Intel Skylake / Kaby Lake

[itsmydamnation]

Tpu
"At its top overclock of 4.50 GHz, the i7-6950X achieved a Cinebench R15 score of 2327 points"

https://www.techpowerup.com/222230/intel-core-i7-6950x-tested-against-i7-5960x

What is happening with that new cache?

It's there to facilitate core count increase while at the same time reducing pressure on the L3 from high read and write requirements caused by AVX-512. I pointed it out ages ago in this thread that the cache changes didn't really look like it would make that much of a difference to absolute performance.

But people were convinced there are all these workloads that have really poor ILP but enough predictability and reused enough to be in the cache system, but not enough to be in the L1 or L2 (256kb size).

 

[TahoeDust]

http://www.tweaktown.com/news/57814/skills-16gb-ddr4-ram-hits-4800mhz-intel-core/index.html

G.Skill DDR4 4800 is known to work on ASRock X299 OC Formula. It seems like these new boards are ultra stable at least with Core -7 7740K.

Fuuuuuu....I just ordered 32gb of Corsair 3200 for my x299 build.

 

[TheF34RChannel]

Fuuuuuu....I just ordered 32gb of Corsair 3200 for my x299 build.

Send it back G.Skill offers both the highest frequency and tightest timings. Klevv has got a nice white LED kit as well with the triangles but only DDR3 darn it. Going for white LED Corsair vengeance myself, 16GB. I like the space ship look of the lights. Like G.Skill trident Z better but the LEDs are awful imo.

 

[blue11]

https://www.youtube.com/watch?v=kyz2gkyIoXI

7900X@4.5
cinebench R15
2419 multi
198 single
temps during cinebench - 85c on air tower cooler (tragic).

Any heavy loads would send that chip into thermal throttling. Not sure how much water would help since the heat simply can't escape the die fast enough due to crappy TIM.

Rather than thermal throttling, it would be better to say that they would run at base frequency on some workloads. Those loads would mainly involve AVX, and it is rather silly to quibble over a few MHz when you are doing twice (or 4x) the work per cycle. The base frequency (3.3 GHz) is essentially the same as the all-core turbo of the outgoing 6950X (3.4 GHz), so there should not be any backsliding in performance.

Is it possible to set different overclocking targets based on the active core count? Now that we are supposedly getting unlocked 18-core CPUs, per-bin overclocking is going to be much more effective (and energy-efficient) than trying to put every core at 5 GHz. If the 7900K and higher SKUs can be set to 4-core turbo at 5 GHz, they should be able to match the 7700K in desktop performance.

 

[moonbogg]

Tpu
"At its top overclock of 4.50 GHz, the i7-6950X achieved a Cinebench R15 score of 2327 points"

https://www.techpowerup.com/222230/intel-core-i7-6950x-tested-against-i7-5960x

What is happening with that new cache?
This is a 4% perf uplift at same freq. But i guess that is expected or what?

Skylake has always been about 5% max IPC improvement over Broadwell at same clocks. That's why X299 was hinging so critically on overclock performance. All of the 14nm CPU's are within spitting distance of each other. That's why people are obsessed with Ryzen clock speed, Skylake-X overclocking ability, Broadwell's sub ideal overclocking ability. Its because overclocking is the only real thing that sets any of these architectures apart from each other in a meaningful way.
If Ryzen overclocked to 4.5, skylake X would be an overpriced dud. Its that close of a race. That's why x299 TIM is a travesty. The 6 and 8 core versions of SkyX could have been phenomenal!!! But nope. Big nope.

 

[jpiniero]

Pretty disappointing scores. That's not even faster than mainstream Skylake. Might even be a regression.

Is there any information on how this was done (what cooling, anything)? Color me intrigued!

Intel could have done 4 Ghz FCT on Broadwell-E for sure. Maybe even Haswell-E. The power draw only starts to really get cray cray once you get past that point.

 

[blue11]

Pretty disappointing scores. That's not even faster than mainstream Skylake. Might even be a regression.

HEDT/server always has lower IPC than client. The uncore has to satisfy far more cores with nearly the same bandwidth, the L3 latency is twice as high, and the memory latency is 10-20 ns higher (because of L3). I have no idea how the whole forum managed to delude themselves into thinking that Skylake-X/SP would somehow have 10% IPC over client.

By the way, "IPC" is not some multiplier that you can use to claim some processor is 5% faster than another. Many client workloads are already running at their ILP limit, and no processor design is going to extract any more IPC when latency-bound. The main bottleneck Skylake alleviates from Haswell/Broadwell is L2/L3 cache bandwidth, so you should see sizable gains on image manipulation and scientific applications.

 

[Shivansps]

i whould have expected that moving from 4-way L2 to 16-way L2 whould bring in some improvements.

 

[Rngwn]

https://www.youtube.com/watch?v=kyz2gkyIoXI

7900X@4.5
cinebench R15
2419 multi
198 single
temps during cinebench - 85c on air tower cooler (tragic).

Any heavy loads would send that chip into thermal throttling. Not sure how much water would help since the heat simply can't escape the die fast enough due to crappy TIM.

On the same clip, what is the 2442 cb score doing here with the 7900X on 2:27? Would that suggest this is the pure 7900X performance when the thermals has yet to kick in during the first run?

 

[imported_ats]

As I understand it, due to mechanical constraints, among other reasons, laptop CPUs do not have a heatspreader attached. The cooling systems in laptops, are direct-to-die, on both the CPU and GPU. (Usually heatpipes, linked to a heatsink and fan elsewhere in the chassis.)

IHS were primarily added in for the retail/DIY market and other markets that do human based assembly in order to protect the CPU die during assembly from cracking and secondarily because the heat spreader functionality on heat sinks is generally uneven in quality. For the laptop market, the CPUs are BGAs directly attached to the boards by robot and often the whole thermal solution is also attached via robot and the heatsink solutions are all OEM designed for the application makes the heat spreader functionality pretty robust, therefore a IHS would largely be redundant and unnecessary.

 

[moonbogg]

On the same clip, what is the 2442 cb score doing here with the 7900X on 2:27? Would that suggest this is the pure 7900X performance when the thermals has yet to kick in during the first run?

I don't know. Wouldn't be surprised if the thing throttled just from doing a couple cinebench runs.

 

[IntelUser2000]

It's there to facilitate core count increase while at the same time reducing pressure on the L3 from high read and write requirements caused by AVX-512. I pointed it out ages ago in this thread that the cache changes didn't really look like it would make that much of a difference to absolute performance.

True.

Cinebench doesn't care about cache sizes though.

The better "IPC" comparison is with single thread:

6950X 4.5GHz: 184
http://www.pcworld.com/article/3075...-core-enthusiast-cpu-is-a-monster.html?page=3

7900X 4.5GHz: 198

That's 7.6%. Ok, that's probably just the core changes but not so bad as it looked on MT.

I'd think you'd see 3-5% change in average, nowhere near the super-optimistic 8-13% Anandtech article was predicting. This gain you'll get is probably almost as big as what you'll get with Icelake on a new uarch so take what you can get.

 

[csbin]

True.

Cinebench doesn't care about cache sizes though.

The better "IPC" comparison is with single thread:

6950X 4.5GHz: 184
http://www.pcworld.com/article/3075...-core-enthusiast-cpu-is-a-monster.html?page=3

7900X 4.5GHz: 198

That's 7.6%. Ok, that's probably just the core changes but not so bad as it looked on MT.

I'd think you'd see 3-5% change in average, nowhere near the super-optimistic 8-13% Anandtech article was predicting. This gain you'll get is probably almost as big as what you'll get with Icelake on a new uarch so take what you can get.

6950X@4.5+DDR4 3200

ST 195
MT 2390

 

[LTC8K6]

6950X@4.5+DDR4 3200

ST 195
MT 2390

Presumably the 7900X will be able to use much faster memory.

http://www.tweaktown.com/news/57814/skills-16gb-ddr4-ram-hits-4800mhz-intel-core/index.html

 

[Ajay]

I think I'll hold out for reviews to get an overall sense of SKX's capabilities. These leaks are concerning, but a couple of data points are useless. Silicon Lottery is contemplating delliding all the SKX CPUs it sells and are suggesting that these suckers are going to run hot (overclocked). Glad companies like this exist so I don't potentially destroy a brand new CPU. The downside is that that's going to push a
7800X cost almost $450. I can get an HWE 5820K for $300 right now plus 30 bucks off with a motherboard bundle!

http://www.overclock.net/t/1631319/skylake-x-binning

 

[csbin]

Presumably the 7900X will be able to use much faster memory.

http://www.tweaktown.com/news/57814/skills-16gb-ddr4-ram-hits-4800mhz-intel-core/index.html

i9 7900X@5755MHz + DDR4 3820mhz CL12 (3181 cb ）World Records
http://hwbot.org/submission/3563150_elmor_cinebench___r15_core_i9_7900x_3181_cb


i7 6950X@5323MHz + DDR4 3550mhz CL11（2878 cb ）World Records
http://hwbot.org/submission/3468936_xtreme_addict_cinebench___r15_core_i7_6950x_2878_cb/

 

[LTC8K6]

For CPU-Z, do we know if they were all running 1.79.1 version with the new version 17.01 benchmark?

 

[LTC8K6]

i9 7900X@5755MHz + DDR4 3820mhz CL12 (3181 cb ）World Records
http://hwbot.org/submission/3563150_elmor_cinebench___r15_core_i9_7900x_3181_cb


i7 6950X@5323MHz + DDR4 3550mhz CL11（2878 cb ）World Records
http://hwbot.org/submission/3468936_xtreme_addict_cinebench___r15_core_i7_6950x_2878_cb/

Well, don't much care about LN2 benches.

 

[wildhorse2k]

I think I'll hold out for reviews to get an overall sense of SKX's capabilities. These leaks are concerning, but a couple of data points are useless. Silicon Lottery is contemplating delliding all the SKX CPUs it sells and are suggesting that these suckers are going to run hot (overclocked). Glad companies like this exist so I don't potentially destroy a brand new CPU. The downside is that that's going to push a
7800X cost almost $450. I can get an HWE 5820K for $300 right now plus 30 bucks off with a motherboard bundle!

http://www.overclock.net/t/1631319/skylake-x-binning

https://siliconlottery.com/ is overhyped. There is very little benefit from buying an OC pretested CPU from them for average customer. Performace gain from those extra 200Mhz you could gain is minimal. You have to pay more and don't get warranty. You could buy from any seller, be lucky and get exactly the same OC. Their cheaper CPUs have very high OC probability so its totally pointless to buy them. Buying these OC pretested CPUs only makes sense for some hard core overclockers trying to beat world records.

Intel pricing is quite aggressive for the top CPUs. It may be simply better to pay $200 more to Intel to get 2 more cores and get much better performance than from those extra 200Mhz from silicon lottery.

 

[TheF34RChannel]

https://siliconlottery.com/ is overhyped. There is very little benefit from buying an OC pretested CPU from them for average customer. Performace gain from those extra 200Mhz you could gain is minimal. You have to pay more and don't get warranty. You could buy from any seller, be lucky and get exactly the same OC. Their cheaper CPUs have very high OC probability so its totally pointless to buy them. Buying these OC pretested CPUs only makes sense for some hard core overclockers trying to beat world records.

Intel pricing is quite aggressive for the top CPUs. It may be simply better to pay $200 more to Intel to get 2 more cores and get much better performance than from those extra 200Mhz from silicon lottery.

Definitely the better option!

 

[MrTeal]

https://siliconlottery.com/ is overhyped. There is very little benefit from buying an OC pretested CPU from them for average customer. Performace gain from those extra 200Mhz you could gain is minimal. You have to pay more and don't get warranty. You could buy from any seller, be lucky and get exactly the same OC. Their cheaper CPUs have very high OC probability so its totally pointless to buy them. Buying these OC pretested CPUs only makes sense for some hard core overclockers trying to beat world records.

Intel pricing is quite aggressive for the top CPUs. It may be simply better to pay $200 more to Intel to get 2 more cores and get much better performance than from those extra 200Mhz from silicon lottery.

How is that at all comparable? Silicon Lottery charges $50 for a delid. Even at $100 for Skylake-X, that's good value. If you did it yourself, not only would you almost certainly need to buy a specialized tool to do it, but you still seriously risk destroying a $1000+ processor. Warranty is a non-issue since you loose it when you delid anyway; at least with SL you would know the chip post-delidding works.

 

[TheF34RChannel]

How is that at all comparable? Silicon Lottery charges $50 for a delid. Even at $100 for Skylake-X, that's good value. If you did it yourself, not only would you almost certainly need to buy a specialized tool to do it, but you still seriously risk destroying a $1000+ processor. Warranty is a non-issue since you loose it when you delid anyway; at least with SL you would know the chip post-delidding works.

Your arguments are only valid if you had planned on delidding anyway. I wouldn't buy it for the OC since the same chip OCs differently on different (individual?) boards.

 

[wildhorse2k]

How is that at all comparable? Silicon Lottery charges $50 for a delid. Even at $100 for Skylake-X, that's good value. If you did it yourself, not only would you almost certainly need to buy a specialized tool to do it, but you still seriously risk destroying a $1000+ processor. Warranty is a non-issue since you loose it when you delid anyway; at least with SL you would know the chip post-delidding works.

Delid is not something I was discussing in my post. So you are not actually reacting to my post at all. When you buy OC pretested CPU from them you don't get automatic delid.

Delidding can be purchased with this CPU. Silicon Lottery recommends having Kaby Lake processors delidded for the best overclocking experience.

My argument is if you are willing to pay extra and are an average customer, you perhaps should be buying a more core CPU from Intel or even get Threadripper for much more cores.

 

[blue11]

Delid is not something I was discussing in my post. So you are not actually reacting to my post at all. When you buy OC pretested CPU from them you don't get automatic delid.

Silicon Lottery claim to be bundling delid and a 1-year aftermarket warranty on their Skylake-X offerings.

http://www.overclock.net/t/1631319/skylake-x-binning

 

[ManyThreads]

Does anyone know when the review/press embargo is lifted for Skylake-X? Or in other words, when we should see benchmarks and reviews?

All I really care about is how it performs relative to Ryzen and BW-E in the relevant benchmarks, and what the OC potential is.