Intel Broadwell Thread

[frozentundra123456]

Broadwell doesnt compete with ARM. It runs x86 apps so is much more useful in that regard for business/productivity. Atom competes with ARM for cheap media consumption devices. Saying Broadwell doesnt compete with ARM is like saying a Lexus doesnt compete with a Yaris because it costs more. Duh, that is the point, it is a premium device.

Personally, after living with an android tablet for 2 years, I would *never, ever* buy another one. Since I dont use it for business though, I would get a cheap windows tablet or something like the Asus Transformer T-100.

 

[jdubs03]

Unfortunately, it appears that being on SATA2 is holding me back quite a bit, so I'm itching for an upgrade, even though I am more than fine with my i5 750 as well.

I totally agree with you on laptops though. Haswell's big enough of an upgrade, even for Ivy Bridge users. Obviously, most people wouldn't be down to buy a new laptop only one year later, but there would actually be tangible gains for them to do so, which generally isn't the case.

At this point though, I think it'd be better waiting for Broadwell. Waiting 6(?) months for the kind of battery life and graphics performance it'll be bringing should be fairly reasonable. I'm also a bit pessimistic when it comes to Skylake for mobile, as the loss of the IVR is a big bummer.

With Skylake we know very little, but I think that with Broadwell-Y going down to such a low power envelope, Skylake-Y can be focused mainly for performance, along the lines with Broadwell-U and onward. But yea, waiting for Broadwell can't hurt, the benefit from the shrink will allow for a sizable improvement. I'm curious about the U-series, because it would make sense for them to focus on performance. So that'll be interesting.

 

[cbn]

agreed. these guys are basically blasting the low end of the china market. playing in the dirt with allwinner etc.

http://online.wsj.com/articles/for-...henzhen-holds-key-to-tablet-market-1405106599

Of course but to get into the Shenzen whitebox tablet makers that's what it takes. Getting some market share and getting their supply chain straightened out so they can compete with the tablet chip makers is really all that Intel can do at this point since the smart phone market is out of reach until Intel can finally sort out a decent integrated LTE modem which is 2015 at the earliest and that will only be good enough for non western markets.

Yes, with regard to atom, I feel Intel should compete with the likes of AllWinner. (Hopefully the Rock chip alliance works well enough for this purpose)

Obviously it is not going to be immediately profitable, but I believe it is the sort of thing that helps Intel ensure the long term viability of x86.

 

[Shmee]

I wonder how they will OC?

 

[jihe]

I wonder how they will OC?

Will it blend?

 

[witeken]

DirectX 12 will be a great combination with Core M. It should be much better to feed the 24EUs without the CPU consuming half the 4.5W TDP.

So to put the significance of this in numbers, there will be a 2X efficiency from going to 14nm and Broadwell, the Gen8 architecture will be maybe 50% more efficient (which seems a reasonable assumption compared to Maxwell's 100%), and DirectX 12 gives us another 2X improvement, for a total of a 6X improvement.

 

[Enigmoid]

DirectX 12 will be a great combination with Core M. It should be much better to feed the 24EUs without the CPU consuming half the 4.5W TDP.

So to put the significance of this in numbers, there will be a 2X efficiency from going to 14nm and Broadwell, the Gen8 architecture will be maybe 50% more efficient (which seems a reasonable assumption compared to Maxwell's 100%), and DirectX 12 gives us another 2X improvement, for a total of a 6X improvement.

That seems almost ludicrously optimistic.



(FYI maxwell is not 100% more efficient on 28nm. GM107 is almost 2x the power of GK107 but uses quite a bit more power on a larger die).

 

[witeken]

That seems almost ludicrously optimistic.



(FYI maxwell is not 100% more efficient on 28nm. GM107 is almost 2x the power of GK107 but uses quite a bit more power on a larger die).

Okay, it's not 100%, it's 82%.

Given that Gen7 is a worse architecture than Kepler, 50% doesn't seem optimistic at all.

 

[pw257008]

though there's no solid evidence on improvements in Broadwell iGPU from Intel

 

[antihelten]

Okay, it's not 100%, it's 82%.

Given that Gen7 is a worse architecture than Kepler, 50% doesn't seem optimistic at all.

Actually it's only about 55%.

That graph mixes power consumption from only metro: last light with average performance across their entire test suite.

If you only look at power consumption in metro: last light combined with the performance in just metro: last light, you end up with the the 750 ti being about 55% more efficient than the best Kepler card (650 ti).

Interestingly 780 is almost exactly the same efficiency as 650 ti, indicating that at least for Kepler efficiency can be kept constant even with increased die size, if the card is balanced optimally (780ti and titan aren't far behind).

Edit: Admittedly if you only compare GM107 with GK107 it's about 79%, but GK107 is not a particularly efficient card (once you have something that slow, components on the card other than the die itself arguably start to dominate the total power draw) and not really representative of what the Kepler architecture is capable of.

 

[Phynaz]

though there's no solid evidence on improvements in Broadwell iGPU from Intel

I have to ask where you've been if you haven't seen it.

 

[antihelten]

I have to ask where you've been if you haven't seen it.

I think it's fairly obvious that pw257008 made that statement in regards to Witekens claim of a 50% efficiency improvement for Gen8, and whilst we have certainly seen plenty of stuff that would indicate improved performance for Gen8, I don't think we've necessarily seen anything that would indicate improvements in efficiency (the stuff about DCC in the Anandtech preview, would mainly affect idle power use).

 

[pw257008]

I think it's fairly obvious that pw257008 made that statement in regards to Witekens claim of a 50% efficiency improvement for Gen8, and whilst we have certainly seen plenty of stuff that would indicate improved performance for Gen8, I don't think we've necessarily seen anything that would indicate improvements in efficiency (the stuff about DCC in the Anandtech preview, would mainly affect idle power use).

yep. i wasn't clear. i meant concrete evidence on the level of performance and efficiency improvements (in response to witeken's extrapolations on extrapolations).

 

[mikk]

that would indicate improvements in efficiency (the stuff about DCC in the Anandtech preview, would mainly affect idle power use).

Bigger caches, higher texture throughput, microarchitecture improvements for Geometry, Z, Pixel Fill is an indicator.

 

[witeken]

Bigger caches, higher texture throughput, microarchitecture improvements for Geometry, Z, Pixel Fill is an indicator.

I hadn't noticed this yet.

The ramifications of this is that not only is the total number of EUs increased by 20% from 20 to 24, but Intel has greatly increased the ratio of L1 cache and samplers relative to EUs. There is now 25% more sampling throughput per EU, with a total increase in sampler throughput (at identical clockspeeds) of 50%.

 

[mikk]

Yes and there is also an older report from CPU-world.

The list of improvements includes increased size of various GPU caches

http://www.cpu-world.com/news_2014/2014050101_Intel_Broadwell_graphics_enhancements.html


Intel didn't go in detail about Gen8, maybe at IDF. For example they made lots of changes within one EU according to a Linux driver dev. All these detailed information are missing on the slide. But even without a deeper look into Gen8 is a statement like there is nothing that would indicate improvements in efficiency bizarre. 25% more sampling throughput per EU is one improvement that should be a big help because with MSAA and high graphics settings Intel suffered a lot with Gen7. It could also improve the EU scaling for GT3.

 

[witeken]

From a previous message that I posted:

There are also many things about graphics, with the most interesting one probably being The Compute Architecture of Intel® Processor Graphics Gen8:

This session describes the architecture’s compute components in detail. It is an ideal foundation for anyone seeking to better understand or better optimize compute applications for Intel® Processor Graphics.

Topics include:
• Description and explanation of the architecture’s components and scaling for different products
• Explanation of the architecture’s connection to other SoC components.
• Explanation of the architecture’s memory
• Explanation of new components for supporting programmability features such as shared virtual memory

Another notable session is: Intel® Processor Graphics: Power Optimization in Graphics Architecture

As computing devices get smaller, optimizing the platform for great performance in smaller power budgets is key to delivering the best user experience.

This session will:
• Explain why new architectural techniques are important to deliver better performance in lower power:
The physics of power consumption
Implications for power and performance vs. device size
• Discuss and demonstrate some features of the Intel® Graphics Architecture which make it ideal in those environments

So we should hear a lot about Gen8 at IDF indeed.

 

[antihelten]

Yes and there is also an older report from CPU-world.

http://www.cpu-world.com/news_2014/2014050101_Intel_Broadwell_graphics_enhancements.html


Intel didn't go in detail about Gen8, maybe at IDF. For example they made lots of changes within one EU according to a Linux driver dev. All these detailed information are missing on the slide. But even without a deeper look into Gen8 is a statement like there is nothing that would indicate improvements in efficiency bizarre. 25% more sampling throughput per EU is one improvement that should be a big help because with MSAA and high graphics settings Intel suffered a lot with Gen7. It could also improve the EU scaling for GT3.

Stuff like improved sampling throughput, bigger caches, etc, will undoubtedly improve performance, but that doesn't mean it will necessarily improve efficiency (i.e. performance/watt).

Note, I'm not saying that those things can't or won't improve efficiency, merrily that we can't say for certain with the available information.

 

[Enigmoid]

Stuff like improved sampling throughput, bigger caches, etc, will undoubtedly improve performance, but that doesn't mean it will necessarily improve efficiency (i.e. performance/watt).

Note, I'm not saying that those things can't or won't improve efficiency, merrily that we can't say for certain with the available information.

To be fair they are moving to 14 nm from 22 nm. That won't increase the architectural efficiency but will reduce the consumption of the igp at a given performance level.

 

[antihelten]

To be fair they are moving to 14 nm from 22 nm. That won't increase the architectural efficiency but will reduce the consumption of the igp at a given performance level.

Absolutely, but we're talking about efficiency gains purely from the architectural changes from Gen7.5 to Gen8 here (since Witeken claimed that a 50% improvement purely from going to Gen8, was a reasonable assumption)

 

[Wall Street]

Looking at the die sizes of the high end desktop offerings we see:

Lynnfield 296 mm2
Sandy Bridge 216 mm2
Ivy Bridge 160 mm2
Haswell 177 mm2

Do you guys think that the High End Desktop (HEDT) version of Broadwell will be 4 cores and ~100 mm2 for $200-300? If so, joke is on us enthusiasts as Intel is now selling 1/3 of the die area for the same price (or $20-40 more) as they did five years ago.

 

[jpiniero]

Sandy Bridge-E is 270 mm and 416 mm (for the 8 cut down to 6).
Ivy Bridge-E is 259 mm

 

[AtenRa]

SB-E Quad Core (Core i7 3820) die is 294mm^2, 8 Core die size is 435mm^2

 

[jpiniero]

Absolutely, but we're talking about efficiency gains purely from the architectural changes from Gen7.5 to Gen8 here (since Witeken claimed that a 50% improvement purely from going to Gen8, was a reasonable assumption)

Intel did say that Gen8 can also maintain it's clock speed much better when thermally constrained.

 

[witeken]

Absolutely, but we're talking about efficiency gains purely from the architectural changes from Gen7.5 to Gen8 here (since Witeken claimed that a 50% improvement purely from going to Gen8, was a reasonable assumption)

Exactly. This architectural efficiency discussion started after I did a 6X total GPU power efficiency improvement estimate, with 2x from DirectX 12, 2x from 14nm and >1.5x from Gen8.

Looking at the die sizes of the high end desktop offerings we see:

Lynnfield 296 mm2
Sandy Bridge 216 mm2
Ivy Bridge 160 mm2
Haswell 177 mm2

Do you guys think that the High End Desktop (HEDT) version of Broadwell will be 4 cores and ~100 mm2 for $200-300? If so, joke is on us enthusiasts as Intel is now selling 1/3 of the die area for the same price (or $20-40 more) as they did five years ago.

Don't forget that the cost per mm² also increased by some factor. I don't know by how much, but if it's say 50%, that's only a drop of less than 1/2th in equivalent die area (296/(111*1.5)=1.77). BTW, if you want more die area, Intel will be happy to sell Iris Pro Broadwell-K to you.