Intel Skylake / Kaby Lake

[krumme]

That's probably the most optimistic outlook I've seen of Intel's roadmap to date. You are aware of all the delays and so forth, no? Coffeelake shouldn't even exist. Desktop users are supposed to be on Cannonlake right now. Hell even Kabylake wasn't supposed to exist! Everything is in doubt now. It is up to Intel to show us that they can execute on some node other than 14nm.

Intel management bs. I dont know how they got themselves into this idiotic lie they could execute like Moores law was eternal. While they were cutting capex !. What a disgrace to the Intel engis working on ground floor. Self destructive management behavior imo.

What is happening now was evident 4 years ago even from the outside. When some of us spoke about it we took tons of flak for years now we aparently have to lisnt to years of complaining.

If we get back to actual product on the market it looks to me all product in Intels portfolio have nice proces performance. Fmax for cfl is super skl serverline is fine. Dense and performant stuff.

The challenge is the basic business model for sharing proces cost and to a certain degree to wide fp core vs a leaner ryzen product. But that is unrelated.

 

[DrMrLordX]

Intel management bs.

That doesn't make it any better that they've sold us the same core in three different iterations. And it looks like Cannonlake is the same core, die-shrunk. Icelake may not be radically different. Though I wish that it were.

Intel is leaving themselves open to competition by rehashing Skylake over and over again. And it is not AMD that should concern them.

 

[VirtualLarry]

That doesn't make it any better that they've sold us the same core in three different iterations. And it looks like Cannonlake is the same core, die-shrunk. Icelake may not be radically different. Though I wish that it were.

Intel is leaving themselves open to competition by rehashing Skylake over and over again. And it is not AMD that should concern them.

This is much akin to how AMD rested on their process laurels, around 90nm or 65nm, and delayed the introduction of a new process, and then fell behind in the market. I feel that with what you've explained above, that Intel is doing much the same with their architecture lead.

 

[DrMrLordX]

AMD did rest on 90nm for too long. They made the mistake of not investing enough in advanced process nodes (it was actually more complicated than that, but essentially, that is what happened to them). AMD's 65nm process was a major disappointment. It was not long after that that they spun off their entire fab business. Hooray for Hector.

Intel is investing a good bit into advanced process nodes, though doing this "while cutting capex" (to quote krumme) may be part of the problem. Whatever else may be going wrong in the foundry portion of their business, I do not know. I look at TSMC, Samsung, and the GF/IBM hybrid and I see progress. It could be that the latter two entities are tech-sharing to get around their problems; depending on whom you believe, TSMC and Samsung may be sharing tech involuntarily. Who is stealing more from whom is unknown to me.

 

[TahoeDust]

Intel is leaving themselves open to competition by rehashing Skylake over and over again. And it is not AMD that should concern them.

Is there another player looking to jump into the fray?

 

[IntelUser2000]

Is there another player looking to jump into the fray?

Yes, always. Consumers. Most business seem to be naturally lazy. They turn into profit-eking machines.

ARM Vendors and Apple should factor in as well, though it should be secondary. While it looks at the moment it doesn't matter because the OS doesn't overlap, long term technical deficiencies turn into profit and revenue deficiencies.

Fmax for cfl is super skl serverline is fine. Dense and performant stuff.

Cannonlake flop turned out to be Kabylake and Coffeelake. Actually, we should have seen Cannonlake instead of Kabylake, and right now we should have been waiting for Icelake any day now. AMD always does their big jump when Intel starts flopping.

The whole hoopla around 2 extra cores should not exist even to Anandtech Forum users. Cannonlake would have probably brought 3-5% perf/clock improvement, and Icelake another 7-10%. That's 10-15% improvement everyone can use. Extra cores it would have brought is just an icing on the cake.

For us it does not matter much. But for Intel it sucks much for them that their most profitable lines can't use 10nm Tock core. The Gen 11 GT2 on Icelake probably has 48EUs and perform nearly 2x what it can do now, while iVR starting on CNL would have brought 20-30% improvement in battery life for laptops.

Skylake-SP is OK. I'm a bit surprised how well EPYC is doing. Only the top end Platinum 8180 can beat top of the line EPYC, and only by ~10%. They messed up the lineup by needing you to buy $7k CPUs if you want more performance than top of the line Xeon E5s. Execution wasn't perfect there either. Their problems extend into 3D XPoint, and even FPGA integration judging by Cascade Lake.

The reason the 14nm modem that should have been here nearly 2 years ago is instead coming next year is because of their execution issues and process.

They could have had 10nm Knights Hill, which would have brought significant perf/watt and absolute performance improvements, rather than the patchwork "for DL" they call Knights Mill.

 

[krumme]

Process cost goes up exponentially.
But intels portfolio cant carry those cost. Therefore they adapted capex a few years ago as they should. Years before a11 ryzen whatever. It's unrelated.

At the same time mobile kept growing and we saw eg Samsung r&d explode. People here said no its not for process it's for ehh...well it primarily was and now Samsung is forerunner for euv. For good reason.

Intel clearly had no interest in euv because they are competitive with complex tighter standard stuff. But obviously there is only so much you can go down that tighter road while adapting cost. At the same time mobile just pushes more money into a by definition more cost effective ecosystem. You are bound to look worse vs that.

It's not like people on the floor and first and second level management at Intel is running slower or beeing more stupid and taking worse decisions. They are probably better than they have ever been. With the business structure it's bound to be like it is.

Imo it's more like they have been a bit complacent on the arch design side. Especially the gpu have been stagnating since ivy bridge.

 

[LTC8K6]

That doesn't make it any better that they've sold us the same core in three different iterations. And it looks like Cannonlake is the same core, die-shrunk. Icelake may not be radically different. Though I wish that it were.

Intel is leaving themselves open to competition by rehashing Skylake over and over again. And it is not AMD that should concern them.

If 10nm didn't work out well at the first try, what else could they do but try to tweak a proven design?

 

[Zucker2k]

This is much akin to how AMD rested on their process laurels, around 90nm or 65nm, and delayed the introduction of a new process, and then fell behind in the market. I feel that with what you've explained above, that Intel is doing much the same with their architecture lead.

None of what you said is true.

 

[Bouowmx]

Kaby Lake-R Core i7-8550U: https://www.notebookcheck.net/Dell-XPS-13-i7-8550U-QHD-Laptop-Review.257650.0.html
20 W power limit (up from 15): 2.5 GHz in Prime95. Don't know which version (for type of SIMD used, likely latest with AVX2), or type of Prime95 stress test used

 

[DrMrLordX]

If 10nm didn't work out well at the first try, what else could they do but try to tweak a proven design?

Well, nothing, but that's a poor plan as well. Desperation is still desperation.

None of what you said is true.

No, he had a point.

AMD delayed their introduction of 65nm back in '05/'06 with plans to do nothing but sell k8 derivatives against P4 variants (notably Cedarmill and Presler, if I recall correctly). It wasn't until late 2006 that they finally started to sell 65nm chips, and what we got was horribly process-limited. I had one of the 65nm x2s and it wouldn't do more than 3200 MHz, whereas older 90nm x2s were hitting 3500-3600 MHz. Plus the newer 65nm variants had l2 cache size issues and weird RAM multiplier idiosyncracies based on CPU multiplier.

Anyway it wasn't the upgrade they needed to take on Conroe. At all. It was just more of the same.

 

[The Stilt]

Finally got my 7960X system running and the performance is rather impressive.
There is no "magical" AVX2 performance increase (as rumored) from the additional resources of the core thou.

RZN vs. CFL vs. SKL-X in X265 2.5+31:

RZN: /w AVX2 = 100.00%, /wo AVX2 = 105.21%
CFL: /w AVX2 = 130.61%, /wo AVX2 = 101.13%
SKL-X: w/ AVX2 = 135.47%, /wo AVX2 = 105.21%

Ryzen's performance without AVX2 is impressive, but it is sad to see that there is still a penalty (like on Excavator) when running 256-bit code.

 

[Zucker2k]

Finally got my 7960X system running and the performance is rather impressive.
There is no "magical" AVX2 performance increase (as rumored) from the additional resources of the core thou.

RZN vs. CFL vs. SKL-X in X265 2.5+31:

RZN: /w AVX2 = 100.00%, /wo AVX2 = 105.21%
CFL: /w AVX2 = 130.61%, /wo AVX2 = 101.13%
SKL-X: w/ AVX2 = 135.47%, /wo AVX2 = 105.21%

Ryzen's performance without AVX2 is impressive, but it is sad to see that there is still a penalty (like on Excavator) when running 256-bit code.

For the "/wo AVX" scores which chip is the reference? I see no 100% in there. The picture is still clear, just curious.

 

[The Stilt]

For the "/wo AVX" scores which chip is the reference? I see no 100% in there. The picture is still clear, just curious.

RZN /w AVX2 = 100.0%

I didn't test without AVX because AVX in X265 basically makes no difference.

 

[el etro]

Next year Xeon/HEDT update could be Kaby-Lake E, again, but this time at LCC/HCC.

 

[jpiniero]

Next year Xeon/HEDT update could be Kaby-Lake E, again, but this time at LCC/HCC.

It's called Cascade Lake and yeah they are updating it to 14++. I think higher clocks and proper support for Optane DIMMs is it.

 

[el etro]

It's called Cascade Lake and yeah they are updating it to 14++. I think higher clocks and proper support for Optane DIMMs is it.

Nice, but 14+ have better power characteristcs for server processors.
14++ is better for HEDT processors.

I'm just wish i could buy a m.2 optane SSD at a sane price, optane SSDs are great.

 

[Eug]

OK, a little off topic but... I just learned something. There is no lake officially called Kaby Lake. It's short for Kabinakagami Lake. Dunno if this was posted already, but there ya go.

 

[wingman04]

Here is a link to official Kaby Lalke. https://ark.intel.com/products/97144/Intel-Core-i5-7600K-Processor-6M-Cache-up-to-4_20-GHz

 

[The Stilt]

X265 just received it's first roll of AVX512 assembly.

The performance improvement over AVX2 is ~6.5% for MT and ~5.1% for ST, or IPC (1080, Medium-preset, CRF RC).
Meanwhile the power wise (average) the increase is around 3.4%.

This difference is not going to show up in reviews which test SKU vs. SKU at their true default settings, as Skylake-X CPUs have 200-500MHz lower frequency during AVX512 compared to AVX2.
Since AVX512 instructions are executed relatively rarely in the current implementation, the frequency will vary between AVX512 - AVX2 and SSE set points during execution. Essentially AVX512 will cause a slight reduction to the average clock compared to AVX2,
however obviously the difference isn't as great as the difference between the nominal AVX2 and AVX512 frequencies would indicate.

 

[IntelUser2000]

Stilt: Intel claims that on Xeon Scalable chips, AVX-512 specific optimizations grant 18% gains in addition to the gains Skylake-SP changes bring.

Perhaps the consumer chips are limited by bandwidth?

 

[The Stilt]

Stilt: Intel claims that on Xeon Scalable chips, AVX-512 specific optimizations grant 18% gains in addition to the gains Skylake-SP changes bring.

Perhaps the consumer chips are limited by bandwidth?

Bandwidth is the limit on many AVX512 workloads, but not with the X265.
384-bit memory interface definitely helps.

 

[IRobot23]

Bandwidth is the limit on many AVX512 workloads, but not with the X265.
384-bit memory interface definitely helps.

So future probably brings on die RAM (HBM) for CPUs or iGPU acceleration.

IMC/IF is probably most important thing why Ryzen does so well in AVX.

 

[The Stilt]

IMC/IF is probably most important thing why Ryzen does so well in AVX.

??

 

[IRobot23]

??

Ryzen bandwidth is great. Meaning is does efficiently use FPU (AVX). I remeber slide where AMD said (or YT video) to something like this : "why using AVX3 if even AVX/2 is not efficiently used."
Comparing ryzen to coffee on AVX there is really small difference.