Intel Skylake / Kaby Lake

[TheF34RChannel]

The main thing I am looking for from Icelake is a major change in uarch. Intel needs to outgrow Core. Apple's chip developments show this already. And no I don't think it's x86 legacy "holding back" Intel except maybe at the very low end. What they need to do is stop leaning exclusively on process advantage. They need to beat Apple and the ARMy with superior design.

Music to my ears! It is time for a uarch change and this is something I'll be excitingly following.

 

[coercitiv]

What I would like to see, is two chips PRICED about the same competing. So what if one has more cores. Also, whats the power consumption of each ? My wallet does most of the thinking, and for me power consumption also affects my wallet (DC computing 24/7)

I just saw an outstanding claim, that Zen can match Skylake in IPC, and was asking for further clarification and sources if possible. We don't judge value based on IPC, it's still an interesting metric though.

 

[dullard]

Are you sure about that? Coffee lake is basically Skylake IPC and Skylake chips when clocked at the same as Ryzen often lose in many tests. clock per clock Ryzen has great IPC.

Are you sure you are comparing apples to apples with the same number of cores and the same number of threads at the same clock speed? Or were you comparing something like a 16 thread Ryzen vs a 8 thread Kaby Lake?

To do an IPC comparison, single thread benchmarks are probably the most fair.
https://www.pcper.com/image/view/78910?return=node/67215

But even if you go with multi-threaded benchmarks, Ryzen with the same core count and even with a bit faster clock isn't quite as fast as Kaby Lake. Compare the 4C/4T Ryzen 1300X at 3.7 GHz turbo to the 4C/4T Intel 7400 at up to 3.5 GHz turbo. The 7400 has a slight lead even with the turbo speed disadvantage:
http://images.anandtech.com/doci/11658/mt_cpu.png

 

[IntelUser2000]

Music to my ears! It is time for a uarch change and this is something I'll be excitingly following.

New architecture won't be enough. You can always end up with a mediocre one. The problem is due to management. Technical knowledge and throwing money at R&D can only get you so far. They have lost a lot of people with decades of invaluable experience in the last few years.

Next few years will determine whether they can recover from this or the decline will accelerate. CanardPC translation says Icelake generation will be the fork in the road. It will determine whether they have something to continue. It will determine whether Apple will abandon their chips.

The worst is probably the situation becoming bad to a point where less knowledgeable people start pointing fingers at x86 being the source of their problems. The company gets split into manufacturing and architecture groups. A decade down, both groups falter and both get sold off for IP and patents.

 

[tamz_msc]

Are you sure you are comparing apples to apples with the same number of cores and the same number of threads at the same clock speed? Or were you comparing something like a 16 thread Ryzen vs a 8 thread Kaby Lake?

To do an IPC comparison, single thread benchmarks are probably the most fair.
https://www.pcper.com/image/view/78910?return=node/67215

But even if you go with multi-threaded benchmarks, Ryzen with double the cores isn't double the speed at the same clocks:
https://www.pcper.com/image/view/78909?return=node/67215

Talks about multi-threaded performance - shows us benchmarks for MP3 encode.

 

[Anarchist Mae]

Are you sure you are comparing apples to apples with the same number of cores and the same number of threads at the same clock speed? Or were you comparing something like a 16 thread Ryzen vs a 8 thread Kaby Lake?

To do an IPC comparison, single thread benchmarks are probably the most fair.
https://www.pcper.com/image/view/78910?return=node/67215

But even if you go with multi-threaded benchmarks, Ryzen with double the cores isn't double the speed at the same clocks:
https://www.pcper.com/image/view/78909?return=node/67215

Here's my own bench of my old i3-6100 and my R7 1700 both at the same clock speed:

Single threaded: http://openbenchmarking.org/result/1706134-STOA-170613721
Multi threaded: http://openbenchmarking.org/result/1706134-STOA-170613437

The same memory was used in both systems.

 

[DrMrLordX]

New architecture won't be enough. You can always end up with a mediocre one. The problem is due to management.

It remains to be seen whether Krzanich can lead Intel well enough to launch an entirely new uarch. Most of the current signs point to "no", but remember, we are talking about Intel here. They do have vast resources at their disposal. If they really need to poach the talent necessary to get things done, they have the cash, and they can probably sandbox the design effort to prevent Krzanich from screwing it up entirely.

Any company that can go from Pentium D to Core 2 is capable of a massive turnaround, even under the same executive team.

 

[JoeRambo]

Any company that can go from Pentium D to Core 2 is capable of a massive turnaround, even under the same executive team.

With C2D things were different, as they had kept good old P6 core alive with Pentium-M Dothan/Banias derivatives. So taking that, making it wider and tuning for power efficiency gave them ability to turn around and then destroy AMD by adding IMC to core. In hindsight, Sandy Bridge was a true redesign that probably was started during P4 dark days. ( as in Nehalem was C2D with IMC bolted and caches rebalanced toward small L2 + L3 slices, while SB was a true new core with P4 lessons learned, with major design breaks from Core2).

Compared to that, Intel currently has no alternative high perf CPU design, is stuck on 14nm, with 10nm having worse perf characteristics. Much harder to turn things around.

 

[PeterScott]

With C2D things were different, as they had kept good old P6 core alive with Pentium-M Dothan/Banias derivatives. So taking that, making it wider and tuning for power efficiency gave them ability to turn around and then destroy AMD by adding IMC to core. In hindsight, Sandy Bridge was a true redesign that probably was started during P4 dark days. ( as in Nehalem was C2D with IMC bolted and caches rebalanced toward small L2 + L3 slices, while SB was a true new core with P4 lessons learned, with major design breaks from Core2).

Compared to that, Intel currently has no alternative high perf CPU design, is stuck on 14nm, with 10nm having worse perf characteristics. Much harder to turn things around.

Yes things were different. Back in those days, Intel was actually behind, unlike now where they are still ahead. Despite the hype, Ryzen has worse IPC and runs at lower clock speeds, so Intel is hardly floundering like in the P4 dark days. They are still in the lead, with Coffee lake around the corner.

Intel can just keep incrementally improving like they have and stay ahead of AMD, they don't need a revolution to save them.

 

[Edrick]

Intel can just keep incrementally improving like they have and stay ahead of AMD, they don't need a revolution to save them.

If AMD can release Ryzen II at 7nm before Intel gets anything out the door at 10nm, then Intel will need a revolution to save them.

 

[PeterScott]

If AMD can release Ryzen II at 7nm before Intel gets anything out the door at 10nm, then Intel will need a revolution to save them.

The odds of that happening are similar to me winning powerball.

 

[dullard]

Here's my own bench of my old i3-6100 and my R7 1700 both at the same clock speed:

Single threaded: http://openbenchmarking.org/result/1706134-STOA-170613721
Multi threaded: http://openbenchmarking.org/result/1706134-STOA-170613437

The same memory was used in both systems.

Thanks for the data. The stock Skylake i3 was essentially tied with the turbo Ryzen 1700 in single thread (within the standard error of the tests). From that test, the IPC is the same. Of course you are comparing a new Ryzen with an old i3 though.

From your multi-thread test, the turbo Ryzen was 3.67x faster with 4x more threads. Do you know the penalty from having 16 threads vs 4 threads in that particular benchmark? Without knowing the penalty in that particular test, I we can't draw specific conclusions.

 

[dullard]

Talks about multi-threaded performance - shows us benchmarks for MP3 encode.

I took it from a website under the multi-thread title. Sorry that I didn't look it up closely enough. My fault. To correct my mistake, I put in today's Anandtech multithread test result in it's place.

The 4C/4T Ryzen 1300X 3.5 GHz base 3.7 GHz turbo slightly under-performed the very similar 4C/4T Intel 7400 3.0 GHz base 3.5 GHz turbo. That is true even though the 1300X had a slight clock speed advantage. It isn't quite the same clock, but at least it is the same core and thread count.

http://images.anandtech.com/doci/11658/mt_cpu.png

Heck, to be even more fair, the Ryzen 1200X had closer clocks to the Intel 7400, but I won't go there.

Ryzen was close, but still not quite on par on average on the multi-threaded tests. Ryzen just doesn't quite have the same IPC as Kaby Lake. Ryzen's advantages are more cores, more threads, and usually a good value. Intel's advantage is more IPC and sometimes higher clocks.

Do you care to counter with any clock-for-clock and thread-for-thread comparisons of your own?

 

[crashtech]

@dullard , I think most of us agree that "old" Skylake has the same IPC as Kaby Lake, the difference in the latter being a speed bump and IGP changes.

 

[tamz_msc]

I took it from a website under the multi-thread title. Sorry that I didn't look it up closely enough. My fault. To correct my mistake, I put in today's Anandtech multithread test result in it's place.

The 4C/4T Ryzen 1300X 3.5 GHz base 3.7 GHz turbo slightly under-performed the very similar 4C/4T Intel 7400 3.0 GHz base 3.5 GHz turbo. That is true even though the 1300X had a slight clock speed advantage. It isn't quite the same clock, but at least it is the same core and thread count.

http://images.anandtech.com/doci/11658/mt_cpu.png

Heck, to be even more fair, the Ryzen 1200X had closer clocks to the Intel 7400, but I won't go there.

Ryzen was close, but still not quite on par on average on the multi-threaded tests. Ryzen just doesn't quite have the same IPC as Kaby Lake. Ryzen's advantages are more cores, more threads, and usually a good value. Intel's advantage is more IPC and sometimes higher clocks.

Do you care to counter with any clock-for-clock and thread-for-thread comparisons of your own?

It is very clear from the numbers that the 1500X got in the AnandTech review - the reason why the Ryzen 3's sometimes fall much behind the 1500X is because of the reduced L3 cache. 2x8MB L3 helps improve performance because because Ryzen relies on inter-core communication within a CCX's L3 cache to improve hit rates. It isn't 2.5MB/core like Broadwell-E or 1.375 MB/core Skylake-X, it's 8MB per CCX. But I digress. The best quad-core Ryzen that you can buy right now is the 1500X, if it can be found at a price that sufficiently separates it from the 1600.

Clock-for-clock and thread-for-thread comparisons? Here you go(French):
http://www.hardware.fr/articles/964-6/piledriver-zen-broadwell-e-skylake-x-3-ghz.html

 

[Yakk]

If AMD can release Ryzen II at 7nm before Intel gets anything out the door at 10nm, then Intel will need a revolution to save them.

That would mostly depend on Samsung I guess, then GF would license it or AMD maybe contract Samaung directly and pay the penalty(?) if they absolutely must have it quickly.

 

[scannall]

That would mostly depend on Samsung I guess, then GF would license it or AMD maybe contract Samaung directly and pay the penalty(?) if they absolutely must have it quickly.

7 nm is an IBM high performance process, that will be glofo only. Samsung and TSMC are working on lower clocked for mobile processes at 7 nm. The IBM 'LP" (Lead Performance) is targetting 5 Ghz. Considering IBM has produced 12 core, 48 thread monsters at 5.5 Ghz I'd say that's a goal within reach.

 

[Yakk]

7 nm is an IBM high performance process, that will be glofo only. Samsung and TSMC are working on lower clocked for mobile processes at 7 nm. The IBM 'LP" (Lead Performance) is targetting 5 Ghz. Considering IBM has produced 12 core, 48 thread monsters at 5.5 Ghz I'd say that's a goal within reach.

Ah ok, I had it reversed thanks. In any case I see it as certainly feasible.

 

[dullard]

The best quad-core Ryzen that you can buy right now is the 1500X, if it can be found at a price that sufficiently separates it from the 1600.

Clock-for-clock and thread-for-thread comparisons? Here you go(French):
http://www.hardware.fr/articles/964-6/piledriver-zen-broadwell-e-skylake-x-3-ghz.html

The 1500X is a good value. But personally, I think the 1600X chip is the best up-front value of all chips at the moment. The 1600X has faster speed and 2 more cores than the 1500X, for only $40 more. But if you take total cost of ownership into account, the lower power 1600 is probably the best buy. To keep to the topic of the thread though, the upcoming Intel 8600 might win that best buy award from me if the rumored speeds are true (unlike what you appear to be, I'm vendor neutral).

Back to the IPC claims, I don't speak French, but a quick glance looked like out of 22 tests the Ryzen won 3, was slightly behind but I'll call them a tie in 3, and was significantly behind in 16. That is why I say the Ryzen IPC isn't quite yet up to par with Intel's processors. Close, but not quite there.

 

[PeterScott]

That would mostly depend on Samsung I guess, then GF would license it or AMD maybe contract Samaung directly and pay the penalty(?) if they absolutely must have it quickly.

No it would depend on meteors destroying all of Intels 10nm production facilities.

Krzanich already showcased a Canon Lake 10nm in a laptop. They are almost certainly ramping production right now.
http://fortune.com/2017/01/05/intel-ces-2017-moore-law/
"But I thought best to just bring (a) 10 nanometer Canonlake product on stage and let you see it live."

 

[Arachnotronic]

No it would depend on meteors destroying all of Intels 10nm production facilities.

Krzanich already showcased a Canon Lake 10nm in a laptop. They are almost certainly ramping production right now.
http://fortune.com/2017/01/05/intel-ces-2017-moore-law/
"But I thought best to just bring (a) 10 nanometer Canonlake product on stage and let you see it live."

Samples on stage != ramping production now.

Yields could be (and probably were) very bad at the time.

 

[PeterScott]

Samples on stage != ramping production now.

Yields could be (and probably were) very bad at the time.

Demo on stage 7 months ago and plans to release by end of the year, do indicate ramping production around now.

The argument is that AMD could release 7nm Zen before Intel releases 10nm Cannon Lake.

It should be blindingly obvious that Intel is much closer to that goal. Or did I miss the 7nm Zen demos?

 

[tamz_msc]

The 1500X is a good value. Back to the IPC claims, I don't speak French, but a quick glance looked like out of 22 tests the Ryzen won 3, was slightly behind but I'll call them a tie in 3, and was significantly behind in 16. That is why I say the Ryzen IPC isn't quite yet up to par with Intel's processors. Close, but not quite there.

You mixed up gaming and applications. Since "gaming IPC" is a vague term, looking at only the application numbers we have:

Spoiler

Excluding 7-Zip, WinRAR and Lightroom, which is completely fair because Skylake-X is behind Broadwell-E in these tests, we have Skylake-X 8-core roughly 8% ahead of Ryzen 8-core, clock-for-clock, thread-for-thread.

Hardly what I'd call 'significant'.

 

[Zucker2k]

You mixed up gaming and applications. Since "gaming IPC" is a vague term, looking at only the application numbers we have:

Spoiler

Excluding 7-Zip, WinRAR and Lightroom, which is completely fair because Skylake-X is behind Broadwell-E in these tests, we have Skylake-X 8-core roughly 8% ahead of Ryzen 8-core, clock-for-clock, thread-for-thread.

Hardly what I'd call 'significant'.

It will, if you factor in time. For example, rendering and encoding jobs could take hours at a time, if not days.

 

[dullard]

Excluding 7-Zip, WinRAR and Lightroom, which is completely fair because Skylake-X is behind Broadwell-E in these tests, we have Skylake-X 8-core roughly 8% ahead of Ryzen 8-core, clock-for-clock, thread-for-thread.

Hardly what I'd call 'significant'.

So, Let me get this right:
1) You are excluding the non-gaming tests where Zen was behind Skylake by a large amount (28.1%, 40.9%, and 23.1%).
2) Then you throw out gaming tests.
3) Then you conclude that since the remaining tests show Zen behind by 8% that there is no difference?

Do you expect anyone to take you seriously?