Search CPU-World

Search site contents:

Some features of Skylake graphics architecture

Westmere was the first Intel x86 microarchitecture to integrate a graphics unit together with CPU cores, since then the company made quite an effort to improve graphics performance. Intel implemented sharing of last level cache between the CPU and GPU, increased the number of shaders, or execution units (EUs), and optimized them. Current Haswell architecture introduced GT3 graphics, that increased the number of EUs to 40. It is anticipated that upcoming Broadwell GT2 and GT3 GPUs will have the number of execution units increased by 20% to 24 and 48. The next generation Skylake graphics will have up to 72 execution units, which signifies a 50% increase over Broadwell. As was already reported, Skylake processors will have several variations of die, incorporating 2 or 4 CPU cores, and GT2, GT3 or GT4 graphics. We presume that the GT4 has the maximum number of execution units (72).

Skylake GPUs will be built on "Generation 9 Low Power" architecture, that should substantially improve performance and reduce power consumption. Like Haswell and Broadwell based processors, some Skylake SKUs will have up to 128 MB of embedded DRAM, used as the last level cache for both the CPU and GPU. The graphics units will also add support for Shared Virtual Memory. As we understand, and this is not confirmed, this feature will allow the CPU and GPU to share system memory, which should boost performance of heterogeneous applications.

Skylake integrated video encoder/decoder will be enhanced, and it will support new codec types. The hardware decoder will work with JPEG, JMPEG, MPEG2, VC1, WMV9, AVC, H264, VP8 and HEVC/H265 video and image formats. The encoder will be support JPEG, MPEG2, AVC, H264, VP8 and HEVC/H265 standards. It is possible that Intel will add support for encoding/decoding of VP9 video. There will be new video filters as well.

Skylake microprocessors will come with DisplayPort, embedded DisplayPort, DVI and HDMI interfaces. The maximum display resolution of DP and eDP ports is 4096 x 2304 at 60 Hz using 24 bits per pixel, and 4096 x 2160 at 24 or 60 Hz, at 24 bpp for HDMI interface. The maximum resolution will be slightly lower on "Y" series chips.

Skylake GPU will support DirectX 11.2 and earlier, OpenGL 5.0 and OpenCL 2.1/1.2 APIs.

Related News (newer articles):

Related News (older articles):

Comments

There are 9 comments posted

Skylake EUs

2014-06-10 13:37:37
Posted by: gmb

72 EUs is odd. On HSW and BDW GT3 is a doubling over GT2 and GT4 should be a further doubling. It would mean for Skylake

GT2 18 EUs
GT3 36 EUs
GT4 72 EUs

GT2 only 18 EUs seems odd or maybe one EU is much stronger in Skylake.

SKL GPUs

2014-06-11 00:04:12
Posted by: Jon

You only think it's odd because you're looking at it wrong.
Each GPU isn't a discrete unit, they're made up of one or more "slices", because I can't remember the word Intel uses (Nvidia calls them SMXs).

In IVB each "slice" was made up of 8 EUs plus associated hardware.
GT1 has one slice(6EUs because it's cut down), GT2 has 2 slices(16EUs).

With HSW, each slice is made up of 10 EUs + other HW
GT1 has 1 slice(10EUs), GT2 has 2 slices(20EUs), and GT3 has 4 slices(40EUs).

For BDW, and, seemingly, SKL each slice has 12 EUs + addl. HW.
GT1 has 1, GT2 has 2(24 EUs), GT3 has 4(48EUs), and GT4 has 6(72 EUs).

Once you get past 2 "slices" you don't need to double the number of EUs each time, any more than AMD or Nvidia need to double the number of cores for each larger GPU.

You can be disappointed that it's not 24, 48, 96, but there's no reason to expect anything other than 24, 48, 72 configurations.

opencl

2014-06-10 14:35:16
Posted by: nahiiin

OpenCL 2.0 Drivers are going to be out in Dec at the latest.
So there is no such thing as OpenCL 2.1 let alone support for it in Skylake.

Apple introduced Metal API which used Shared Memory aka OpenCL 2.0. So it seems Skylake will be the first Intel GPU to support Metal. Unless OpenCL 2.0 can be supported in Broadwell.
Which should be possible since GPU is sharing main memory.

This would also be sign of death of discrete GPU unless
they magically some fancy memory and massive bandwidth.

 

2014-06-10 15:40:54
Posted by: gmb

This is a mid 2015 CPU and it wouldn't surprise if there is already work ongoing for a future OpenCL 2.1 version behind the scenes.

 

2014-06-10 22:37:10
Posted by: Nahiiin

The design for Skylake should be set in stone already.
It takes close to a year just to have drivers
support the latest iteration of opencl.

So Intel will bring out OpenCl 2.0 or just skip to 2.1 in
less than 6 months.

 

2014-06-10 20:12:26
Posted by: klon

I think the common denominator here is 24 EU for each GT step. So ...

GT2: 24 EU
GT3: 48EU
GT4: 72EU

GT1 if they still bother with that 12 EU.

Quicksync

2014-06-11 16:04:35
Posted by: Marcin

Are there any informations about quality improvment of quicksync encoding? I mean h265 is nice, but if the encoder has the same quality like h264, then it is nothing worth

 

2014-06-12 01:06:04
Posted by: Albert

By the time Skylake implement's HSA via shared virtual memory, AMD's Kaveri would still be 10 years ahead in its level of HSA development. I just wish more software developers would take it up and see their performance skyrocket as is shown in Libreoffice.

 

2014-06-26 22:47:12
Posted by: Patrick Proctor

You'd be wrong. Intel started doing heterogeneous research ages ago, hence the Larabee project. AMD is nowhere close to getting ahead. They just sunk money into it and got enough software people behind it Intel is finally putting research into production.

The 4790k crushes the a10 7850k in games even with Mantle. How's that for sour grapes? Not to mention the greater core count makes for some nifty mid-scale gpu acceleration to tinker with.

Terms and Conditions · Privacy Policy · Contact Us (c) Copyright 2003 - 2010 Gennadiy Shvets