidia, Apple and Qualcomm behind Intel?

I'm talking about architecture, IPC, raw max performance per clock, per benchmark.

When you look at benchmarks, the graphs are solely dominated by Intel's CPUs.

 

[Fjodor2001]

I'm talking about architecture, IPC, raw max performance per clock, per benchmark.

When you look at benchmarks, the graphs are solely dominated by Intel's CPUs.

If you sum up all aspects and metrics of the CPU, ARM does better in the low power segment. If not, why do you think basically all phone and tablet manufacturers use ARM, despite Intel's contra revenue program?

 

[witeken]

If you sum up all aspects and metrics of the CPU, ARM does better in the low power segment. If not, why do you think basically all phone and tablet manufacturers use ARM, despite Intel's contra revenue program?

Stop moving the goalposts. This thread and discussion is about 91W Skylake, it has nothing to do with mobile. That's a whole other discussion.

 

[Idontcare]

I don't know if Sandy Bridge's time is up now, but for me it was up when 22nm IB was released. At least 22nm IB gave a decent power reduction for same performance over SB.

But this 14nm Skylake launch just has me puzzled (business-wise, not technology-wise). Its like 14nm gave Skylake no benefit over 22nm Haswell. All the performance gains are architectural, designed in by some hard working engineers.

In the meantime, 14nm looks to be basically 100% cost-reduction focused (areal shrinkage maximized, electrical parametrics be damned) versus 22nm and 32nm.

So here is my half-baked assessment at this point - Whether or not Sandy Bridge is made obsolete by Skylake, the progression of diminishing returns with Intel's Tick-Tock tells me, as a consumer, that I may as well buy Skylake now because there is pretty much no damn good reason to hold out for any future Intel processor(s) as they are also just as likely to deliver, at best, a solid 5% performance bump in 1, 2, or 3 years time.

Intel's Skylake is telling us that this is not as good as it is going to get, but at the same time they are communicating that 10nm and 7nm are going to give us CPUs that might be, if we are lucky, another 10-20% over what Skylake is giving us today.

Buy one now, log out of your Anandtech forums account, and don't worry about the semiconductor world for at least 5 years. Save yourselves hundreds, if not thousands, of hours to go do something more productive with your lives in the meantime. Maybe 5nm will come along in 2022 and give us a grand total of 25% improvement in performance/watt over Skylake, then its time to consider an upgrade.

 

[ShintaiDK]

Don't know what is more disappointing: Skylake, or Anandtech's "review".

I expected this website's quality to fall after Shimpi sold out, just not this sharply. I would go in detail about all that can be improved, but why bother. Would be an utter waste of my time. Besides you can always go back and read earlier generation CPU reviews (like the Sandy Bridge review, that was a good one) and see how the mighty have fallen.

What is up with that laughable "Sandy Bridge, Your Time Is Up" warcry towards the end. Do Intel pay extra for publishing such twaddle? Perhaps they hope that people who do not read the benchmarks carefully will throw their perfectly decent Sandy Bridge systems and buy Skylake?

Rather underwhelming. I expected more. Both from Intel and Anandtech.

As for Skylake, where to begin. Another CPU arch that could have delivered so much more if not crippled in design. Intel keep cheapening out on basic things like PCIe lanes. They give their new generation CPUs absurdly low memory speeds, when anyone could see months/year ago that 2133 MHz was aiming too low. It's as if they want to kill desktop. Worst there is no alternative, crippled-by-design / mental retardation it might be on Intel's part and yet Skylake is still better than anything from AMD.

Here's hoping next launch (Kaby?) actually delivers something tangible and desirable. New CPU launches -on both sides, Intel and AMD- continue to get more and more disappointing. These days we live in hope for a better future, because actual CPUs are so damn underwhelming. :|

In terms of memory its not Intels fault. Its the memory industry that is getting further and further behind the curve. You cant validate something with DDR running out of spec one way or the other. If I recall right, DDR4 2400 at spec settings is just released short before Skylake launch.

In 2011 with SB, we had DDR3 1333 CL9 or so. With Skylake we got DDR4 2133 CL15 4 years later. Memory manufactors had dreamed of DDR4 3200 by now but forgot to deliver. (Modules running out of spec doesnt count).

 

[Ghaaandi]

Is there any Info out on the release date of mobile skylake processors?

 

[Fjodor2001]

Stop moving the goalposts. This thread and discussion is about 91W Skylake, it has nothing to do with mobile. That's a whole other discussion.

It was you who brought up the discussion, saying that "ARM, Nvidia, Apple, AMD and Qualcomm have tried catching up and are still far behind".

Now of those only AMD makes CPUs in the server and desktop segment, the rest only make CPUs in the low power phone and tablet segment. So there's nothing else to compare with really, than the mobile segment. I.e. if you need someone to blame for discussing mobile CPU performance of Intel vs the competition, blame yourself.

 

[ShintaiDK]

So no hdmi 2.0 confirmed? Looks like I'm sticking to current platform and just waiting for a GTX 950 then...

HDMI 2.0 is done via Alpine Ridge LSPcon. So look for that on a motherboard.

 

[CHADBOGA]

Buy one now, log out of your Anandtech forums account, and don't worry about the semiconductor world for at least 5 years. Save yourselves hundreds, if not thousands, of hours to go do something more productive with your lives in the meantime. Maybe 5nm will come along in 2022 and give us a grand total of 25% improvement in performance/watt over Skylake, then its time to consider an upgrade.

This is why I have rediscovered my love of bicycling and am just about to do a total rebuild of my old bike, turning it into one hell of an urban weapon.

Learning about derailleurs and gear ratio's and tyre characteristics has replaced what computers use to give me as a hobby.

The desktop scene is moribund.

 

[Idontcare]

In terms of memory its not Intels fault. Its the memory industry that is getting further and further behind the curve. You cant validate something with DDR running out of spec one way or the other. If I recall right, DDR4 2400 at spec settings is just released short before Skylake launch.

In 2011 with SB, we had DDR3 1333 CL9 or so. With Skylake we got DDR4 2133 CL15 4 years later. Memory manufactors had dreamed of DDR4 3200 by now but forgot to deliver. (Modules running out of spec doesnt count).

You raise an extremely important point.

Skylake was probably 70% of the way through electrical validation 12 months ago. Maybe even 85%.

To get to that stage of validation, a necessary stage so system integrators (board developers) can reliably develop products that incorporate engineering sample characteristics as being native and not as outlyer or bugs, Intel would have needed prototype memory for inhouse testing at least 18 months ago.

So what did the memory makers have to offer Intel 18 months ago? Even if just a handful of samples, prototypes of DDR4 products they intended themselves to make available to the market in 18 months time?

I doubt they (Intel or the memory companies) had access to early prototype DDR4-3000 or DDR4-3200. So how could Intel possibly build a roadmap to HVM of Skylake that incorporated validation with memory that simply wasn't available (even in one-off samples) 12-18 months ago?

 

[witeken]

I can only wonder what TuxDave is thinking. Those guys spend so many effort in those architectures and then they have to face multitudes of moaning and whining and disappointed people..

Can't wait for some deep dive details at IDF.

In the meantime, 14nm looks to be basically 100% cost-reduction focused (areal shrinkage maximized, electrical parametrics be damned) versus 22nm and 32nm.

Have you read the 14nm IEDM paper?

http://www.eetimes.com/document.asp?piddl_msgorder=%2D1%27&doc_id=1325023&page_number=2

http://www.intel.com/content/dam/www/public/us/en/documents/pdf/foundry/intel-14nm-iedm-2014-presentation.pdf

 

[ShintaiDK]

Is there any Info out on the release date of mobile skylake processors?

Yes.

 

[Idontcare]

This is why I have rediscovered my love of bicycling and am just about to do a total rebuild of my old bike, turning it into one hell of an urban weapon.

Learning about derailleurs and gear ratio's and tyre characteristics has replaced what computers use to give me as a hobby.

The desktop scene is moribund.

:thumbsup:

Good on you, great use of one's time

 

[witeken]

I don't know if Sandy Bridge's time is up now, but for me it was up when 22nm IB was released. At least 22nm IB gave a decent power reduction for same performance over SB.

But this 14nm Skylake launch just has me puzzled (business-wise, not technology-wise). Its like 14nm gave Skylake no benefit over 22nm Haswell. All the performance gains are architectural, designed in by some hard working engineers.

In the meantime, 14nm looks to be basically 100% cost-reduction focused (areal shrinkage maximized, electrical parametrics be damned) versus 22nm and 32nm.

So here is my half-baked assessment at this point - Whether or not Sandy Bridge is made obsolete by Skylake, the progression of diminishing returns with Intel's Tick-Tock tells me, as a consumer , that I may as well buy Skylake now because there is pretty much no damn good reason to hold out for any future Intel processor(s) as they are also just as likely to deliver, at best, a solid 5% performance bump in 1, 2, or 3 years time.

Intel's Skylake is telling us that this is not as good as it is going to get, but at the same time they are communicating that 10nm and 7nm are going to give us CPUs that might be, if we are lucky, another 10-20% over what Skylake is giving us today.

Buy one now, log out of your Anandtech forums account, and don't worry about the semiconductor world for at least 5 years. Save yourselves hundreds, if not thousands, of hours to go do something more productive with your lives in the meantime. Maybe 5nm will come along in 2022 and give us a grand total of 25% improvement in performance/watt over Skylake, then its time to consider an upgrade.

I'd rather see your analysis as a semiconductor engineer . I think you give Intel's process nodes too little credit, cause they are definitely better than 25%. Broadwell-y was at least 60%.

 

[ShintaiDK]

You raise an extremely important point.

Skylake was probably 70% of the way through electrical validation 12 months ago. Maybe even 85%.

To get to that stage of validation, a necessary stage so system integrators (board developers) can reliably develop products that incorporate engineering sample characteristics as being native and not as outlyer or bugs, Intel would have needed prototype memory for inhouse testing at least 18 months ago.

So what did the memory makers have to offer Intel 18 months ago? Even if just a handful of samples, prototypes of DDR4 products they intended themselves to make available to the market in 18 months time?

I doubt they (Intel or the memory companies) had access to early prototype DDR4-3000 or DDR4-3200. So how could Intel possibly build a roadmap to HVM of Skylake that incorporated validation with memory that simply wasn't available (even in one-off samples) 12-18 months ago?

Another example of how far DDR falls behind performance wise is the Purley/Xeon Phi platforms. DDR4 channels have been raised from 4 to 6. While HMC as local cache gets more and more popular.

The obvious question would then be, why didnt Intel add 1-2 more channels to Skylake-S. But the reason is cost.

Until we get HMC/HBM on CPUs, its just going to be a painful experience. Where we see good leaps in IPC in processing limited applications. But hardly anything in memory limited.

A last example would be the Broadwell-C and its eDRAM. Complete game changer as well in memory stuck games and applications.

 

[witeken]

This is why I have rediscovered my love of bicycling and am just about to do a total rebuild of my old bike, turning it into one hell of an urban weapon.

Learning about derailleurs and gear ratio's and tyre characteristics has replaced what computers use to give me as a hobby.

The desktop scene is moribund.

And it is why I completed yesterday The Idea Factory, a stellar book about all the innovation at Bell Labs in the 20th century.

 

[ShintaiDK]

I checked the DDR4 history line.

Memory makers had expected 2133Mhz DDR4 as standard, while 3200Mhz as enthusiast in 2012. And in 2013 they had expected 1V DDR4.

No wonder we are so far behind the curve.

 

[Brunnis]

Buy one now, log out of your Anandtech forums account, and don't worry about the semiconductor world for at least 5 years. Save yourselves hundreds, if not thousands, of hours to go do something more productive with your lives in the meantime. Maybe 5nm will come along in 2022 and give us a grand total of 25% improvement in performance/watt over Skylake, then its time to consider an upgrade.

Thanks for very thoughtful post. I've been vigorously reading tech forums for more than 15 years. I've spent more time reading about tech stuff than I care to think about. And I agree with you. While there was a lot going on in the desktop world in the 2000s, it's just not the same anymore. The desktop world is very mature and unexciting at this point and I keep returning more out of habit than anything else, really.

I will probably end up with a 6700K system that I'll enjoy for at least as long as my current i7-860. Feels pretty nice, actually.

 

[mrmt]

So here is my half-baked assessment at this point - Whether or not Sandy Bridge is made obsolete by Skylake, the progression of diminishing returns with Intel's Tick-Tock tells me, as a consumer, that I may as well buy Skylake now because there is pretty much no damn good reason to hold out for any future Intel processor(s) as they are also just as likely to deliver, at best, a solid 5% performance bump in 1, 2, or 3 years time.

From my perspective here every penny at Intel is being channeled at servers and mobile devices. In terms of server we have significant performance uplifts happening, and Intel is taking on the memory issues with their own hands, and mobile gets basically all the node improvements.

So yes, if a desktop is what turns you on, better to find a new hobby. I'm betting that my 4770 will stay with me safely for the next 5-6 years without much trouble. only disagree with you that we'll see 5nm on 5 years. The companies working on it are being a lot subjective and a lot less sure of what they can deliver and how they can deliver, let alone when they will deliver. I think we'll be lucky if we get to see it on 8 years from now on, with 7nm being a hell of a long node.

Ed: don't you think that a cost optimized 10nm would be interesting for reenter the memory business?

 

[TechFan1]

Don't think I've seen it mentioned in this thread yet, but it was kind of cool to see Intel advance the feature level support of their igpu so much. It was amusing to me to see Intel have DX feature level 12.1 support, and that it had more feature support than amd and NVidia for the moment.

 

[Dresdenboy]

Don't think I've seen it mentioned in this thread yet, but it was kind of cool to see Intel advance the feature level support of their igpu so much. It was amusing to me to see Intel have DX feature level 12.1 support, and that it had more feature support than amd and NVidia for the moment.

I think I saw 12_1 mentioned by Nvidia, but I'm not sure about that.

Are those features useful if one has to use lower settings?

 

[cantholdanymore]

I don't know if Sandy Bridge's time is up now, but for me it was up when 22nm IB was released. At least 22nm IB gave a decent power reduction for same performance over SB.

But this 14nm Skylake launch just has me puzzled (business-wise, not technology-wise). Its like 14nm gave Skylake no benefit over 22nm Haswell. All the performance gains are architectural, designed in by some hard working engineers.

In the meantime, 14nm looks to be basically 100% cost-reduction focused (areal shrinkage maximized, electrical parametrics be damned) versus 22nm and 32nm.

So here is my half-baked assessment at this point - Whether or not Sandy Bridge is made obsolete by Skylake, the progression of diminishing returns with Intel's Tick-Tock tells me, as a consumer, that I may as well buy Skylake now because there is pretty much no damn good reason to hold out for any future Intel processor(s) as they are also just as likely to deliver, at best, a solid 5% performance bump in 1, 2, or 3 years time.

Intel's Skylake is telling us that this is not as good as it is going to get, but at the same time they are communicating that 10nm and 7nm are going to give us CPUs that might be, if we are lucky, another 10-20% over what Skylake is giving us today.

Buy one now, log out of your Anandtech forums account, and don't worry about the semiconductor world for at least 5 years. Save yourselves hundreds, if not thousands, of hours to go do something more productive with your lives in the meantime. Maybe 5nm will come along in 2022 and give us a grand total of 25% improvement in performance/watt over Skylake, then its time to consider an upgrade.

I'm totally following IDC advice, but waiting for the 6 core skylake-E

 

[witeken]

only disagree with you that we'll see 5nm on 5 years. The companies working on it are being a lot subjective and a lot less sure of what they can deliver and how they can deliver, let alone when they will deliver. I think we'll be lucky if we get to see it on 8 years from now on, with 7nm being a hell of a long node.

Ed: don't you think that a cost optimized 10nm would be interesting for reenter the memory business?

We'll still get new nodes from Intel every 2-3 years for many generations to come.

 

[mrmt]

We'll still get new nodes from Intel every 2-3 years for many generations to come.

Not sure about that Witeken. That Intel delayed two nodes in a row should mean an structural shift in both business and R&D model, and the culprits will only get worse at 7nm and 5nm.