Intel Skylake / Kaby Lake

[TheF34RChannel]

Very crude math here:

I'll assume there are no hidden secrets and thus CFL is basically SL and KL on a more refined 14 nm process. I'll also assume that Intel is correct that the 14++ process has 23% more efficient transistors than the 14+ process. Finally, I'll also assume the rumors that CFL is 6C/12T with a 95W TDP (a slight increase over SL and KL) are correct. Based off of that information alone, and using the standard approximate power formula for a CPU core (P = f*C*V^2) you end up with this:

• A 6 core CFL with hyperthreading at 95 W could reach 3.98 GHz all core turbo at stock. Round that up to 4.0 GHz.

Unless Intel has an ace up their sleeve, then 6 core CFL all-core turbo will be just shy of the SL all-core turbo. I honestly think it'll be slightly less than 4.0 GHz (say 3.9 GHz) turbo unless Intel has something else that is not yet known.

The previous leak already showed the part @ 4.2GHz turbo, so... (see below)

Is the die bigger? Would that give us better cooling, or worse? Or is it a wash with more cores and a bigger die such that we can just leave it out of our guesses?
Intel could also solder the IHS?

Indeed, there are still many uncertainties that would affect the final frequencies.

[Bold] That cruel joke would make my day!

Spoiler

 

[dullard]

Is the die bigger? Would that give us better cooling, or worse? Or is it a wash with more cores and a bigger die such that we can just leave it out of our guesses?
Intel could also solder the IHS?

A bigger die and/or solder would give us better cooling. But that affects temperature and not stock frequency. Yes, they are related concepts if you overclock (where you allow the power to increase far above the rated TDP power). But as long as Intel is limiting the stock processor to an average power cap, then a larger die or solder won't affect the stock frequencies that Intel is guaranteeing.

 

[dullard]

The previous leak already showed the part @ 4.2GHz turbo, so... (see below)

Thanks, I had not seen that particular leak. Intel's 14nm++ may be slightly better than the 23% to 24% more efficient transistors than they claimed this spring. 27% more efficient would hit 4.2 GHz all core turbo. As would slightly reducing yield on the top chip to just the best performers. There are several variables to play with outside of simple math.

 

[TheF34RChannel]

Thanks, I had not seen that particular leak. Intel's 14nm++ may be slightly better than the 23% to 24% more efficient transistors than they claimed this spring. 27% more efficient would hit 4.2 GHz all core turbo. As would slightly reducing yield on the top chip to just the best performers. There are several variables to play with outside of simple math.

Exactly, and these are not necessarily the final clocks, being an ES (could go both ways, however usually these leaks this close upon release show either the retail or slightly below retail clocks if memory serves me well).

 

[AdamK47]

I've been pretty impressed in how far Kaby Lake X and X299 can push dual channel memory. I bought 16GB of TridenZ DDR4 4266 memory not really knowing if it would run it well or not. It seems to run without a hitch with the XMP profile at it's rated 4266 speed. The memory is currently at 4340 running some stress testing. I'll see if I can push it further after that.

 

[frozentundra123456]

Intel needs an improvement in architecture and lower prices if they want to remain competitive against the AMD roadmap. When is the new arch supposed to be released ? I don't care about their confusing names.

Not really. Intel still has an ipc advantage in most apps and at least a 20% or more clockspeed advantage. So they certainly are more than "competitive" in performance. In all the Ryzen hype, people seem to think they have taken the performance lead, but they really havent, only the price/performance lead if you can use a lot of cores.

 

[Lodix]

Not really. Intel still has an ipc advantage in most apps and at least a 20% or more clockspeed advantage. So they certainly are more than "competitive" in performance. In all the Ryzen hype, people seem to think they have taken the performance lead, but they really havent, only the price/performance lead if you can use a lot of cores.

I don't disagree with anything you said. But Zen being a totally new architecture has a lot of room of improvements and AMD is releasing in the following years new revisions of the architecture. I would be surprised if by Zen2 they haven't surpassed current Intel core IPC ( the different right now is around 10% as I have seen, could be wrong). And they will use higher performance nodes with 7nm targeting 5GHz frequencies with Zen 2 in early 2019. So yes, I think intel will lose their perfomance lead as they have alredy lost their process lead if they don't come up with something new soon. I asked before when are they supposed to do the next "tock" ?

 

[Markfw]

Not really. Intel still has an ipc advantage in most apps and at least a 20% or more clockspeed advantage. So they certainly are more than "competitive" in performance. In all the Ryzen hype, people seem to think they have taken the performance lead, but they really havent, only the price/performance lead if you can use a lot of cores.

Lets see, they have the price/performance lead and in the case of EPYC, the performance lead, and probably threadripper, and they use 2/3rds of the power, so who cares that its more cores ?

 

[ManyThreads]

Just got my system, running 7820X and Asus TUF MK1. I'm having problems with my Samsung 960 pro though, hopefully someone can help. I cannot get it anywhere near it's advertised speeds. Best I can do is about 2300 MB Read/ 1700 MB write. I changed the Mircosoft NVME driver to the Samsung one and there was no improvement whatsoever. Am I missing something? I'm using Samsung Magician software to run a performance benchmark. Tried it in Crystaldisk too and it's still super slow, over 1000MB/s slower in 2-3 areas compared to everyone else who I see screenshots from.

 

[frozentundra123456]

Lets see, they have the price/performance lead and in the case of EPYC, the performance lead, and probably threadripper, and they use 2/3rds of the power, so who cares that its more cores ?[/QUOTE}

People who like the best single core performance????? (Rhetorical question BTW) I also thought it was not allowed to speculate about unproven performance, even about an intel chip in an intel thread, much less an AMD chip in an Intel thread.

 

[Rngwn]

Anybody know if the Skylake Xeon is soldered, not toothedpasted to the die just like SKL-X?

 

[Markfw]

People who like the best single core performance????? (Rhetorical question BTW) I also thought it was not allowed to speculate about unproven performance, even about an intel chip in an intel thread, much less an AMD chip in an Intel thread.

See the OTHER threads with BENCHMARKS on TR and EPYC.

 

[TheF34RChannel]

Anybody know if the Skylake Xeon is soldered, not toothedpasted to the die just like SKL-X?

Good question! I haven't heard anything about it yet and I'm curious too.

 

[moonbogg]

I thought I saw a chart that showed overclocking the mesh improved gaming performance. Anyone know what the current consensus is on this? Here are my current beliefs about the status of Skylake-X gaming performance. Check them and let me know where I have errors, and if possible, correct them please (preferably with graphs if possible, but if not no biggy)

Beliefs:
1: Skylake-X performs worse than Broadwell-E in games at the same clocks due to mesh being slow and maybe due to cache structure playing a role as well.
2: Game performance improves when overlcocking the mesh, but performance is still slower than expected.
3: Skylake-X is slow in gaming for similar reasons as Ryzen's strange gaming behavior; its caused by the new way both companies connect their cores.

I am confused about the gaming performance status of skylake-X.

 

[wildhorse2k]

I would say the increased cache latencies, slow L3 cache, L3 cache being victim is what hurts gaming performance. L3 cache latency can be partially fixed with mesh overclock as you say. Usage of fast low latency memory helps as well (3600 is best, 3200 second best). Higher skylake-x frequencies also help. So in the end it shouldn't be worse than broadwell-e in games.

VRM issues are fixed by using a fan or alternative VRM heatsink. Phantom throttling can be fixed by increasing some limits in BIOS. The only issue that remains is really the poor M.2 performance.

 

[RichUK]

I would say the increased cache latencies, slow L3 cache, L3 cache being victim is what hurts gaming performance. L3 cache latency can be partially fixed with mesh overclock as you say. Usage of fast low latency memory helps as well (3600 is best, 3200 second best). Higher skylake-x frequencies also help. So in the end it shouldn't be worse than broadwell-e in games.

VRM issues are fixed by using a fan or alternative VRM heatsink. Phantom throttling can be fixed by increasing some limits in BIOS. The only issue that remains is really the poor M.2 performance.

Poor M.2 performance?

 

[wildhorse2k]

Poor M.2 performance?

Yes its been reported by couple of people. I think also one review found it too, but I don't remember which one. Most reviews didn't test m.2 performance. You would have to go through review links posted in this topic.

 

[AdamK47]

Tested my 512GB 960 Pro using AS-SSD. Looks okay.

Also tested my four 2TB 850 EVOs in RAID-0 (8TB).

The X299 PCH seems to do a great job with SSDs.

 

[JoeRambo]

The X299 PCH seems to do a great job with SSDs.

But Your CPU is Kabylake 7740X and not Skylake-X. Problems could be legit and connected to power management or NVM drivers not liking Skylake-X cache changes.
Setting power to high performance, and minimum Processor state to 100% and rerunning the M.2 tests for Skylake-X would remove Windows power mgmt factor.

 

[lobz]

The previous leak already showed the part @ 4.2GHz turbo, so... (see below)

True, but noone knows how much power did that chip consume, and if it's within launchable levels.


Sent from my VTR-L09 using Tapatalk

 

[AdamK47]

But Your CPU is Kabylake 7740X and not Skylake-X. Problems could be legit and connected to power management or NVM drivers not liking Skylake-X cache changes.
Setting power to high performance, and minimum Processor state to 100% and rerunning the M.2 tests for Skylake-X would remove Windows power mgmt factor.

Another reason to like my CPU choice.

 

[wildhorse2k]

Tested my 512GB 960 Pro using AS-SSD. Looks okay..

Here are links to show I didn't make it up:

https://www.bit-tech.net/reviews/tech/asus-rog-strix-x299-e-gaming-review/11/

Asus ROG Strix X299-E

M.2 SSD speed was a few hundred megabytes per second lower than we'd expect from the Samsung 960 Evo. We've had to manually tweak settings in the EFI for Asus boards before, but this time we could find nothing specifically relating to the M.2 port. We tried using Samsung's NVMe driver, a 960 Pro, and both ports, but we were still met with sub-par read speeds. Asus has told us it's looking into the issue

Seq read:
X299 Strix Gaming - 2941MB/s
Z270 Maximus ix hero - 3396MB/s

http://www.tweaktown.com/reviews/8260/gigabyte-x299-aorus-gaming-9-motherboard-review/index9.html

They found 4K read speeds to be lower than expected.

For reference, compare read speeds with http://www.tweaktown.com/reviews/7777/asrock-x99-taichi-intel-motherboard-review/index9.html as they test the same SSD.

X299:
Asus Prime - 49MB/s, 2000MB/s
Aorus 9 - 43MB/s , 2000MB/s

X99:
Taichi - 57MB/s, 2300MB/s
Rampage 10 Extreme - 56MB/s, 2300MB/s

 

[TheF34RChannel]

True, but noone knows how much power did that chip consume, and if it's within launchable levels.


Sent from my VTR-L09 using Tapatalk

I would be very surprised if that frequency put it outside tolerable levels.

 

[lobz]

I would be very surprised if that frequency put it outside tolerable levels.

But it could still be much higher, than whta the current i7 lineup has. I just wanted to point out, that many people (not you) act with coffee lake as it was also a node shrink and I have no idea why.

Sent from my VTR-L09 using Tapatalk

 

[TheF34RChannel]

But it could still be much higher, than whta the current i7 lineup has. I just wanted to point out, that many people (not you) act with coffee lake as it was also a node shrink and I have no idea why.

Sent from my VTR-L09 using Tapatalk

Ah got it. Yeah, it's a mere optimization.