Intel Skylake / Kaby Lake

[Edrick]

Unless I'm missing something, I'm not sure you'll save much on power / heat by turning off HT. No material benefit, at least.

Before I had my 6600K, I had a 4770K. I ran tests with both HT on and off. With HT off, I was able to get the same OC with between 0.10 and 0.15 less vCore voltage and temps were between 3-5 degrees (F) lower. Not a huge difference, but noticeable. And I typically did not see much performance increase with HT on, so that is why when I went Skylake, I decided to go for the 6600K instead of the 6700K.

Edit - And my power draw was between 8-10 watts lower at the same OC. (Tested in AIDA)

 

[vissarix]

Damn Intel did an excellent job with Skylake X..immagine the 18 core part overclocked to 4.5 - 4.7ghz...What a beast of a cpu

 

[Markfw]

Damn Intel did an excellent job with Skylake X..immagine the 18 core part overclocked to 4.5 - 4.7ghz...What a beast of a cpu

With more cores, the top speed generally goes down. My guess is 3.4 -3.8 for the 18 core ?

Anybody else guess ?

 

[DrMrLordX]

Imagine the heat and power consumption of the 18c part @ 4.5 GHz. Whew.

 

[raghu78]

With more cores, the top speed generally goes down. My guess is 3.4 -3.8 for the 18 core ?

Anybody else guess ?

This is the first time ever that Intel is using a MCC die for their HEDT lineup (thanks to Ryzen Threadripper). So it would be interesting to see how max clocks are affected by a big increase in die size. LCC - 12 cores , MCC - 20 cores. 66% more cores. I think 3.8- 4.0 Ghz should be possible on the MCC since LCC is able to do 4.6 Ghz without delidding. I think the limiting factor could be the power circuitry on the motherboard and how much power can be safely driven through the LGA 2066 socket.

 

[moonbogg]

This is the first time ever that Intel is using a MCC die for their HEDT lineup (thanks to Ryzen Threadripper). So it would be interesting to see how max clocks are affected by a big increase in die size. LCC - 12 cores , MCC - 20 cores. 66% more cores. I think 3.8- 4.0 Ghz should be possible on the MCC since LCC is able to do 4.6 Ghz without delidding. I think the limiting factor could be the power circuitry on the motherboard and how much power can be safely driven through the LGA 2066 socket.

I've heard people say things like this. "Able to do". People might assume this means "If you buy this chip, you will do 4.6ghz without delidding". I think what it really means is something more like, "The chip will do 4.6ghz without throttling", which means you get 4.6ghz at 100c under load. So yeah, I guess you can say it will do it, but who's going to be happy with a chip at 100c?
I don't get excited when people say the chip can do this or do that and fail to mention temps and the kind of load in the same sentence.

 

[aigomorla]

With more cores, the top speed generally goes down. My guess is 3.4 -3.8 for the 18 core ?

Anybody else guess ?

nah... intel has yet to bring out a cpu that cant do 4.0...

So 4.0ghz -> 4.3ghz... like how all the other previous flagships have been averaging without the need of insane voltages.

Im not going to say i like that 2000 dollar price tag however... :T

 

[krumme]

The ipc is a letdown and efficiency suffers. Powerconsumption is bad.
Are we sure this a cpu and not an american car?

 

[Carfax83]

I'm guessing these samples must be engineering parts with plenty of bugs, and or the BIOS still needs lots of work. Either way, I'm glad I decided to hold off on Skylake-X this time and wait for the inevitable shrink with better performance, fewer bugs, and hopefully PCIe-4.0

 

[TheF34RChannel]

Memory latency could be a result of the changed cache hierarchy as well. If these benchmarks are legit the new cache structure is worse than the old. Because with Skylake it should have a constant IPC advantage over Broadwell, even if it is tiny. In some of these benchmarks IPC is way below Broadwell, especially for gaming because games tend to be more cache/latency sensitive. I would like to hear some statements from Intel, I wonder what they have to say about this.

Fortunately Intel isn't going to use this structure for the mainstream (definitely not for Coffeelake and Icelake doesn't use it either based on early Geekbench entries).

This has got me worried (gaming is my thing) and baffled with Intel claiming SKL-X is aimed at gamers (probably just advertising, admittedly). If this turns out to be the case it will put me off the platform and put me back on CFL-S with its familiar cache system; I'd be trading in the additional 2C/4T SKL-X offers for the better gaming experience (which would make sense for me personally; I'm after better performance with getting HT back and more cores). Argh I'm pulled in all directions ha ha! Better to wait until both have launched/all the facts are in. The cache restructure for the HEDT platform is interesting to the point of intrigue, imo, and may just well only offers a real advantage in prosumer workloads. You lot make for interesting discussions I'm always eager to read up on, keep it going!

 

[krumme]

Is there a possibility the new cache is made to better reach higher freq at lower process cost? Or in someway its more a cost reducing meassure?

Frankly the new cache looks like a failure regardless of workload but its hardly possible its not intended because the performance cant in anyway be a surprise. Its a planned development and result.

 

[itsmydamnation]

Is there a possibility the new cache is made to better reach higher freq at lower process cost? Or in someway its more a cost reducing meassure?

Frankly the new cache looks like a failure regardless of workload but its hardly possible its not intended because the performance cant in anyway be a surprise. Its a planned development and result.

The cache design is clearly for HPC and large core counts. Having double the vector throughput and more cores means a more then doubling of the cache/interconnect performance is required. Nothing is free and intel have kept the die size the cache/memory system at about the same percentage as broadwell-E while allowing for that big increase in throughput, there is no question that "some" workload might "suffer". The questions are how many workloads are "some" and how much is "suffer".

I wouldn't call it a failure it looks very impressive to me. I just don't think intels direction with Skylake-X is right.

 

[jpiniero]

Is there a possibility the new cache is made to better reach higher freq at lower process cost? Or in someway its more a cost reducing meassure?

Frankly the new cache looks like a failure regardless of workload but its hardly possible its not intended because the performance cant in anyway be a surprise. Its a planned development and result.

The issues might simply be coming up because this is being released in June instead of August. Hexus said the FPS jumped from 48.2 to 74.9 in TW3, and presumably that was from a bios update they got during testing. This sounds like the Ryzen release, doesn't it?

I will say this, Intel's future processors are going to be optimized solely for server workloads in the future. You probably won't see the real impact of this until Sapphire Rapids though. Whether this ends up being bad for gaming remains to be seen.

 

[krumme]

Is lcc cpu size hpc and double vector througput market?
I cant make any sense of this.
What hpc market? Its a cpu build for the future not the past but what is the future here?

 

[moonbogg]

I'm guessing these samples must be engineering parts with plenty of bugs, and or the BIOS still needs lots of work. Either way, I'm glad I decided to hold off on Skylake-X this time and wait for the inevitable shrink with better performance, fewer bugs, and hopefully PCIe-4.0

If that 10 core was soldered and cost $800 would you be interested then?

 

[krumme]

The issues might simply be coming up because this is being released in June instead of August. Hexus said the FPS jumped from 48.2 to 74.9 in TW3, and presumably that was from a bios update they got during testing. This sounds like the Ryzen release, doesn't it?

I will say this, Intel's future processors are going to be optimized solely for server workloads in the future. You probably won't see the real impact of this until Sapphire Rapids though. Whether this ends up being bad for gaming remains to be seen.

I never read and put any emphasis on gaming bm right at release. They are to erradic. That said i dont expect sklx to have remotely the same issues as ryzen. If any.

Surely if the entire cache is tailored eg to push those heavy vectors then it comes comes at a cost. But man its expensive priority even for many professional loads.

 

[jpiniero]

Oh now that explains it; Hexus and Bit-Tech's reviews mention they did not get the board+processor by Intel. So they may not be under NDA. Of course they may be having problems because of it too.

 

[TahoeDust]

Jesus...Jet.com has m.2 960 Evo 500gb for $199.99. Had to order another one. m.2 raid 0 here I come...lol

 

[Carfax83]

Are we sure this a cpu and not an american car?

You're still stuck in the 1990s I see.

 

[Carfax83]

If that 10 core was soldered and cost $800 would you be interested then?

No, probably not. Buying CPUs on the first go around, rather than the re-spin generally isn't a good idea from my experience. The revisions are always going to be better. Take Haswell-E for instance. The DDR4 memory controller wasn't properly optimized for write performance, but this problem was fixed with Broadwell-E. That's part of why I sold my 5930K and bought my 6900K, so I could skip the first Skylake-X parts, and go with the die shrink later on which will be even faster and more efficient.

 

[csbin]

Uh?

I was comparing Broadwell to Skylake here:

Not that if you took the 1800X it would change much: 1614/8x10 = 2017.5 (and I assumed linear scaling to a 10 core Ryzen) then 2017.5/3.6x4 = 2242 (to match 7900X clock).

This nets a 2.5% advantage in total performance but IPC is less for Zen because it has higher SMT scaling (~10%?) in this test, so IPC would be somewhat less, closer to Broadwell/Haswell.

7900X all cores 4.0 =2169
1800X all cores 4.0 =2191 (10 core Ryzen)

 

[pantsaregood]

Two observations here.

First, the two published reviews at the moment lead me to believe that the overclock were solely thermally limited, not limited by instability. Skylake-x cores tend to see reasonable gains all the way up to around 1.45v, so we could reasonably see these CPUs hitting 4.9-5.1 GHz with proper cooling after being delidded.

The latency oddities may be a product of the restructured cache - faster RAM could certainly help with this. Cache misses are an issue, but that could ultimately result in Skylake-X being very bandwidth hungry.

 

[IntelUser2000]

7900X all cores 4.0 =2169
1800X all cores 4.0 =2191 (10 core Ryzen)

6950X score for Bit-Tech is too optimistic in Cinebench R15. Other sites get 1800-1850. Bit-Tech gets 2080.

In comparison, the 7900X scores are consistent for the two sites. One is 2169, one is 2187.

Scores are also strange for VRMark. They both use 32GB memory and GTX 1080.

Hexus gets 10,191 for 7900X and point out the earlier versions with earlier BIOS got 4,015, and 6950X gets 11,422. Bit-Tech gets 9,678 for 7900X and 6950X gets 11,782.

I'd look at more review sites before drawing my conclusion.

 

[TheGiant]

Oh now that explains it; Hexus and Bit-Tech's reviews mention they did not get the board+processor by Intel. So they may not be under NDA. Of course they may be having problems because of it too.

Nice find. Missed that one.

I hope your are right. My decision was to buy the 7820X but ..lets wait for official benches

 

[TheF34RChannel]

Nice find. Missed that one.

I hope your are right. My decision was to buy the 7820X but ..lets wait for official benches

Smart decision!