Intel Skylake / Kaby Lake

[formulav8]

Could someone please explain to me this CPU soldering, and why it's so bad if Skylake-X doesn't have it?

Search for the bad problems some are having with very high cpu temps overclocking their 7700k cpu. Many are delidding and getting huge improvements in temps. Intel cheaped out on the paste they used between the lid and die. Sandy Bridge was the last I recall from Intel doing the lid the better way by soldering instead of cheapo paste. Devils Canyon seems to have been decent paste but they may have went back to the cheaper stuff again.

AMD solders their Ryzen cpu's which is the more costly, but better quality way to join the lid/heatspreader and cpu core.

 

[tamz_msc]

And? It's just one specific workload. Can't make up BS claim about AMD CPU's somehow avoiding NUMA specific issues just by enabling Interleaving.

The biggest NUMA specific issue is memory latency, and in the most extreme example which is highly sensitive to latency, AMD takes a lesser hit in performance when number of threads exceed the number of physical cores by interleaving, which is a way around NUMA to be clear.

There are lots of applications where raw FP throughput does not matter so much as memory latency.

 

[Zucker2k]

Looks like Christmas came early this year. Thank you AMD!

Now, how soon can I lay my filthy paws on a 7820X?

 

[ManyThreads]

That pricing wouldnot make much sense. I mean the difference between 8 and 10 core is way, way too large.

I'm in Canada but the 8-Core 6900K is $1500 and The 10-core 6950K is $2400. Almost $1000 for 2 more cores. Now, with the 8-core getting dropped to 28 PCI lanes I could see them making the price gap even larger while they position the 8-core Skylake-X SKU to compete more closely with the Ryzen 1800X.

 

[Zucker2k]

Yeps AMD competing helps even those who only support Intel as well. A 3rd player would make it even better.

So we can all agree the ONLY reason why Intel kept us starved for moar cores was because AMD failed to offer any tangible competition for a whole decade? Mmkay!

 

[formulav8]

Also looks like why Intel could be offering better pricing than expected.

 

[RichUK]

That pricing wouldnot make much sense. I mean the difference between 8 and 10 core is way, way too large.

Not really, take a look at the price difference between the current 8 and 10 core (6900, 6950) - the gap is even larger.

 

[JoeRambo]

The biggest NUMA specific issue is memory latency, and in the most extreme example which is highly sensitive to latency, AMD takes a lesser hit in performance when number of threads exceed the number of physical cores by interleaving, which is a way around NUMA to be clear.

How can one even arrive to conclusion that make no sense? Without perf counters it is pretty much impossible to say exactly what was happening in that case and we can only speculate. Even Anandtech guys noticed that in Cinebench workload performance improved by just 5% by enabling Interleave.

The lack of scaling was classic result of 48 threads jumping on one node and its mem controller and swamping it with reads/writes since all memory was allocated in it and due to lack of proper multisocket MP support it was overwhelmed ( something that does not happen on Intel 4S CPUs of that timeframe).


Still how one can arrive to conclusion that Zen is less NUMA scaling problem prone ???

 

[tamz_msc]

How can one even arrive to conclusion that make no sense? Without perf counters it is pretty much impossible to say exactly what was happening in that case and we can only speculate. Even Anandtech guys noticed that in Cinebench workload performance improved by just 5% by enabling Interleave.

Cinebench historically isn't sensitive to memory, except with Ryzen Cinebench R15 which shows a slight improvement with faster memory. However one look at any algorithm for a PDE solver shows that it is sensitive to memory, and I bring these benchmarks up again and again because CFD and other similar stuff are among the the most relevant applications for testing how well your NUMA implementation works.

Still how one can arrive to conclusion that Zen is less NUMA scaling problem prone ???

How can one arrive to the conclusion that Naples will have the same severity of latency issues that plagued NUMA architectures in the past, just because it is a 4-die MCM?

 

[swilli89]

So we can all agree the ONLY reason why Intel kept us starved for moar cores was because AMD failed to offer any tangible competition for a whole decade? Mmkay!

Right, we can all agree Intel has been artificially milking their high end socket for years now, yes.

 

[R0H1T]

So we can all agree the ONLY reason why Intel kept us starved for moar cores was because AMD failed to offer any tangible competition for a whole decade? Mmkay!

Alternatively - Intel has been greedy as f*** all these years & enthusiasts who upgrade regularly have been fueling this greed incessantly.

 

[lolfail9001]

Alternatively - Intel has been greedy as f*** & all the enthusiasts who upgrade regularly have been fueling this greed incessantly.

Intel has every right to be greedy as much as it wants to be. And enthusiasts have every right to fuel that greed if they want to.

So, yeah, AMD to blame for failing to compete.

 

[R0H1T]

Intel has every right to be greedy as much as it wants to be. And enthusiasts have every right to fuel that greed if they want to.

So, yeah, AMD to blame for failing to compete.

Yeah sure the public gets what they paid for, like the recent elections even some outside the US, doesn't mean they are right though.

 

[ehume]

Rats. 6-core Coffee Lake is 95W. The higher-watt i9 will be a 7xxx part with no inbuilt GPU -- will require a dGPU, right? If true, this is bad news for those of us that review heatsinks and AIO's.

 

[jpiniero]

That Intel partner slide suggests that there might be some Coffee Lake locked models released in August too, plus perhaps a refresh of the entire lineup, including K models in January.

 

[Bouowmx]

Rats. 6-core Coffee Lake is 95W. The higher-watt i9 will be a 7xxx part with no inbuilt GPU -- will require a dGPU, right? If true, this is bad news for those of us that review heatsinks and AIO's.

Nothing is radically different from the previous generation: mainstream processor with integrated GPU ~ 95 W (using thermal interface compound between die and heat spreader), and high-end without integrated GPU > 140 W (using solder).

EDIT: Well, I got pranked by Intel xdd. Skylake-X does not use solder.

 

[wildhorse2k]

Search for the bad problems some are having with very high cpu temps overclocking their 7700k cpu. Many are delidding and getting huge improvements in temps. Intel cheaped out on the paste they used between the lid and die. Sandy Bridge was the last I recall from Intel doing the lid the better way by soldering instead of cheapo paste. Devils Canyon seems to have been decent paste but they may have went back to the cheaper stuff again.

AMD solders their Ryzen cpu's which is the more costly, but better quality way to join the lid/heatspreader and cpu core.

My speculation is that it is way better not to solder as you can delid and use the best suitable paste. You could even have CPU cooler directly on the core. It was done like this in the past and little CPU damage occurred (although coolers were smaller).

If you use water cooling, CPU water blocks are tiny and light, should pose no risk to damaging CPU once delidded properly. It makes no sense to keep the heat spreader on.

 

[coercitiv]

That Intel partner slide suggests that there might be some Coffee Lake locked models released in August too, plus perhaps a refresh of the entire lineup, including K models in January.

The August launch definitely includes i7 and i5 locked models by the looks of that slide, maybe even i3. Meanwhile, only Z370 chipset will be available in 2017. Do we have any clear info on CFL lack of support for 200 series chipsets? I find it very hard to believe they would launch cheaper mainstream parts without a proper chipset to match.

 

[ImSpartacus]

Silly question:

What are the anticipated core counts for Skylake-EP's LCC, MCC & HCC (XCC?) die?

I know Broadwell-EP appears to be sporting up to 10C, 15C and 24C on their respective die (though I don't think those all get used for yield purposes.

 

[ehume]

Nothing is radically different from the previous generation: mainstream processor with integrated GPU ~ 95 W (using thermal interface compound between die and heat spreader), and high-end without integrated GPU > 140 W (using solder).

"Nothing is radically different" -- and this is bad news. We can't really test the noise a heatsink makes unless we use a passive GPU. And any GPU -- especially a passive one -- will add heat to the system that will drift up to interfere with a heatsink -- unless we test the system in an open setting and blow air past it to make sure a GPU's heat does not make it to the heatsink. But then if we blow air in, that will affect a heatsink. The only workaround will be to overclock the system and hope that the CPU can support the higher wattages required.

I was hoping to avoid all that by using a new CPU with both iGPU and 140W+ TDP. Looks like that won't happen.

 

[coercitiv]

My speculation is that it is way better not to solder

Come on, delid that $1700 monster, only non-enthusiast pussies hang on to their warranty!

 

[Sweepr]

Silly question:

What are the anticipated core counts for Skylake-EP's LCC, MCC & HCC (XCC?) die?

I know Broadwell-EP appears to be sporting up to 10C, 15C and 24C on their respective die (though I don't think those all get used for yield purposes.

LCC = 10C, MCC = 18C, HCC = 28C

BTW with Core i9-7800X costing only ~7% more than Core i7-7700K (pre-launch listing, MSRP could be closer) we can hopefully put the talk about 6C Coffee Lake-S being much more expensive than its predecessor to rest.

 

[coercitiv]

BTW with Core i9-7800X costing only ~7% more than Core i7-7700K (pre-launch listing, MSRP could be closer) we can hopefully put the talk about 6C Coffee Lake-S being much more expensive than its predecessor to rest.

I can't wait for that 7740K that's going to clock higher and cost ~15% less than 7700K! Wait, something's wrong...

 

[wildhorse2k]

A load of X299 motherboards: http://wccftech.com/intel-x299-motherboards-pictured-asus-asrock-gigabyte-biostar-evga-msi/

The ASRock OC Formula is interesting - it only has 4 DIMM slots.

Some very interesting ASRock motherboards there.

My favorite will probably be ASRock X299 PROFESSIONAL GAMING i9 (more feature loaded Taichi version). ASUS ROG Rampage VI Extreme or Gigabyte X299 AORUS Gaming 9 would probably be an option too. OC Formula has too few DIMM slots, although I like the number of PCIe slots.

Interesting is claim for DDR4 4400+ support. This could mean G.Skill DDR4 4000 could actually work fine. Poor Ryzens lucky if 3466 works with new BIOSes.

 

[Sweepr]

I can't wait for that 7740K that's going to clock higher and cost ~15% less than 7700K! Wait, something's wrong...

...or perhaps, there isn't? You'll see tomorrow.

Ps: They really need to get Coffee Lake-S out soon.

Edit: Sorry, missed the sarcasm.