Intel Skylake / Kaby Lake

Page 555 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

jpiniero

Lifer
Oct 1, 2010
14,509
5,159
136
Intel should not shove aside the 1151 owners. KL is only a few months old and what upgrade will there be beyond 7700k?

You really think people would want to drop another $350+ on Coffee Lake that quickly? That's why the upgrade path is overrated, by the time you'd want to upgrade you'd probably want the new features anyway.
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
1,540
136
You really think people would want to drop another $350+ on Coffee Lake that quickly? That's why the upgrade path is overrated, by the time you'd want to upgrade you'd probably want the new features anyway.

I would bet 99%+ of motherboards only ever have one CPU in them. It's one of those complaints to have something to complain about, and will be utterly forgotten a couple of months after launch of the new socket standard.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,708
3,554
136
Why don't we equalize clocks between Ryzen and SKL-X at 4.6Ghz on water cooling then? It won't be a fair fight, right? I hope you see that even the decision to run clocks at 3Ghz is an indirect recognition of the limitations of one design compared to the other?
Performance does not always scale linearly with clock speeds - if it were then we would have no problem running them at 4.6GHz, as long as both can attain that speed. But one of them doesn't, hence comparing at 3 or 3.5GHz is perfectly fine because server CPUs have been running at those frequencies for a long time. IPC is loosely used to talk about intra and inter-generational architectural comparisons and these will happen whenever a new a new architecture is released.
 
  • Like
Reactions: Drazick

gx_saurav

Senior member
Dec 5, 2012
247
61
101
about.me
Desire to upgrade from Core i7 7700k to CFL will always be there. Need to upgrade isn't for next few years.

Games are getting multi thread aware but still only few games use all 8 threads of a Core i7 7700k. That CPU is plenty powerful to hold for 4, 5 years considering games are GPU limited and not CPU.

Buying a 4C-4T CPU today to keep for next few years of gamimg isn't a wise decision. Buying a 4C-8T isn't a wise decision now that CFL is so close to release. Better to wait and get a Coffee Lake CPU in few months unless you have no PC today and must buy.
 

tamz_msc

Diamond Member
Jan 5, 2017
3,708
3,554
136
Let's say this. It's not zero as you imply. The usage scenarios that benefit are probably a lot less than it used to with AVX and AVX2, but it exists.

If you look at the Xeon Scalable news items from Intel, they point out the general increase is 1.4-1.6x. There are few that reaches in the 1.7-2x range. Those if you look at the presentation are due to AVX-512 optimizations. They were showing DL/AI workloads doing 2-2.2x due to architecture + AVX-512. And then Tencent with video stitching performance improved by 67% using AVX-512 optimizations. Another company was talking about Data Analytics.

So the Flops increase in processors that are limited by memory bandwidth is now more of a architectural checkbox, as improved L3 caches for example are. In processors like Xeon Phi where it has much more BW to work with, it can make a difference.

I wouldn't discount 20% gain for Xeon Phi. 20% for a single feature is actually a lot. Also, the 20% number is on the same core. It still has the 2x64B load + 1x64B store units. It still has MCDRAM. But in the case of Skylake-SP versus Broadwell-EP, you're talking about 2x load/store bandwidth in addition to 2x FP. The gains are incredibly varied depending on code. Real world codes are far more complicated which is why its hard to get across the board large gains. Usually you need to double EVERYTHING to get 2x the performance. General rule of thumb I had for graphics were: 30% for double bandwidth, 30% for doubled texture filtering, 30% for shaders(Flops).
I believe you were looking for L3 bandwidth numbers of Skylake-X. The version 5.92 of AIDA64 gives those numbers, as you can see from the posts of owners in this thread:
https://forums.anandtech.com/thread...hange-your-vote.2504706/page-41#post-38974137
Then there's this :
http://www.hardware.fr/articles/964-3/nouveaux-caches-nouvelle-interconnexion.html
Before giving you the bandwidth figures below, we must tell you that we are not the only ones to have had little time to work on our article. The authors of the test tools that we use regularly, whether Hwinfo64 or Aida64 have also had little time to adapt to Skylake-X's particularities that change a lot, particularly in how to correctly detect the frequency of Processor and the BCLK.

Intel has added new mechanisms that require the processor to report the information itself, but these are poorly documented, which can lead to errors in some tools that do not necessarily read the frequencies correctly. We once again wish to thank the persistence and patience of the authors of the said software, which have amply helped us once again!

In the same way, bandwidth tests have not necessarily been "optimized" even if they seem correct at first glance.

For this test, we have configured the 7900X in "8 cores" mode by deactivating two of its cores. We disable HyperThreading, and the chip frequency is set to 3 GHz. This allows us to get a direct comparison with the Core i7-6900K and the Ryzen 1800X.
Let's put aside for a moment the latency we marked orange to compare the read bandwidths between 7900X and 6900K.

In terms of bandwidth, on the L1, one is in practice slower by 11% in reading and a little bit more in writing. Nothing serious, if we compare to Ryzen for example, Intel continues to have L1 excessively fast.

On the other hand when looking at the L2, things are all different. Of course, the write / copy bandwidth is only 12/13% lower, but the read bandwidth is divided by three! We have not given it back on this chart, but this figure of read bandwidth is even less than ... Piledriver. One will remain cautious, as we will see a little below the L2 seems to have a different operation than what the raw features provided by Intel can suggest, and this probably impacts this bandwidth value.

But if we now look at the L3, things are in the same line. This time the reading / copying operations are twice slower on Skylake-X than on Broadwell-E, and in reading, performance is divided once again by three.

In other words, these figures are not good, and if once again we will remind you that it is possible that Aida64's measurement tools do not correctly detect certain details, or are not adapted to certain features of Skylake- X, the general order of magnitude seems to be this.

As far as latency is concerned, here too we are cautious. The latency of the L1 is similar, but that of the L2 increases, lining up a little more on Zen than on Broadwell-E. And for L3, again, the latency is very high, almost 50% higher than that of Broadwell-E.
On the L1, there is no surprise, we find ourselves at a coherent level and does not move (latency is 4 cycles). On the L2, we ask ourselves a few questions. The "fast" benchmark of AIDA64 indicated a fairly high value, which corresponds more or less to accesses on a little less than one MB. Normal is the size of the L2 advanced by Intel. Except that what is surprising is that this L2 has a latency variable according to the size of the accesses. The first 256 KB (historical size of the Intel L2) have a very low latency while beyond latency rises. This is intriguing, and shows maybe that this L2 would function a little more complex than usual.

As far as L3 is concerned, it can also be seen that the latency is not the same at 1.5 MB as above, at 12 MB, this latency increasing fairly sharply (past 14 MB, it is in the main memory). There would therefore be an effect, although with a very slight impact, on the latency.
As far as AVX512 is concerned it is foolish to make grossly oversimplified statements like one core with AVX512 is worth four cores without AVX512, as some people like to do. I do not have access to the tests done by multi-billion dollar companies, but from personal experience running Einstein@Home doing tasks that use AVX, a Kaby Lake i3(which is at least 2.5x faster than a Conroe Dual-Core on its own, in x264 benchmark) is only a little more than 2X faster in completing such tasks than Conroe, which in theory would suggest a 4X speedup.
So in the real world where code is this variable, which requires specific per-use-case-tuning, one should not make claims like GPU vendors do in comparing against CPUs.

"Our new GPU can do 26TFLOPS! Oh wait, it's only half-precision"

"Great! What's your memory bandwidth?"

"Plenty - around 500GB/s. Should be enough right?"

":D"
 
  • Like
Reactions: Drazick

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,680
136
Base frequencies are for extreme scenarios when running heavy AVX loads or CPU + iGPU heavy loads.
I think there's been a shift in Intel's turbo policy for desktops. Look at i7 8700 specs: 4.3Ghz hexa core turbo and 65W TDP? Either 14nm++ is fantastic or the 4.3Ghz only applies to lighter loads. (gaming included)
 

dfk7677

Member
Sep 6, 2007
64
21
81
The only questions now are for how long are they going to sustain turbo frequency and what are their prices?
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
They would need multiple new chips for Macs, those would be low volume. iPhone chips won't replace desktop chips.

All they have to do is follow Intel's lead by extending the same core into different segments. Tweak the core counts, clockspeeds, and power saving features. Maybe remove some fixed-function hardware and/or graphics IP. Done deal. Still not a zero-cost venture, but much easier than designing an entirely new uarch.
 
Aug 11, 2008
10,451
642
126
You really think people would want to drop another $350+ on Coffee Lake that quickly? That's why the upgrade path is overrated, by the time you'd want to upgrade you'd probably want the new features anyway.
Yea, exactly. With the slow pace in cpu advancement, a cpu should last a minimum of 2, probably 3 to five years. I dont really see it as a big deal for most users if one has to drop a couple hundred for a new motherboard after that amount of time.
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
1,540
136
Yea, exactly. With the slow pace in cpu advancement, a cpu should last a minimum of 2, probably 3 to five years. I dont really see it as a big deal for most users if one has to drop a couple hundred for a new motherboard after that amount of time.

Some people feel the need to always keep up, or ahead of the Joneses.

I'm still proudly rocking my C2Q. :D
 

arandomguy

Senior member
Sep 3, 2013
556
183
116
Yea, exactly. With the slow pace in cpu advancement, a cpu should last a minimum of 2, probably 3 to five years. I dont really see it as a big deal for most users if one has to drop a couple hundred for a new motherboard after that amount of time.

Most users buying the mainstream platform are not even buying motherboards that cost a couple of hundred dollars. Those that are buying those super high end boards are mostly not very price sensitive since they are way off the cliff in terms of value to cost.
 

gx_saurav

Senior member
Dec 5, 2012
247
61
101
about.me
In real world out of this forum and the kind of enthusiasm we show, regular people buy and keep a PC for many years as long as it works fine.

Many users I know are using Core i3 Sandy Bridge. They purchased with 2 GB DD3800 RAM and slow upgraded to 8 GB and a SSD and still are happy with the system.

The upgrade bug is real and only affects us Enthusiasts. The market reacts to majority of users who are still going to use Quad Core CPUs for next few years and slow upgrade to 6C-12T in the next 3 - 4 years.
 

SpoCk0nd0pe

Member
Jan 17, 2014
26
11
46
I very much hope for a new socket because it means I could still hope for 20 pci-e lanes.

16 just isn't enough. I don't get why not more people are annoyed by this. I think a 350$ CPU without enough PCI-e lanes to run a GPU with 16 and an m.2 is a serious issue. Even AMD's 110$ ryzen 3 can do it (and you save on the boards too).
 
  • Like
Reactions: pcp7

gx_saurav

Senior member
Dec 5, 2012
247
61
101
about.me
I very much hope for a new socket because it means I could still hope for 20 pci-e lanes.

16 just isn't enough. I don't get why not more people are annoyed by this. I think a 350$ CPU without enough PCI-e lanes to run a GPU with 16 and an m.2 is a serious issue. Even AMD's 110$ ryzen 3 can do it (and you save on the boards too).

I second you on that. 20 PCIe lanes should be minimum in a CPU going forward. With 20, the bare minimum requirements of a system are met.

+ The chipset will give some lanes to. At max I can think of 24 PCIe lanes in a system actually being used

1. x16 for GPU
2. X4 for M.2 SSD
3. 2 x1 PCIe slots for Sound card, LAN card, TV Tuner etc if needed. But these are better purchased in USB.
 
  • Like
Reactions: pcp7

R0H1T

Platinum Member
Jan 12, 2013
2,582
162
106
I very much hope for a new socket because it means I could still hope for 20 pci-e lanes.

16 just isn't enough. I don't get why not more people are annoyed by this. I think a 350$ CPU without enough PCI-e lanes to run a GPU with 16 and an m.2 is a serious issue. Even AMD's 110$ ryzen 3 can do it (and you save on the boards too).
With PCIe 4.0 launching soon (hopefully) & then PCIe 5.0 supposed to be fast tracked by 2019-20 this question could well be moot, unless someone intentionally wants to gimp their latest platform to 8x PCIe 4/5.0 lanes :rolleyes:
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
1,540
136
I second you on that. 20 PCIe lanes should be minimum in a CPU going forward. With 20, the bare minimum requirements of a system are met.

+ The chipset will give some lanes to. At max I can think of 24 PCIe lanes in a system actually being used

1. x16 for GPU
2. X4 for M.2 SSD
3. 2 x1 PCIe slots for Sound card, LAN card, TV Tuner etc if needed. But these are better purchased in USB.


I think this is kind of Moot. Skylake has 16 dedicated PCIe lanes and 4 general purpose that go the chipset and are multiplexed pushing the budget beyond 24, for the combo of CPU and Chipset.
 

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
With PCIe 4.0 launching soon (hopefully) & then PCIe 5.0 supposed to be fast tracked by 2019-20 this question could well be moot, unless someone intentionally wants to gimp their latest platform to 8x PCIe 4/5.0 lanes :rolleyes:
Indications are that pcie 4 is going to be skipped, with pcie 5 showing up in 2020. Think about how much it would suck to spend a bunch on a new CPU and motherboard, only to have it outdated in a year or two and needing yet another new cpu and motherboard to stay current.
 

R0H1T

Platinum Member
Jan 12, 2013
2,582
162
106
Indications are that pcie 4 is going to be skipped, with pcie 5 showing up in 2020. Think about how much it would suck to spend a bunch on a new CPU and motherboard, only to have it outdated in a year or two and needing yet another new cpu and motherboard to stay current.
Assuming PCIe 5.0 final specs are released within a year or so, otherwise it'd be a waste of time waiting for the next major PCIe version to be finalized. The next upgrade (4.0) was supposed to happen a couple of years back, if they can't finalize 5.0 in a timely fashion & release compatible products on time then it'd be another major fail IMO, heck we're now (not right now per se) getting 20Gbps USB 3.2 whilst PCI SIG is twaddling their thumbs on the next major revision for close to a decade now!
 

Insomniator

Diamond Member
Oct 23, 2002
6,294
171
106
What is the big deal about pcie lanes? Is that an actual issue or just something people on this forum think they should care about?

Do GPU's even saturate x8 now? I remember x4 not even making a difference a few years ago.
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
I very much hope for a new socket because it means I could still hope for 20 pci-e lanes.

16 just isn't enough. I don't get why not more people are annoyed by this. I think a 350$ CPU without enough PCI-e lanes to run a GPU with 16 and an m.2 is a serious issue. Even AMD's 110$ ryzen 3 can do it (and you save on the boards too).
What are you running that is being hampered by 16 lanes? Certainly not a gpu and an m2 drive?
 
  • Like
Reactions: Drazick

TheF34RChannel

Senior member
May 18, 2017
786
309
136
I very much hope for a new socket because it means I could still hope for 20 pci-e lanes.

16 just isn't enough. I don't get why not more people are annoyed by this. I think a 350$ CPU without enough PCI-e lanes to run a GPU with 16 and an m.2 is a serious issue. Even AMD's 110$ ryzen 3 can do it (and you save on the boards too).

Lanes for your M.2 come from the chipset, not the CPU so I'm not sure what you're on about? Please do the research first before saying something randomly. Ah I only just now see you end with how Ryzen is better, so that's your aim. Well, thanks for your thoughts and enjoy being a glorified bug and beta tester for AMD :p
 
Last edited:
  • Like
Reactions: Pilum and CHADBOGA