Intel Skylake / Kaby Lake

[Dave2150]

135,15GOPS vs 130,30GOPS = 4%.

?? As the others said, provide sources for your numbers, or they are completely pointless.

 

[JoeRambo]

135GOps for SKylake is here:

http://www.sisoftware.eu/rank2011d/...d4e3d7e6d3e1d2f486bb8badc8ad90a086f5c8f0&l=de

Cant find relevant score for Haswell

 

[Arachnotronic]

Nice find.
13.1% better performance from Core i7 6700K at stock (overall).
The integer scores are particularly impressive.

Core i7 6700K (up to 4.2GHz):
Multi-Media Integer 390,86Mpix/s (19% advantage)
Multi-Media Long-int 181,37Mpix/s (25.6% advantage)

Core i7 4790K (up to 4.4GHz)
Multi-Media Integer 328,08Mpix/s
Multi-Media Long-int 144,12Mpix/s

There's also per clock results in the individual results section.

Core i7 6700K: 79,73Mpix/s/GHz
Core i7 4790K: 67,42Mpix/s/GHz

18.25% better performance per clock.

If we assume BDW is 5.5% better per clock than HSW, then this implies SKL is ~12% better per clock than BDW. Looks pretty reasonable.

 

[Sweepr]

135GOps for SKylake is here:

http://www.sisoftware.eu/rank2011d/...d4e3d7e6d3e1d2f486bb8badc8ad90a086f5c8f0&l=de

Cant find relevant score for Haswell

Here's Core i7 4790K scores, average score is 130.30 GOPS (8 Threads).
It should be noted that this is a single pre-launch Core i7 6700K score. You can easily find single Core i7 4790K scores below 125 GOPS, especially pre-launch scores like the Skylake result we're using as reference.

Just a few examples (Core i7 4790K - 113-122 GOPS):
http://www.sisoftware.eu/rank2011d/...d4e2d1e0d4e4ddfb89b484a2c7a29faf89fac7ff&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e9dee8dafc8eb383a5c0a598a88efdc0f8&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d1e1d1e3dbfd8fb282a4c1a499a98ffcc1f9&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e9d9efd7f183be8ea8cda895a583f0cdf5&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e7d0e4d7f183be8ea8cda895a583f0cdf5&l=en
http://www.sisoftware.eu/rank2011d/...d4e3d7e1d6e1d1f785b888aecbae93a385f6cbf3&l=en

 

[Abwx]

Where did you find those numbers?

Here :

6700k: http://www.sisoftware.eu/rank2011d/...d5e0d5e0d8efdbfd8fb282a4c1a499a98ffcc1f9&l=de
4790k: http://www.sisoftware.eu/rank2011d/...d5e0d5ecd8e8dcfa88b585a3c6a39eae88fbc6fe&l=de

SKL

http://www.sisoftware.eu/rank2011d/...b68bba9cf4c9fcdaa29fae88ed88b585a3d0edd5&l=de


HSW


http://www.sisoftware.eu/rank2011d/...bd80b197ffc2f7d1a994a583e683be8ea8dbe6de&l=de

 

[Fjodor2001]

Looks like you are right. In that case Skylake is even a regression compared to Haswell in terms of perf/watt for the models being compared.

 

[Dave2150]

Looks like you are right. In that case Skylake is even a regression compared to Haswell in terms of perf/watt for the models being compared.

Indeed, lets hope that figure is wrong D:

 

[Sweepr]

Looks like you are right. In that case Skylake is even a regression compared to Haswell in terms of perf/watt for the models being compared.

Lol. Even Broadwell beats Haswell this benchmark, a Core i7 5775C manages to outperform 200MHz higher clocked Core i7 4770K.

http://media.bestofmicro.com/T/Z/497447/original/61-Arithmetic-MT.png

Do you really believe that Skylake is a regression compared to Broadwell in both IPC and perf/watt? You shouldn't draw conclusions based on a single pre-launch Core i7 6700K score vs an average of Core i7 4790K scores 1 year after launch. As shown above some of the first Core i7 4790K scores were quite a bit lower than today's.

Oh, and the Multi-Media score shows a different picture.
Core i7 6700K: 335,52Mpix/s
Core i7 4790K (average): 296,66Mpix/s

 

[jpiniero]

There's something screwy about those scores. The 5930K shouldn't be that much faster than the 5920K for instance.

 

[Abwx]

Looks like you are right. In that case Skylake is even a regression compared to Haswell in terms of perf/watt for the models being compared.

We ll know more once there s power number at Hardware.fr or eventualy THG, what is sure is that the 95W rating wasnt randomly increased over HW s 84/88W, not sure that it will be less efficient, though, more likely that the progress is not huge to say the least.

 

[Abwx]

If we assume BDW is 5.5% better per clock than HSW, then this implies SKL is ~12% better per clock than BDW. Looks pretty reasonable.

Only if new instructions, wich are not available in Haswell, are used, so that s certainly not in the majority of softwares.

 

[Arachnotronic]

Only if new instructions, wich are not available in Haswell, are used, so that s certainly not in the majority of softwares.

What major new instructions did Intel introduce in BDW that were not available in HSW?

 

[Abwx]

What major new instructions did Intel introduce in BDW that were not available in HSW?

Not a lot isnt it, that means that the 5.5% is rather a corner case, so far Intel didnt provide any bench, it s like they think that we must take those numbers at face value ...

Intel ADX: ADOX and ADCX for improving performance of arbitrary-precision integer operations

RDSEED for generating 16-, 32- or 64-bit random numbers from a thermal noise entropy stream, according to NIST SP 800-90B and 800-90C

PREFETCHW instruction

Broadwell's Intel Quick Sync Video hardware video decoder adds VP8 hardware decoding support

Add also the dual FMA (uarchitectural improvement) and improved AES encryption...
http://en.wikipedia.org/wiki/Broadwell_(microarchitecture)

 

[Arachnotronic]

Not a lot isnt it, that means that the 5.5% is rather a corner case, so far Intel didnt provide any bench, it s like they think that we must take those numbers at face value ...




Add also the dual FMA (uarchitectural improvement) and improved AES encryption...
http://en.wikipedia.org/wiki/Broadwell_(microarchitecture)

Pretty sure dual FMA came as part of HSW.

 

[Abwx]

It s written in the slide above that it s an improvement brought by BDW...

 

[inf64]

If "dual FMA" was anything but a marketing talk then we would see 2x FP throughput claims. So Arachnotronic is correct, Haswell already has "dual" FMA units and also has AVX2.0 support.

 

[Abwx]

If "dual FMA" was anything but a marketing talk then we would see 2x FP throughput claims. So Arachnotronic is correct, Haswell already has "dual" FMA units and also has AVX2.0 support.

So those slides are complete rubbish.?..

But then why should the IPC be accurate, as said there s not a single bench, at least to my knowledge, that has been published by Intel.

 

[Edrick]

Looks like BDW fixed TSX at least.

 

[Fjodor2001]

Looks like BDW fixed TSX at least.

Yes, but it's not fixed in all Broadwell models. See this:

"This erratum was evidently discovered quite recently, too late for a fix to be included in the first revision of Intel's upcoming Broadwell Y-series chips."

 

[ShintaiDK]

So those slides are complete rubbish.?..

Nope.

 

[ShintaiDK]

Looks like BDW fixed TSX at least.

Yep. Important for Xeons.

 

[SAAA]

Here's Core i7 4790K scores, average score is 130.30 GOPS (8 Threads).
It should be noted that this is a single pre-launch Core i7 6700K score. You can easily find single Core i7 4790K scores below 125 GOPS, especially pre-launch scores like the Skylake result we're using as reference.

Just a few examples (Core i7 4790K - 113-122 GOPS):
http://www.sisoftware.eu/rank2011d/...d4e2d1e0d4e4ddfb89b484a2c7a29faf89fac7ff&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e9dee8dafc8eb383a5c0a598a88efdc0f8&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d1e1d1e3dbfd8fb282a4c1a499a98ffcc1f9&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e9d9efd7f183be8ea8cda895a583f0cdf5&l=en
http://www.sisoftware.eu/rank2011d/...d4e2d0e7d0e4d7f183be8ea8cda895a583f0cdf5&l=en
http://www.sisoftware.eu/rank2011d/...d4e3d7e1d6e1d1f785b888aecbae93a385f6cbf3&l=en

I did a quick comparison of the individual scores, int and float subtests, using those and other scores, they present a much better picture than the GOPS value:

Name Skylake Haswell→
GHz     4,21    4,4     4,59    4,2     4,4     4,4     4,42    4,3     4,4     4,3     4,6
Integer 198,58  178,56  200,09  183,17  191,35  180,88  135,41  136,96  140,41  141     147,1
Float32 108,55  111,51  127,07  115,48  121,13  115,76  113,25  113,23  118,53  116,92  121,5
Float64 77,94   77,35   86,79   78,38   82,93   79,24   78,8    80,19   82,34   81,42   85,17
 
Name            Skylake Haswell Increase:
INT/GHz         47,16   37,15   1,27
FLOAT/GHz       25,78   26,69   0,97
                18,51   18,46   1,00

So a lot faster in integer and about the same in floating operations. AVX3 should adress that, on Xeons at least...

 

[mikk]

Yes, but it's not fixed in all Broadwell models. See this:

"This erratum was evidently discovered quite recently, too late for a fix to be included in the first revision of Intel's upcoming Broadwell Y-series chips."

The only affected Broadwell is Core M in E0 Stepping. Core M F0 and all later Broadwell SKUs have been fixed.

 

[paul878]

I am too lazy to read.
So is Skylake a worthy upgrade from my 2600k @ 4ghz?

 

[SAAA]

I am too lazy to read.
So is Skylake a worthy upgrade from my 2600k @ 4ghz?

It depends on definition of worthy: are you unsatisfied with your current CPU in some activities? If you need more than 4 cores now then Skylake won't be much better for sure, go with any hex core or better. But it will be good for gaming and lowly threaded applications, think Nehalem-Sandy jump putting together all the IPC increases since your generation.