Intel Skylake / Kaby Lake

[mikk]

What ? would you stop putting words in my mouth ? I said that since a 14core 2.5 ghz Intel with a 35 meg cache (E5-2683v3) only bests my 1700x at stock by a few %, that a 6 core skylake-X I don't think could beat it. Nothing about HEDT. And 6 core 12 thread Ryzens are $219 ! So stop the FUD until you get something to prove your point.

Comparing price points of some value consumer CPUs versus higher end Server CPUs never made sense, it's just silly. You might look for examples on server parts from Zen, if available, it would make a lot more sense. Expect a different price point for this sector. However Zen for server won't come up before Skylake server, any comparison in this regards with Broadwell is meaningless.

 

[Nothingness]

TSX was enabled on the Haswell Xeon E7 chips, but it was disabled on Xeon E5 Haswell (as well as client Haswell). It works on Broadwell-EP, Skylake, and Skylake Xeon.

Didn't know about that, thanks! But then your claim that Intel has had TSX for four years is still wrong

 

[raghu78]

Its the extra 50% of active transistors that will be the challenge, 5ghz 7700k is what about 140watts with highend cooling, so @ 6 cores even with process improvement your going to be very close to 200 watts, if your running high end air its probably going to be even higher as your temps will be higher.

Thats a good point. 6 core Coffeelake seems to come at around 150 sq mm so even though power draw will go up by 40-50% at max clocks die size is going up by around 20%.

http://wccftech.com/intel-coffee-lake-2018-cpu-details/

Thus Coffeelake 6C/12T will have a lot higher heat density / sq mm to cool compared to 7700k. Now we need to see how that affects max clocks. But I think 5 Ghz on 14++ even for 6C/12T should be possible.

 

[SAAA]

Improved drive current vs leakage current characteristics does not necessarily imply even higher clocks in the 5GHz territory, given that we don't know the operating voltages. People seem way overexcited from mere transistor performance curves and are already celebrating 5.2GHz for 6 cores when nothing is known for certain regarding voltages that might be required for such clock speeds.

6700K could reliably do 4.6-4.7 GHz, even 4.8GHz. 7700K reliably manages 4.8-4.9GHz. So that's a mere 200MHz considering 6700K@4.7GHz vs 7700K@4.9GHz; the latter comes at the cost of crazy temperature spikes and voltage requirements. But people still expect a 6 core CPU that is barely 1/6th bigger in size compared to its 4 core equivalent to hit 5GHz or more.

So much for rationality in some of the posts in this thread.

You guys expecting Coffee Lake (much less, Skylake-X) to be 5GHz are going to be disappointed. The 7700K was already outside of its ideal voltage to attain its stock clocks. Reaching 5 GHz on even 6 cores would be looking at 200W+ TDP...

~200W in 150mm2 shouldn't be a problem to cool when they weren't years ago:

3770K is 160mm2, close enough, delid and do the trick.

Thats a good point. 6 core Coffeelake seems to come at around 150 sq mm so even though power draw will go up by 40-50% at max clocks die size is going up by around 20%.

http://wccftech.com/intel-coffee-lake-2018-cpu-details/

Thus Coffeelake 6C/12T will have a lot higher heat density / sq mm to cool compared to 7700k. Now we need to see how that affects max clocks. But I think 5 Ghz on 14++ even for 6C/12T should be possible.

Yes the die overall will have more power to dissipate, but the core themselves are the same (?) and each one should run with slightly less power (14nm+ → 14nm++) thus there shouldn't be localized hot spots. And 200W+ quads were never a problem until Intel switched to damned paste... I seriously want to see them excuse again with "small die might crack under cycle load" after Zen uses solder and has a relatively small die too. That's definitely greed and nothing else.

Coming back to the point, Coffe silicon doing 5GHz is certain:

https://forums.anandtech.com/thread...ng-august-2017.2428363/page-404#post-38865823

~1/3 to 1/2 of simulated Kaby Lake 6 cores should reach 5GHz, I don't see how a better stepping/process can't make that happen.

 

[Markfw]

Comparing price points of some value consumer CPUs versus higher end Server CPUs never made sense, it's just silly. You might look for examples on server parts from Zen, if available, it would make a lot more sense. Expect a different price point for this sector. However Zen for server won't come up before Skylake server, any comparison in this regards with Broadwell is meaningless.

You are missing the point. Ryzen is a mainstream consumer CPU, but has the POWER to match both Intel mainstream AND some server CPU's, both of which are more expensive, up to 3 times the price.

 

[lolfail9001]

I seriously want to see them excuse again with "small die might crack under cycle load" after Zen uses solder and has a relatively small die too.

Zen uses die that is damn close in size to Broadwell-EP LCC, you know that, don't you?

 

[frozentundra123456]

I think people find it hard to let go the preconceived notions they have regarding certain products. For example AMD has a stigma of having poor IPC (ie. Bulldozer) hence Ryzen is "slow". When in fact Ryzen is a masterpiece of perf./watt efficiency and a very high IPC uarchitecture that only lacks full AVX execution width- which currently means jack sh*t for average Joe consumer. Even in server space it (HPC) accounts for roughly 15% of total market. AMD is positioned really well, best in the last 10+ years to attack intel's x86 dominion. I have a good feeling about their x86 cores this time around .

Off topic. This is a skylake thread. Plenty of other threads to extol the virtues of Ryzen.

 

[crashtech]

Its the extra 50% of active transistors that will be the challenge, 5ghz 7700k is what about 140watts with highend cooling, so @ 6 cores even with process improvement your going to be very close to 200 watts, if your running high end air its probably going to be even higher as your temps will be higher.

Another challenge may be to find motherboards that can supply the requisite current. This may also be an impediment to backwards compatibility as well. I'd assume any 300 series chipset intended for a hexacore will have correspondingly beefy VRMs, beefier than just about anything intended for a quad.

I really need to buy a motherboard as my current Z170 is partially defective, but if even a high-end Z270 with big VRMs won't work with 6c Coffee Lake, it might be time to just change out my entire platform instead of incrementally moving forward. I wish a few more leaks were out there, all I've seen is the one possibly dodgy SiSoft Sandra screenshot.

 

[lolfail9001]

Another challenge may be to find motherboards that can supply the requisite current.

Any $120+, even ITX one is ridiculous overkill for 7700k VRM-wise already. You just use those VRMs and you're happy with CFL-S OCd.

 

[IEC]

I personally opted to get rid of my Skylake system while it still had value, and got a good amount for it locally. I would be surprised if Coffee Lake ends up working on Z270 motherboards as each time I've bought an Intel processor I've had to buy a new motherboard. In my case with a Z170 motherboard I would almost certainly have to change motherboards anyways so clearing out my Skylake rig frees me to start from scratch with Coffee Lake if the performance and value are there.

 

[crashtech]

Any $120+, even ITX one is ridiculous overkill for 7700k VRM-wise already. You just use those VRMs and you're happy with CFL-S OCd.

You can't possibly know that. One spec sheet I recently read for a high-end Z270 claimed that the VRMs were good for 160W. That's below what I estimate an OCed 6C Coffee Lake could draw.

 

[lolfail9001]

You can't possibly know that.

Well, for those motherboards i checked it did. Granted, the list did not have any MSI boards.

One spec sheet I recently read for a high-end Z270 claimed that the VRMs were good for 160W

160W for what? For CPU? At what voltage, then? Anyways, i dare to guess it was MSI board because it looks about where their VRMs tend to blow up.

 

[IEC]

You can't possibly know that. One spec sheet I recently read for a high-end Z270 claimed that the VRMs were good for 160W. That's below what I estimate an OCed 6C Coffee Lake could draw.

VRM ratings also change with temperature and the vCore you are pumping through them. With MSI's affinity for using inefficient Niko Semi PKs, 160W+ to the CPU at 1.3V+ vCore would be... optimistic on some of their Z270 boards. For long term use, anyways.

 

[crashtech]

VRM ratings also change with temperature and the vCore you are pumping through them. With MSI's affinity for using inefficient Niko Semi PKs, 160W+ to the CPU at 1.3V+ vCore would be... optimistic on some of their Z270 boards.

Right, so even if we buy the rumor that Coffee lake will be backwards compatible, motherboard selection is likely to be critical for good results.

 

[tamz_msc]

~200W in 150mm2 shouldn't be a problem to cool when they weren't years ago:

3770K is 160mm2, close enough, delid and do the trick.



Yes the die overall will have more power to dissipate, but the core themselves are the same (?) and each one should run with slightly less power (14nm+ → 14nm++) thus there shouldn't be localized hot spots. And 200W+ quads were never a problem until Intel switched to damned paste... I seriously want to see them excuse again with "small die might crack under cycle load" after Zen uses solder and has a relatively small die too. That's definitely greed and nothing else.

Coming back to the point, Coffe silicon doing 5GHz is certain:

https://forums.anandtech.com/thread...ng-august-2017.2428363/page-404#post-38865823

~1/3 to 1/2 of simulated Kaby Lake 6 cores should reach 5GHz, I don't see how a better stepping/process can't make that happen.

Kaby Lake was touted as having 12% improved performance characteristics over Skylake, but in reality it doesn't amount to much when comparing desktop CPUs.

Looking at AnandTech's Skylake and Kaby Lake results,(motherboards are different so not a direct head-to-head comparison), we see similar power draw across the given frequency range.

Who knows if the 6700K had AVX offset and ability to stick with a bit more voltage, maybe it could also achieve the same 4.9-5.0GHz speeds that the 7700K does?

Even with a delid that drops temperatures by 20 degrees on the 7700K, they just hit a brick wall at 5GHz. Realistically one can expect Coffee Lake 4 core to behave the same at 5.1-5.2GHz, however it does not paint a good picture for the 6-core variant.

14nm vs 32/22nm comparison isn't apt because transistor density is much higher at 14nm. Before we had the cheap TIM controversy with Ivy Bridge, the improved 2700K could do 4.7GHz on 32nm! Therefore, looking back at Intel's uarches over the years, there hasn't been a terribly large improvement in clock speeds achieved with OCing as marketing would have you believe. These improvements will mainly be observed in the sub-45W mobile parts, not in a 95W TDP 6-core desktop CPU.

 

[AtenRa]

Competition is good, we now have 6-Core mainstream CoffeeLake 6 months earlier

 

[Conroe]

I don't think any Z chipset VRM would have a problem running a six core stock. I doubt any have a problem with an overclocked 7700K. The 7700K is not current limited it's temp limited. Adding two more cores will probably still only be temp limited on the higher end Z boards. I can see some manufactures not implementing BIOS if they think it could lead to problems, and even if they do update it for the six core they may not officially support it. It's happened before. My P5WDH got updates to run 45nm but it was not able to support all the features and they locked the bus at 1066. That was a high end board though.

 

[scannall]

I'm sure CL will bring improvements. But I gotta say, some of you may wish to temper your expectations a bit. Intel isn't going to cut their margins by any meaningful amount in the consumer space. They *may* offer bigger 'rebates' to OEM's, but even that's doubtful. And over 5 Ghz for a 6 core? Maybe a golden chip on water.

Leave some expectations on the table. That way you don't set yourself up for disappointment. Or if it does come in better than your tempered expectations, you can be pleasantly surprised.

 

[IntelUser2000]

For example caches dont get bigger for free, you trade latency/power/area/size/associativity, end result is a different set of trade offs

I think the trade-off will be with frequency. I wouldn't expect it to be large though. Maybe 100-200MHz.

The latency isn't necessarily impacted. Across the same design, going to a really large cache like on server has higher latency because the sheer size means it has to travel further. But on consumer designs, and especially with sizes as small as L2 its almost never true.

Core Duo had a 14 cycle latency L2 with 2MB size. Core 2 Duo had the same 14 cycle latency L2 but with double the size at 4MB. Interestingly Pentium M "Dothan" had lower latency at same 2MB indicating design decisions are much more important than physical size.* It's very possible the 1MB L2 might end up being 16-way with same latency as the 256KB 4-way one.

*Here the reasons are likely twofold. One to increase clock headroom and other because it moved to a shared cache between two cores.

I doubt that. I would say more, but then I would have to infract myself.

Haswell-E overclocks more than Broadwell-E and we know Broadwell chips have trouble reaching high frequencies. Just by that Skylake-X should clock higher. I think the 14nm+ benefits regarding max OC frequency will be negated by decisions to increase perf/clock for server designs though. Maybe it'll do 4.5. That's about the same as HSW-E. BDW-E was 4.2.

I'm sure CL will bring improvements.

I am not sure why people are expecting even 100MHz OC bumps for Coffeelake. It took saying "5GHz OC!!" 3 times before it became a reality. Even now, 7700K can't always do 5GHz. You are expecting a 6 core version to increase in frequency, when opposite will likely be true, even with 14++. The 12% gains are for much lower frequency models like laptops(will be running out of headroom soon) and servers.

 

[Sweepr]

Basin Falls SiSoftware Entry (KBL-X)

http://ranker.sisoftware.net/show_s...e4d9e9cfa79aaa8cf4c9f9dfbadfe2d2f487ba82&l=en

 

[Sweepr]

And Skylake-SP breaks another SiSoftware record, 5989.90 Mpix/s @ Processor Multi-Media:

2x Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz (28C 56T 3.8GHz, 2.4GHz IMC, 28x 1MB L2, 38.5MB L3)

 

[jpiniero]

Competition is good, we now have 6-Core mainstream CoffeeLake 6 months earlier

I guess the question is what compromises were made to rush this out.

 

[Markfw]

And Skylake-SP breaks another SiSoftware record, 5989.90 Mpix/s @ Processor Multi-Media:

2x Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz (28C 56T 3.8GHz, 2.4GHz IMC, 28x 1MB L2, 38.5MB L3)

I don't think those are skylake-SP, This is from http://www.tweaktown.com/news/57293/intels-new-xeon-rocks-28c-56t-costs-over-12-000/index.html

"Intel is continuing to show signs that it is scared to its core over AMD's new Ryzen processors, and the threat of the Naples-based 32C/64T processor has Intel pre-emptively launching its new Xeon Gold and Platinum platforms. The new series of CPUs from Intel are based on the new Purley platform, on LGA 3647."

And they cost $12,000 EACH

 

[lolfail9001]

I don't think those are skylake-SP, This is from http://www.tweaktown.com/news/57293/intels-new-xeon-rocks-28c-56t-costs-over-12-000/index.html

Skylake-SP is the name of the line-up/platform.

the individual SKUs are indeed called Xeon Unobtainium.

P. S. There is no real source for price tag, is there.

 

[mikk]

I don't think those are skylake-SP, This is from http://www.tweaktown.com/news/57293/intels-new-xeon-rocks-28c-56t-costs-over-12-000/index.html

Amazing how uninformed he is.




Callouts and insults are not allowed.


esquared
Anandtech Forum Director