Intel Skylake / Kaby Lake

[RichUK]

I'm debating getting an eBay special. 2500k with a planned oc of circa 4.5ghz and 8gb or ram - good enough to see me through on BF1 until the 7700k gets released? I've also got a gtx 960 2GB.

 

[wingman04]

What I feel CPU's need, above all, is more Memory Bandwidth.
More bandwidth is needed to take advantage of the vector units and the GPU (As Vectorization computation unit).

I hope next generation will have triple channel memory with higher speed (Higher speeds are clearly available on market).

I hope the E family will get 6 channels memory controller or maybe eDRAM.

This is what Intel should be doing with eDRAM, no only make it part of Iris Pro, but as a way to overcome the memory bandwidth limitation CPU's have.

So many operation in Data Science / Signal Processing are now Memory Limited that this will be a great thing to do.

How do you feel memory bandwidth?

 

[Arachnotronic]

I'm debating getting an eBay special. 2500k with a planned oc of circa 4.5ghz and 8gb or ram - good enough to see me through on BF1 until the 7700k gets released? I've also got a gtx 960 2GB.

don't waste your money -- just buy the 6700K today if you're gaming today. You can probably score a good deal on one (down to $329 for a boxed proc on Newegg: http://www.newegg.com/Product/Product.aspx?Item=N82E16819117559&cm_re=6700k-_-19-117-559-_-Product), especially as e-tailers clear inventory to make way for 7700K. The performance difference between them at stock won't be much, and that gap should remain similarly small when they're overclocked.

 

[mikk]

14nm yields were garbage when the 6700k launched. They're in good shape today.

Expect a hard launch with solid availability for 7700k.

Exactly this. Intel is already shipping those CPUs over 2 months before the official start is scheduled.

 

[witeken]

14nm yields were garbage when the 6700k launched. They're in good shape today.

Expect a hard launch with solid availability for 7700k.

What's garbage and how does it compare to TSMC ?

Don't forget that it doesn't take a lot of % worse yields to shave off a few % from Intel 65% gross margins (in desktop even higher).

 

[Drazick]

Yeah, that's why for SKL-X Intel goes with 6 channel. It is well proven that for BDW-E and the like memory bandwidth did not matter, though.
And of course there is no point in triple channel on mainstream platform.

Are you talking on Gaming Scenarios?
I was talking about Signal (Image) Processing / Data Science.

Many of them are Memory Bounded (Hence even multi threading is useless).
What the require is better cache system and memory bandwidth.

Easy to show that many algorithms in this world would scale linearly with memory bandwidth.

By the way, are you sure the next iteration of Skylake-E will have 6 memory channels?

 

[lolfail9001]

Are you talking on Gaming Scenarios?
I was talking about Signal (Image) Processing / Data Science.

Many of them are Memory Bounded (Hence even multi threading is useless).
What the require is better cache system and memory bandwidth.

Easy to show that many algorithms in this world would scale linearly with memory bandwidth.

By the way, are you sure the next iteration of Skylake-E will have 6 memory channels?

On about every non-gaming benchmark i have seen as well. Not to mention stuff you describe is better served with true bandwidth beasts: GPUs from what i remember of them.
Not Skylake-E but Skylake-EP, the Purley platform Xeons, by the way.

 

[RichUK]

don't waste your money -- just buy the 6700K today if you're gaming today. You can probably score a good deal on one (down to $329 for a boxed proc on Newegg: http://www.newegg.com/Product/Product.aspx?Item=N82E16819117559&cm_re=6700k-_-19-117-559-_-Product), especially as e-tailers clear inventory to make way for 7700K. The performance difference between them at stock won't be much, and that gap should remain similarly small when they're overclocked.

A 2500k at 4.5Ghz should be comparable to a 6600k / 6700k stock though, right?

I could get a 2500k, decent mobo and 8GB of RAM for the cost of a 6700k easy.

 

[WhoBeDaPlaya]

A 2500k at 4.5Ghz should be comparable to a 6600k / 6700k stock though, right?
I could get a 2500k, decent mobo and 8GB of RAM for the cost of a 6700k easy.

Figure a 25% IPC increase from Sandy -> Skylake, so that would equate to a ~3.6GHz 6600K.
Sounds about right.

If you have a mobo (Z68 / Z77) and/or DDR3 lying around, it's hard to beat the value of a 2500K
if you pick it up cheap. I think the lowest I bought one for recently (for spare gaming rigs) was $50 on CL.
Doesn't take much to cool it either - Hyper212s and CNPS5Xs can handle 4.4 - 4.6GHz OCs.

 

[Arachnotronic]

A 2500k at 4.5Ghz should be comparable to a 6600k / 6700k stock though, right?

I could get a 2500k, decent mobo and 8GB of RAM for the cost of a 6700k easy.

Why buy into old, used technology, especially if your endgame is to get a 7700K? (which is what your post indicated)

DDR4 is likely to be portable to a next gen system; you're at the end of the line with DDR3 (unless your next stop is used Haswell/Z97). 6700K is also faster, supports more threads (as games become more threaded, this will be useful), and comes as part of a more modern platform (PCIe 3.0, NVME, etc.).

 

[Drazick]

On about every non-gaming benchmark i have seen as well. Not to mention stuff you describe is better served with true bandwidth beasts: GPUs from what i remember of them.
Not Skylake-E but Skylake-EP, the Purley platform Xeons, by the way.

Again, those are not the scenarios I was talking about.
Regarding GPU's yes, they are better for those kind of tasks, yet still most of those tasks are done on CPU's.
Intel IPP & Intel MKL are libraries used in the Signal / Data Processing community and many of the functions there are Memory Bounded.

Most of the engineers work at work on their workstation (Usually Core i5 / i7) and they will benefit greatly from a thing like that.

AVX & SSE operations easily saturates the bandwidth of current CPU's.

 

[Soulkeeper]

What I feel CPU's need, above all, is more Memory Bandwidth.
More bandwidth is needed to take advantage of the vector units and the GPU (As Vectorization computation unit).

I hope next generation will have triple channel memory with higher speed (Higher speeds are clearly available on market).

I hope the E family will get 6 channels memory controller or maybe eDRAM.

This is what Intel should be doing with eDRAM, no only make it part of Iris Pro, but as a way to overcome the memory bandwidth limitation CPU's have.

So many operation in Data Science / Signal Processing are now Memory Limited that this will be a great thing to do.

I've felt the same way for awhile myself.
If I think of everything from a throughput perspective it makes the most sense.
I've been choosing my video cards based on bandwidth above all else for over 10yrs now, it's served me well.
I really liked my old core2q with 12MB of L2. It's s shame many CPU's still only have 4MB or 6MB of cache these days
I heard the edram on the intel chips after broadwell is more than just an eviction cache, so that's promising.
HBM is really what gets me excited going forward.

 

[wingman04]

Are you talking on Gaming Scenarios?
I was talking about Signal (Image) Processing / Data Science.

Many of them are Memory Bounded (Hence even multi threading is useless).
What the require is better cache system and memory bandwidth.

Easy to show that many algorithms in this world would scale linearly with memory bandwidth.

By the way, are you sure the next iteration of Skylake-E will have 6 memory channels?

Signal IMAG Processing /Data Science is highly serial operation and would not benefit from lock step memory Data retrieval, parallel bandwidth to you. Graphics benefit from from parallel memory retrieval, to a point.
With 2 channel 128 bit data is sent at one time.
With 4 channel 256 bit data is sent at one time.
Signal (Image) Processing / Data Science only needs 64 bit data is sent at one time or less. Speed of the memory and processor is what makes the difference.

 

[Drazick]

@wingman04 , I think you are wrong.
First of all, Data / Signal / Image processing is highly parallelizable task (Think of Filtering an image, Multiplying Matrix, etc...).

Secondly, if you use AVX (2 Units per core if I'm not mistaken) it means 256 [Bit / AVX Unit] * 2 [AVX Unit / Core] * 4 [Core] = 4096 [Bit] per CPU Cycle.
Memory throughput is far far from this.

Usually when the data is small, the cache system will be enough.
But in our days the data gets larger (Think of the size of modern image).

As a guideline, look how the CPU's FLOPS increased in the last 10 years vs. the memory Throughput.
On the last generation of CPU's, in tasks which uses most of their resources, unless the task is local (Working on L1 / L2 data), the memory can't keep up with them.
This is a concise choice and indeed CPU's should have more data crunch capabilities than the memory (They have Cache).
Yet the gap between the CPU's Number Crunching capabilities and the Memory Throughput is too wise these days.

We need more bandwidth from the memory system.
Both in more memory channels and speed.

 

[nvgpu]

http://www.anandtech.com/show/10802/desktop-kaby-lakes-lineup-base-frequencies-chipset-names

 

[plopke]

6700 will work on 200 series right ? If so then time to keep my eyes open on some 6700 deals next following months ;p.

 

[Majcric]

A 2500k at 4.5Ghz should be comparable to a 6600k / 6700k stock though, right?

I could get a 2500k, decent mobo and 8GB of RAM for the cost of a 6700k easy.

If you're going to be looking into older hardware at least consider the i7-2600k. If your budget allows go ahead and grab the i7-6700k, I upgraded from i5-2500k@4.5 to i7 6700k@4.5 and the difference is very noticeable in open world gaming/ and in min frames in general.

 

[jpiniero]

http://www.anandtech.com/show/10802/desktop-kaby-lakes-lineup-base-frequencies-chipset-names

I guess X299 is the Kaby/Skylake-X chipset.

 

[witeken]

Anyone else think Intel will deliver a revolution with ICL? http://www.fool.com/investing/2016/10/31/intel-corporation-needs-to-deliver-a-revolution-wi.aspx

 

[Sweepr]

Server Skylake (Purley) will be out Q1-2017

In the first quarter of 2017, Intel is set to release several new products for IoT applications including the top-end Purley-based server processors, Lewisburg-based chipsets, Kaby Lake-based S series processors for desktop and M/H/U series processors for mobile devices, new Atom and Quark processors.

http://www.digitimes.com/news/a20161101PD204.html


Preliminary Kaby Lake-S Family Specifications

Lenovo Miix 720 2-in-1 tablet leaks: High-res display, pen support, Kaby Lake CPU

http://winfuture.de/news,94688.html
https://liliputing.com/2016/10/leno...gh-res-display-pen-support-kaby-lake-cpu.html

While Microsoft didn't update the Surface Pro 4, Lenovo delivers.

 

[witeken]

Q1 for Purley, that's fast. If true, nice delievering since that was always planned in the roadmap.

 

[Sweepr]

• Microsoft Lifts the Lid on Some Intel Skylake-EP Details via the Open Compute Project

Adding some details - X299 & KBL PCH-X belong to socket LGA 2066 - the one Skylake-X / Kaby Lake-X will be using next year.

According to DigiTimes the top-end Purley server processors (LGA 3647) are launching in Q1-2017 - not ''well into 2017, if not the year after''.

http://www.anandtech.com/show/10809...kylakeep-details-via-the-open-compute-project


Entry level version of the new 13'' Macbook Pro is using 15W Core i5-6360U (Skylake-U GT3e):

https://browser.primatelabs.com/v4/cpu/915773


On another topic:

Puts in perspective the huge performance per core advantage Intel has right now.

 

[mikk]

Almost twice as fast at the same clock speed.


https://www.computerbase.de/2016-11/intel-core-i7-7500u-test/2/


They tested notebooks with a poor cooling (therefore CPU performance down in some cases), but what's interesting that the GPU Turbo stays so high. They said that a newer driver gives higher priority to the GPU over CPU and that gave them a boost in games on their Kabylake devices. That explains the surprising gaming results on 15W Kabylake. Because in the past for this increase Intel required a new GPU generation.

 

[wingman04]

@wingman04 , I think you are wrong.
First of all, Data / Signal / Image processing is highly parallelizable task (Think of Filtering an image, Multiplying Matrix, etc...).

Secondly, if you use AVX (2 Units per core if I'm not mistaken) it means 256 [Bit / AVX Unit] * 2 [AVX Unit / Core] * 4 [Core] = 4096 [Bit] per CPU Cycle.
Memory throughput is far far from this.

Usually when the data is small, the cache system will be enough.
But in our days the data gets larger (Think of the size of modern image).

As a guideline, look how the CPU's FLOPS increased in the last 10 years vs. the memory Throughput.
On the last generation of CPU's, in tasks which uses most of their resources, unless the task is local (Working on L1 / L2 data), the memory can't keep up with them.
This is a concise choice and indeed CPU's should have more data crunch capabilities than the memory (They have Cache).
Yet the gap between the CPU's Number Crunching capabilities and the Memory Throughput is too wise these days.

We need more bandwidth from the memory system.
Both in more memory channels and speed.

I would be interested to see Data / Signal / Image processing dual channel bench test vs quad channel bench test?

 

[lolfail9001]

Entry level version of the new 13'' Macbook Pro is using 15W Core i5-6360U (Skylake-U GT3e):

Ouch, memory latency on this thing is ridiculously bad.