Intel Skylake / Kaby Lake

[Dave2150]

Competition isnt part of it. The core improvement have pretty much been stable since Core 2.

What do you see as great gains in the future? Specially in a world that is focused on performance/watt. (And dont say moar cores).

In the future, we'll have radical new CPU architecture designs and new materials that will completely outperform silicon.

 

[mikk]

i5-6600 CPU @ 3.30GHz (32SP 4C 3.3GHz, 4GB) (OpenCL)= 85,38Mpix/s
http://www.sisoftware.eu/rank2011d/...efdcefd7e1d3e0c6b489b99ffa9fa292b4c7fac2&l=de


i5-4670 CPU @ 3.40GHz (32SP 4C 3.4GHz, 8GB) (OpenCL)= 74,02Mpix/s
http://www.sisoftware.eu/rank2011d/...efdcefd7e0d4e5c3b18cbc9aff9aa797b1c2ffc7&l=de


i5-4670 result is from me, CPU was running with 3.6 Ghz and DDR3-2133.

 

[Fjodor2001]

Quad-float GP Compute:

4670K: 3,90Mpix/s
6600: 3,64Mpix/s

Isn't that strange, when both Single and Double float were better for Skylake?

 

[Deders]

Kitguru reports that Techbang has reviewed a 6700k sample:

http://www.kitguru.net/components/c...published-up-to-29-faster-than-core-i7-4790k/

http://www.techbang.com/posts/24629...aymore-motherboard-measured-experience?page=2

Mixed results, Single threaded apps are on par with the 4970K, multithreading seems to be quite a bit faster. Firestrike is slightly slower, but other 3Dmark benchmarks show slight improvement.

 

[mikk]

Old news.

 

[Mako88]

Old news.

Doesn't matter. Bottom line is there's no reason to upgrade to Skylake, at all, versus even a nicely overclocked 2500k from two years ago.

Thanks to AMD for being so poor, so absolutely awful, that Intel has been able to essentially sit still with no real innovation. We all suffer when there's no competition.

 

[utahraptor]

This news is fairly sad. I am sitting on a 3770k that I want to replace badly. I now think I could easily wait for skylake e or kabylake.

 

[SAAA]

I wouldn't be so fast to dismiss Skylake especially after seeing this:

http://forums.anandtech.com/showpost.php?p=37461340&postcount=947

The Skylake sample leaked a month ago had better scores than this one in Sandra (yes I know, like 2%) but what you should be looking at are the Haswell scores: earlier results (to the right) were almost 30% worst than the current ones... you can't do IPC comparison when there's that much variation in a part.

Or rather I can suppose the same holds true for Skylake and it could reach much better final results, I estimated 27% better integer and similar floats to Haswell there.

PS: could this be the reason cinebench shows almost no improvement? Is it just a floating point test?

 

[mikk]

In the first leak there was a nice improvement in Cinebench R11.5. Only R15 doesn't show it if the leaks are legit. There is not enough data from real world benchmarks.

 

[ShintaiDK]

3½ weeks more and we know.

 

[inf64]

It would be rather funny situation if Broadwell was to end up faster than Skylake in game benchmarks with dGPU due to eDRAM XD. It has around 15-20% advantage there VS Haswell and I seriously doubt Skylake can make such a jump without that massive cache onboard. Let's wait and see.

 

[ShintaiDK]

It will be a strange transistion with eDRAM vs no eDRAM. I wonder what comes first, full eDRAM product stack or stacked memory. eDRAM is the halfstep there. And with Kabylake we can see a product with 2x128MB eDRAM.

It also shows where the quadchannel+large L3 cache comes to usage with the E series.

 

[inf64]

Honestly if I was a hardcore gamer with 980ti or FuryX in my system, I would just buy Broadwell if it ends up being faster than Skylake with dGPU. It does come with a price premium now but so will 6700K too, so it will be a wash price-wise.

 

[ShintaiDK]

Honestly if I was a hardcore gamer with 980ti or FuryX in my system, I would just buy Broadwell if it ends up being faster than Skylake with dGPU. It does come with a price premium now but so will 6700K too, so it will be a wash price-wise.

It will be very close it seems. Pretty much a wash as you say with one being faster here and the other there. Will be interesting to see how it pans out.

 

[SAAA]

It would be rather funny situation if Broadwell was to end up faster than Skylake in game benchmarks with dGPU due to eDRAM XD. It has around 15-20% advantage there VS Haswell and I seriously doubt Skylake can make such a jump without that massive cache onboard. Let's wait and see.

It will be very close it seems. Pretty much a wash as you say with one being faster here and the other there. Will be interesting to see how it pans out.

You know guys, I seriously think Intel just f* up big this time.
I mean that i5 is on par with a much faster Haswell chip, it has much better graphics than an equal EUs count only for the bandwidth... and we are not going to see L4 in mainstream Skylake?

How big would a 128/64MB, 14nm eDRAM be? Some 40 to 20mm2 and it could easily improve BOTH the CPU and GPU.
Plus they wouldn't even need to make the die bigger to include it: just cut the IGP some 20/30% and get back that performance with the cache alone, while probably having same power consumption, one die only and a handy L4!

I'm not an engineer (still) but something fishy: either that's lack of competition right there or just complacency.

I can just hope current Xeons with eDRAM are such a huge succes that they implement it on all the line and mainstream part as consequence, still it's a sad looking affair from a desktop user perspective.

 

[Sweepr]

It would be rather funny situation if Broadwell was to end up faster than Skylake in game benchmarks with dGPU due to eDRAM XD. It has around 15-20% advantage there VS Haswell and I seriously doubt Skylake can make such a jump without that massive cache onboard. Let's wait and see.

Indeed. I was expecting maybe a 5% bump (between 0-10%) but these results clearly show that the eDRAM is able to improve gameplay experience in CPU-limited titles, maybe more than the usual Intel 'Tock' IPC bump. Skylake-S vs Broadwell-K will be fun to watch.

Repeating what I say on the Broadwell thread:

If Intel keeps 65W Skylake 4C+GT4e (64MB eDRAM) clocks high enough next year then this might be a better gaming chip than 91W 4C Kabylake-K (assuming Kabylake packs higher-clocked Skylake cores, not a 14nm version of Cannonlake cores).

I hope eDRAM (or a similar/better solution) becomes more common in the next generations, including enthusiast chips.

 

[ShintaiDK]

You know guys, I seriously think Intel just f* up big this time.
I mean that i5 is on par with a much faster Haswell chip, it has much better graphics than an equal EUs count only for the bandwidth... and we are not going to see L4 in mainstream Skylake?

How big would a 128/64MB, 14nm eDRAM be? Some 40 to 20mm2 and it could easily improve BOTH the CPU and GPU.
Plus they wouldn't even need to make the die bigger to include it: just cut the IGP some 20/30% and get back that performance with the cache alone, while probably having same power consumption, one die only and a handy L4!

I'm not an engineer (still) but something fishy: either that's lack of competition right there or just complacency.

I can just hope current Xeons with eDRAM are such a huge succes that they implement it on all the line and mainstream part as consequence, still it's a sad looking affair from a desktop user perspective.

Some Skylake SKUs get eDRAM too. 14nm eDRAM doesnt really make sense besides the physical size. The current 128MB parts cost around 3$. However as we saw with Haswell, you couldnt fit GT3e+eDRAM on a LGA socket. Only BGA.

eDRAM is nothing more than a stop gap till stacked DRAM. Currently its still on the experimental stage outside the IGP usage.

The 128MB eDRAM at 22nm uses 976mw just in idle to refresh the DRAM.

 

[SAAA]

Some Skylake SKUs get eDRAM too. 14nm eDRAM doesnt really make sense besides the physical size. The current 128MB parts cost around 3$. However as we saw with Haswell, you couldnt fit GT3e+eDRAM on a LGA socket. Only BGA.

Well that was mostly my point for 14nm: it's so small it would fit on die with a minor IGP cut, a cut to EUs in favor of much better bandwidth anyway.

eDRAM is nothing more than a stop gap till stacked DRAM. Currently its still on the experimental stage outside the IGP usage.

The 128MB eDRAM at 22nm uses 976mw just in idle to refresh the DRAM.

Fair enough, still I'm pretty sure a couple Watts more at idle and load won't be a serious problem on any but smaller -U and -Y chips, those may be different dies anyway, dual core only floorplan.

Stacked DRAM is more like an additional feature to this one, the eDRAM is similar to a classic integration along with North and South bridge chips.

What I mean is that system memory and storage are different now only because hard disk trailed so hard against much faster RAM, but with SSD and flash we have seen what storage can do and I'm pretty sure that one day they will combine one another, either with this or newer technology.
The lesser steps there are in memory the better for the system in any case.

 

[ShintaiDK]

It seems PCM is the integration of memory and storage. Microsoft also got a filesystem for it. Will be interesting to see with the Skylake Purley platform launch that will support it.

I agree they could(and should) have been more aggressive with eDRAM implementation on the S/H lines. While the SKU amounts increase, it could go faster. But as long as we are bound by a certain LGA socket in size then there is also another side of the coin. Not just for current, but also future implementations.

 

[desura]

From the hype it sounded like Skylake would be 20% faster/power efficient than broadwell. If it just barely edges haswell despite having a process advantage...

 

[dark zero]

The Intel version of Bulldozer incomming?

 

[mikk]

The Intel version of Bulldozer incomming?

Unless Skylake IPC drops by 20% with higher power consumption no chance.

 

[MountainKing]

The Intel version of Bulldozer incomming?

That's what I thought

 

[DrMrLordX]

Honestly if I was a hardcore gamer with 980ti or FuryX in my system, I would just buy Broadwell if it ends up being faster than Skylake with dGPU. It does come with a price premium now but so will 6700K too, so it will be a wash price-wise.

Having a competent iGPU under the hood would also be good for whenever developers start making use of them in DX12/Vulkan applications. The downside: availability on Broadwell-C parts seems to be very low.

 

[Sweepr]

Unless Skylake IPC drops by 20% with higher power consumption no chance.

I guess the AMD crowd has to pick on something now that Fury X failed to beat NVIDIA at the same price point and there's nothing interesting from the CPU camp till Q4-2016 (plenty of PowerPoint slides though).

If Skylake is 10% faster than Devil's Canyon across the board at similar/lower power then it's going to be a much better product than Zambezi (Bulldozer) compared to its predecessors.

Also poor Carrizo failed to get significant design wins and will go head to head with Skylake-U after Windows 10 launch.