Intel Skylake / Kaby Lake

[ShintaiDK]

2x128MB would also be twice the bandwidth. Thats really interesting.

 

[Sweepr]

Something about Kabylake.

http://benchlife.info/cannon-lake-postpone-and-kaby-lake-will-replace-skylake-in-2016-06232015/


2x128MB edram for SKL-H looks insane. Overall it looks to me like some sort of refresh. Chipset seems new (USB 3.1 for desktop).

More from the link:

- Quad-core Kabylake + 2x 128MB eDRAM GT4e iGPU (mobile/server)
- Quad-core Kabylake @ 35-65W TDP (locked), 91W TDP (K, unlocked)
- Quad-core Skylake + 64MB eDRAM GT4e iGPU @ 35-65W confirmed for LGA1151 desktops

 

[ShintaiDK]

Its hard not to miss the Goldmont from the site. Tho this is a Skylake thread.

 

[Arachnotronic]

So Kabylake is, what? Skylake CPU with Cannonlake GPU in 14nm? Neat.

 

[Sweepr]

If I'm getting this right there will be:
Skylake 4C+GT4e 128MB eDRAM for mobile (and 1P servers if CPU-World is correct) - expected to be 50% faster than Broadwell-GT3e (@ 3DMark 11 and Cloud Gate).
Kabylake 4C+GT4e 64MB eDRAM for desktops next year (Broadwell-K successor). This one is actually a renamed Skylake part, so 72 EUs (Gen 9).
Kabylake 4C+GT4e 2x 128MB eDRAM for mobile, 72+ EUs (Gen 10?) and that much eDRAM should make it an iGPU powerhouse. Apple will love it.

 

[frozentundra123456]

Come on intel. All this cool sounding stuff but all we have so far is a phantom Broadwell-K that seems very limited in availability (well that and a bunch of 15 watt and lower ultrabook chips). Sometimes it seems like they are just making random combinations of cpus and dgpus and throwing it at the wall to see what sticks. Sort of reminds me of General Motors in the 80s and 90s. So many models that were similar. They tried to cover everything but ultimately just increased production costs and confused the consumer. All they have to do to induce enthusiasts to upgrade is give us "moar cores" (edit: or "moar ipc") instead of "moar igp".

 

[Ajay]

What the fuzzle is going on?! Skylake isn't even out yet and we get this Kabylake 'leak' ^_^

 

[Enigmoid]

I still can't understand why intel is not putting L4 on its server grade CPUs. Even ifs its several hundred dollars that's a drop in the bucket for CPUs costing $5000+. 2 x 128 MB L4 caches have got to help for servers and HPC. Maybe intel is waiting for HMC?

 

[Shehriazad]

What pricing would I possibly have to expect? Anyone got any leads on what to expect?


I want to make another cube PC with a R9 Nano and a 6600K...and of course some DDR4 ram.

I was hoping not having to exceed $/€600 for Board+CPU+Ram...but I am rather unsure.

 

[LTC8K6]

What the fuzzle is going on?! Skylake isn't even out yet and we get this Kabylake 'leak' ^_^

Maybe Intel is making sure AMD doesn't come close... :biggrin:

 

[VirtualLarry]

http://www.phoronix.com/scan.php?page=news_item&px=Intel-SKL-BXT-Firmware-Blobs

Has this been posted yet? Has some info on Skylake's iGPU, from a Linux perspective.

 

[Sweepr]

So what do you guys think Kabylake is?
Simple Refresh, Skylake CPU cores + Cannonlake iGPU or some kind of 14nm version of Cannonlake (new CPU cores + new iGPU)?

 

[Sweepr]

I also wonder how this affects Cannonlake and Icelake launch roadmap. I believe they will skip Cannonlake in desktops now (similar to Broadwell) but mobile might get Cannonlake in early 2017.

 

[Arachnotronic]

I also wonder how this affects Cannonlake and Icelake launch roadmap. I believe they will skip Cannonlake in desktops now (similar to Broadwell) but mobile might get Cannonlake in early 2017.

I think KabyLake will have a relatively short lifetime. I suspect it is being released in order to allow OEMs to release compelling system updates for the back-to-school selling season with something better than just a rehashed Skylake.

I am working under the assumption that KabyLake, as mikk claims, has Gen. 10 GPU/media.

 

[NTMBK]

I think KabyLake will have a relatively short lifetime. I suspect it is being released in order to allow OEMs to release compelling system updates for the back-to-school selling season with something better than just a rehashed Skylake.

I am working under the assumption that KabyLake, as mikk claims, has Gen. 10 GPU/media.

I don't think both of those can be true. Either it's a short lived refresh and it's just a clock-bump of Skylake (with a double Crystalwell SKU to keep Apple happy), or it's a whole new chip with a new graphics engine and is meant to last for minimum 1 year.

The additional cost of porting Kabylake's GPU design back to 14nm, redoing all of the masks for an entire range of chips (2C to 4C, GT1-4!) will be substantial. If this is only a 3-6 month stopgap, I can't see how the economics of that make sense. How many extra sales would they really get over a Skylake refresh with a slightly improved GPU?

Of course, it could be that the economics of 10nm are tricky, and Kabylake will stick around for low margin markets while Cannonlake goes for the high margin markets.

 

[IndyColtsFan]

So, I apologize if I missed this earlier in the thread. What does KabyLake mean for Skylake-E? Is Intel going to push out Skylake-E in 2016 or is it a 2017 part? I'm on an overclocked i7-2600K and was probably going to wait for Cannonlake or Skylake-E, but if Cannonlake is delayed (or is largely mobile-only) and Skylake-E is pushing into 2017, I'm wondering if I should just go ahead and jump to Skylake this fall.

 

[frozentundra123456]

Also, when is intel going to go to high bandwidth memory, whatever their equivalent is, instead of edram. Using edram now gives them an edge compared to AMD, but it seems at best a limited stopgap measure, and will be a poor solution when AMD finally gets HBM on their APUs.

 

[ShintaiDK]

Also, when is intel going to go to high bandwidth memory, whatever their equivalent is, instead of edram. Using edram now gives them an edge compared to AMD, but it seems at best a limited stopgap measure, and will be a poor solution when AMD finally gets HBM on their APUs.

We still have to see when AMD will use HBM on APUs. But the answer is simple for both parties. When the economics allows it. And I think we are close to 2020 before that happens on the mainstream. And then its over with flexible memory. You buy a CPU with x GB. So for example (hypothetical) if you want 16GB you need an i7 or A10 CPU.

Purley looks to be using HMC.

 

[NTMBK]

We still have to see when AMD will use HBM on APUs. But the answer is simple for both parties. When the economics allows it. And I think we are close to 2020 before that happens on the mainstream. And then its over with flexible memory. You buy a CPU with x GB. So for example (hypothetical) if you want 16GB you need an i7 or A10 CPU.

Purley looks to be using HMC.

My guess is that AMD will get it in 2017, when Zen APUs arrive. (I also suspect that this is the reason why Zen APUs are delayed.)

 

[Sweepr]

That remains to be seen (nothing about HBM in the leaked roadmaps yet), though their 2017 APUs will be severely limited if they rely on DDR4. I believe HBM will be limited to high-end APUs at launch (late 2017, 2018, later?), perhaps mobile only (less memory than desktops, 4-8GB total), it might take a while before it makes it to mainstream and low-end. Until then, Intel got the iGPU performance crown.

By the way, how much bandwidth does 2x128MB eDRAM provides?
Could Intel employ HCM in consumer chips later this decade?

 

[Nothingness]

By the way, how much bandwidth does 2x128MB eDRAM provides?

i7-4950 eDRAM bandwidth is 50 GB/s. See this: http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/3

 

[frozentundra123456]

That does not seem like much as they get into gt4e.

How does HCM differ from HBM? Can Intel license HBM or it exclusive to AMD/Hynix?

 

[Shehriazad]

That does not seem like much as they get into gt4e.

How does HCM differ from HBM? Can Intel license HBM or it exclusive to AMD/Hynix?

Nah, not "exclusive" so to speak. AMD just kinda worked closely together with SK Hynix for first gen HBM...which is why they could afford to develop a GPU with it before Nvidia...

But HBM is available to anyone....but research and implementation would probably take some time...which is why Nvidia skipped (another reason would probably be the problematic 4gb limitation that is too much of a hassle to get around)...as for Intel...they can afford not waiting for that and making their own stuff instead...I guess?

Or they might genuinely think that HMC will end up being superior for their personal aspects.

 

[ShintaiDK]

That does not seem like much as they get into gt4e.

How does HCM differ from HBM? Can Intel license HBM or it exclusive to AMD/Hynix?

Anyone can use and make HBM and HMC.

HMC currently ships in 2 and 4GB and 128 and 160GB/sec. While HBM is 1GB and 128GB/sec. Also why you see 16GB on Knights Landing, while only 4GB on Fury.

HBM is supposed to be the cheaper of the 2. And HMC is supposed to be more enterprise oriented. (Wide I/O for smartphones etc btw.)

But else the biggest differences is that HMC is serial based, while HBM is parallel. And HMC just uses PCB while HBM requies interposer.

 

[Ajay]

That does not seem like much as they get into gt4e.

How does HCM differ from HBM? Can Intel license HBM or it exclusive to AMD/Hynix?

Micron developed HMC with input from Intel - so Intel has already made it's choice and won't be using HBM standard.