Skylake Core Configs and TDPs

[stuff_me_good]

I don't see the idea behind wasting so many transistors just for GPU cache? Wouldn't it make the GPU way faster if they just use that cache transistor budget to make more EU's?

 

[ShintaiDK]

I don't see the idea behind wasting so many transistors just for GPU cache? Wouldn't it make the GPU way faster if they just use that cache transistor budget to make more EU's?

Its a seperate chip.

 

[Pottuvoi]

I don't see the idea behind wasting so many transistors just for GPU cache? Wouldn't it make the GPU way faster if they just use that cache transistor budget to make more EU's?

You need to feed those ALUs, ROPs and TEX units somehow to make meaningful computation possible.
Also having the data near the units consumes less power and thus it's possible to have faster chip within power constraints.

Data movement truly is the hard part of computing and data locality is one of the big things needed to decrease power usage and increase performance now that the low hanging fruit of power scaling has ended.

This explains things nicely.
http://techtalks.tv/talks/54110/#

Also the L4 cache is for both CPU & GPU.

 

[NTMBK]

I don't see the idea behind wasting so many transistors just for GPU cache? Wouldn't it make the GPU way faster if they just use that cache transistor budget to make more EU's?

No point adding more execution power if you haven't got the bandwidth to feed it. Look at Kaveri overclocking results- people pushed its GPU from 720MHz to >1GHz, with almost no increase in performance. Why? Because it is bandwidth choked.

Fast in-package "near" memory, backed up by slower DDR3/4 "far" memory seems to be the next step to improving memory bandwidth.

 

[Fjodor2001]

The numbers are copy/paste from other places. First page being one of them. And the 95W is the platform spec. Haswell and Ivy bridge for example was also listed as 95W. It doesnt mean there will be 95W.

And the 64MB have been up and debunked before. Lets see if it happens this time.

How do you know it's due to copy-paste from elsewhere? Also, why mention GT2 specifically if it's 95 W TDP platform spec for desktop Skylake in general, and not the TDP for specific SKU(s)?

Note that the article said "Quad-core processors with GT2 GPU will be also available with 95 Watt TDP."

 

[Homeles]

How do you know it's due to copy-paste from elsewhere? Also, why mention GT2 specifically if it's 95 W TDP platform spec for desktop Skylake in general, and not the TDP for specific SKU(s)?

Note that the article said "Quad-core processors with GT2 GPU will be also available with 95 Watt TDP."

Ivy Bridge sampled at up to 95W. When it launched, it was rated for 77W. Haswell sampled at 95W as well, and launched as an 84W part.

There's historical precedence here. You're not going to get very far by arguing against it.

 

[Fjodor2001]

Ivy Bridge sampled at up to 95W. When it launched, it was rated for 77W. Haswell sampled at 95W as well, and launched as an 84W part.

There's historical precedence here. You're not going to get very far by arguing against it.

You're missing the point, and not answering the question: "why mention GT2 specifically if it's 95 W TDP platform spec for desktop Skylake in general, and not the TDP for specific SKU(s)?"

 

[mavere]

You're missing the point, and not answering the question: "why mention GT2 specifically if it's 95 W TDP platform spec for desktop Skylake in general, and not the TDP for specific SKU(s)?"

The same reason desktop 84W Haswell parts are specifically GT2 and were labeled as 95W?

 

[Fjodor2001]

The same reason desktop 84W Haswell parts are specifically GT2 and were labeled as 95W?

Huh? If it was a platform TDP spec, there would be no reason to mention the GT version in the article, since it would apply regardless of that.

Note that the article states: "There will be three main configurations of "S" desktop parts: 2 CPU cores with GT2 graphics, 4 CPU cores with GT2 graphics, and 4 CPU cores with GT4 graphics and 64 MB of eDRAM. All configurations will be available with 35 Watt and 65 Watt TDP options. Quad-core processors with GT2 GPU will be also available with 95 Watt TDP."

I.e. it is specifically mentioning TDP per SKU, not general Skylake platform TDP.

 

[ShintaiDK]

Huh? If it was a platform TDP spec, there would be no reason to mention the GT version in the article, since it would apply regardless of that.

Note that the article states: "There will be three main configurations of "S" desktop parts: 2 CPU cores with GT2 graphics, 4 CPU cores with GT2 graphics, and 4 CPU cores with GT4 graphics and 64 MB of eDRAM. All configurations will be available with 35 Watt and 65 Watt TDP options. Quad-core processors with GT2 GPU will be also available with 95 Watt TDP."

I.e. it is specifically mentioning TDP per SKU, not general Skylake platform TDP.

The same manufactor posted the same things back then with previous models. Exact same type of model listings.

There are 3 quadcore GT2 models listed. 35, 65 and 95. That funny enough reminds of the GT2 T, S and regular today. While the GT3e today is 65W.

So can we move on?

 

[mikk]

95W is Intels standard TDP, it was the same for SB, IVB, HSW. Have look to earlier IVB and HSW Roadmaps, it was all rated with 95W. Exact SKU TDP is coming later for them.

 

[Homeles]

More of a Broadwell thing, but could very well apply to Skylake too:

2nd Generation Embedded DRAM with 4X Lower Self Refresh Power in 22nm Tri-Gate CMOS Technology
Highlights:
25% increased frequency (2GHz vs 1.6 GHz)
3x improved retention time (300us @95c vs 100us @93c)
4x reduction in charge pump power.
tRC increasing from 6 cycles to 10 cycles, citing power improvements as the reason.
Blurry photos.

14nm Clock Generator:
A 2GHz-to-7.5GHz Quadrature Clock-Generator Using Digital Delay Locked Loops for Multi-Standard I/Os in 14nm CMOS

 

[Fjodor2001]

95W is Intels standard TDP, it was the same for SB, IVB, HSW. Have look to earlier IVB and HSW Roadmaps, it was all rated with 95W. Exact SKU TDP is coming later for them.

Then why mention GT2 specifically, if it's platform TDP spec and does not only apply to a particular GT2 SKU? See earlier post about that for details.

 

[Fjodor2001]

More Skylake config info:

http://wccftech.com/intel-14nm-skylake-leak-multiple-edram-configurations-desktop-variant-tdp/

Basically one of the biggest points of Skylake is that it will feature support for DDR4 Ram and will be based on the 100 Chipset ‘Sunrise Point’ of the ‘Sunrise Bay’ platform.

• Now Variant ‘U’ and ‘Y’ will integrate this PCH (Platform Control Hub) on the die while the ‘S’ and ‘H’ variants will not and require the chip separately.
• The DMI 2.0 is getting replace with the DMI 3.0 Interface that promises speeds of upto 8 GT/s.
• Variants ‘U’ and ‘Y’ will support 1 DIMM Slot per Channel while the ‘S’ and ‘H’ Variants will support 2 DIMMs per channel.
• The IVR (Internal Voltage Regulator) is to be abandoned for Skylake Processors.

The summary of the configurations are as follows:

• SKL-Y Config 1: 2 Cores with GT2 Graphics and LPDDR3 1600Mhz Memory Support. TDP is 4W and eDRAM statistics are unknown, most likely because it will not have eDRAM.
• SKL-U Config 1: 2 Cores with GT2 Graphics and LPDDR3 1600Mhz Memory Support with 15W TDP.
• SKL-U Config 2: 2 Cores with GT3e Graphics and 64MB of eDRAM, supports LPDDR3 1600Mhz Memory and 28W TDP.
• SKL-H Config 1: 4 Cores with GT2 Graphics and DDR4 2133Mhz Memory Support. TDP is 35W.
• SKL-H Config 2: 4 Cores with GT4e Graphics and DDR4 2133Mhz Memory Support. Features 128MB eDRAM and a TDP of 45W.
• SKL-S Processors will support dual memory type namely the DDR3L/DDR3L-RS 1600Mhz and DDR4 2133Mhz. They will also support configurable TDP and come in two models, namely the 35W and 65W designs with the exception of the quad core (GT2 Only) which will come in a 95W TDP Design.
• SKL-S Config 1: 2 Cores with GT2 Graphics and Dual Memory and Dual TDP.
• SKL-S Config 2: 4 Cores with GT2 Graphics with Dual Memory and 95W TDP.
• SKL-S Config 3: 4 Cores with GT4e Graphics with Dual Memory and 64MB of eDRAM. These will feature Dual TDP.

​

Interesting that the IVR will be abandoned for Skylake. If I remember correctly, e.g. ShintaiDK said earlier that it was going to be kept?

​

 

[ShintaiDK]

What you link is wrong. The 44e for example is also under the 95W platform TDP.

 

[Fjodor2001]

What you link is wrong. The 44e for example is also under the 95W platform TDP.

Source? Also, any comment about the IVR being abandoned for Skylake?

 

[Techhog]

Losing IVR would suck a little just because it makes overclocking a little more complicated, but it would also mean higher overclocks... I guess it's okay overall.

 

[Homeles]

Source?

Page 1, post 1... hence the thread title.

 

[jpiniero]

Losing IVR would suck a little just because it makes overclocking a little more complicated, but it would also mean higher overclocks

I don't think you can really say that. I don't think there has been much about why they are removing it other than "The Skylake team doesn't like it".

 

[Homeles]

I don't think you can really say that. I don't think there has been much about why they are removing it other than "The Skylake team doesn't like it".

It runs at 90% efficiency... which means it adds 11% more heat to the chip. Right next to the cores, too.

 

[NostaSeronx]

The FIVR will return with Icelake. Icelake is the 10-nm tock.

Intel is going from a LDO linear regulator to a multi-level switching and buck converter regulator.

 

[mikk]

This is funny, the Oregon team obviously likes the FIVR and isreal team dislikes it.

 

[Homeles]

The FIVR will return with Icelake. Icelake is the 10-nm tock.

Intel is going from a LDO linear regulator to a multi-level switching and buck converter regulator.

Interesting catch... here's the source, for those wondering.

http://www.linkedin.com/pub/mathew-kottoor/3/798/bb5

 

[jpiniero]

It runs at 90% efficiency... which means it adds 11% more heat to the chip. Right next to the cores, too.

Which you would think would be fine... on mobile at least. The benefits would outweigh the heat cost.

Interesting catch... here's the source, for those wondering.
http://www.linkedin.com/pub/mathew-kottoor/3/798/bb5

Now that's funny. I still don't get how people can get away with putting confidential information like this on linkedin. Quick, post this to reddit and get this guy fired!

 

[NostaSeronx]

Intel is very open in regards with future plans.

Haswell/Broadwell;
https://www.google.com/patents/US20140070876

Icelake/7-nm Tick (I think);
https://www.google.com/patents/US20140092574

^-- There will probably be more finalized patents as Icelake comes up in 2017?