Intel Skylake / Kaby Lake

[Sweepr]

Even an i3 6100 is memory limited.

i3 6100 vs the i3 4160. Huge improvement. And in MT an i3 6100 is close to FX6300. 2 vs 6 cores.

Ryse looks really poorly coded. Hence the massive memory requirement.

This >20% performance bump from Haswell Core i3 is quite a game changer. I would probably pick a FX6300 for future proofing instead of a Core i3 back in Haswell days, but not anymore. Skylake Core i3 basically matches matches it in Cinebench MT, beats it in 3DMark Physics and comes real close in x264 encoding where the hexa-core should have a very large advantage (hexa-core vs dual-core with HT). Have you seen the videos? Huge difference in framerates in some titles, especially minimum framerates. Intel's per-core performance advantage is huge right now.

 

[ShintaiDK]

This >20% performance bump from Haswell Core i3 is quite a game changer. I would probably pick a FX6300 for future proofing instead of a Core i3 back in Haswell days, but not anymore. Skylake Core i3 basically matches matches it in Cinebench MT, beats it in 3DMark Physics and comes real close in x264 encoding where the hexa-core should have a very large advantage (hexa-core vs dual-core with HT). Have you seen the videos? Huge difference in framerates in some titles, especially minimum framerates. Intel's per-core performance advantage is huge right now.

Yep. Its a no brainer to pick the Skylake i3. 2 cores is more than 6 cores.

 

[jihe]

Oh right, they "force" you. Because you rather want everyone to pay for your overclocking desires. All about you, you and you.

What's next, 1 CPU SKU? 1 GPU SKU? 1 SSD size, 1 Memory size?

You also forget an overclocking motherboard would always cost more. Because just like Intel, Mobo makers dont work pro bono on a charity concept.

You know it actually costs Intel MORE to locked down and cripple cpus... I wouldn't mind paying for extra features, but it's paying for not having perfectly functional parts gimped that pisses me off.

 

[ShintaiDK]

You know it actually costs Intel MORE to locked down and cripple cpus... I wouldn't mind paying for extra features, but it's paying for not having perfectly functional parts gimped that pisses me off.

So you say segmentation cost them more? Oh I await the documentation. Here I thought it was to hit the optimal price and performance points for all segments.

 

[jihe]

So you say segmentation cost them more? Oh I await the documentation. Here I thought it was to hit the optimal price and performance points for all segments.

Segmentation itself has a cost. The fact that Intel is able to make even more from market segmentation does not change this. How would you feel about a vineyard mixing vinegar into their wines just to create segmentation?

 

[ShintaiDK]

Segmentation itself has a cost. The fact that Intel is able to make even more from market segmentation does not change this. How would you feel about a vineyard mixing vinegar into their wines just to create segmentation?

You can buy cheap and expensive wine cant you? Same with cars and close to about everything else.

Some people will only pay 50$, they get a Celeron. Others are willing to pay 350$. And some willing to pay 1000$.

 

[jihe]

You can buy cheap and expensive wine cant you? Same with cars and close to about everything else.

Some people will only pay 50$, they get a Celeron. Others are willing to pay 350$. And some willing to pay 1000$.

Of couse you can buy cheap wine. I have no problem with Intel selling different products like native dual/quad/hex cores for segmentation. But you don't see vineyard deliberately spoiling their wines, and that what Intel's been doing to perfectly good cpus by fusing off certain features after they've been produced. The only reason Intel is able to get away with it is they have a near monopoly.

 

[ShintaiDK]

Of couse you can buy cheap wine. I have no problem with Intel selling different products like native dual/quad/hex cores for segmentation. But you don't see vineyard deliberately spoiling their wines, and that what Intel's been doing to perfectly good cpus by fusing off certain features after they've been produced. The only reason Intel is able to get away with it is they have a near monopoly.

If it was due to monopoly. We wouldn't see other companies doing it. Yet they do.

Also it happened way before monopoly times too.

 

[Nothingness]

I agree some segmentation certainly makes economical sense.

But instruction set segmentation (such as AVX) is utterly stupid and counter-productive.

 

[ShintaiDK]

But instruction set segmentation (such as AVX) is utterly stupid and counter-productive.

That I fully agree on.

 

[einhoernchen]

Because the 6700K is not available I'm considering buying a 5820K instead. Given that current-gen consoles (PS4, XB1) have 8-core CPUs multicore might become more important in the next years.

Does Broadwell-E feature any major improvements worth waiting 4 additional months?

Considering the platform changes in Skylake, is there any chance 2011-3 boards will be Skylake-E compatible?

 

[frozentundra123456]

I agree some segmentation certainly makes economical sense.

But instruction set segmentation (such as AVX) is utterly stupid and counter-productive.

True, but most users dont care about AVX, or even know if it is enabled or not.

I see not being able to use fast ram with anything but a top end motherboard a much, much bigger issue. This ram issue pretty much invalidates the already paltry performance gains we have seen on the desktop for everything but the K models, which of course will be paired with a Z170 motherboard and fast ram anyway.

 

[frozentundra123456]

Yep. Its a no brainer to pick the Skylake i3. 2 cores is more than 6 cores.

It is also more expensive, requires a top end motherboard, fast ram, and cant be overclocked like the FX6300.

I have always considered the i3 awkwardly priced, and this motherboard and ram issue just magnifies the problem. The "no brainer" decision is that since you have to buy a Z170 motherboard and fast ram to properly utilize the i3, you might as well spend the extra hundred bucks for a 6600k, get 2 more real cores, and a chip that can actually, you know, utilize the overclocking features you are paying for on the motherboard.

 

[Nothingness]

True, but most users dont care about AVX, or even know if it is enabled or not.

Most users don't care about floating-point It's not because you don't know about something or think you don't care that it doesn't matter.

Not having AVX on a vast majority of the CPU means developers don't even try to use it efficiently. It can even create software incompatibilities, or severe discrepancies in performance from CPU to CPU.

 

[ShintaiDK]

It is also more expensive, requires a top end motherboard, fast ram, and cant be overclocked like the FX6300.

I have always considered the i3 awkwardly priced, and this motherboard and ram issue just magnifies the problem. The "no brainer" decision is that since you have to buy a Z170 motherboard and fast ram to properly utilize the i3, you might as well spend the extra hundred bucks for a 6600k, get 2 more real cores, and a chip that can actually, you know, utilize the overclocking features you are paying for on the motherboard.

Not at all. H110/B150+i3 and you got a cheap, yet powerful machine with a very low power usage and all the modern platform goodies.

Gaming wise I am not even sure an overclocked FX6300 is enough.

 

[MrTeal]

Segmentation itself has a cost. The fact that Intel is able to make even more from market segmentation does not change this. How would you feel about a vineyard mixing vinegar into their wines just to create segmentation?

If a vineyard could produce essentially as much top quality wine as they wanted, I would expect them to do just that. There's only so many people you can charge $1000 per bottle for wine, and there's only so many people who will pay $1000 (or even $350) for a CPU. If you're only willing to pay X dollars for B level of performance even though you could pony up Y dollars for A level, it doesn't really matter that the underlying materials are capable of a higher level.

 

[kimmel]

Segmentation itself has a cost. The fact that Intel is able to make even more from market segmentation does not change this. How would you feel about a vineyard mixing vinegar into their wines just to create segmentation?

What segmentation are we talking?

Are we talking about product line segmentation?
Atom,Core,Xeon

Nothing really overlaps there so I'm assuming you don't mean that.

Are we talking about binning?
6600 vs 6700?

Frankly Intel would be ecstatic to be able to sell every part they make as the highest end part. You may be overestimating their ability to do so.

Talking about electrical and functional binning ultimately what you are asking is even more segmentation based on individual per part custom binning so that every part is priced exactly to the manufactured part functionality and performance curve so that you don't "get vinegar" into your more generically binned parts. Now that would be an even more expensive process but you would get exactly what you pay for with less variation. SKU availability would be even more of an issue.

 

[frozentundra123456]

Segmentation is one thing, but we are talking here about the ram issue. Basically for all the desktop chips except the K models, you have to buy a top of the line motherboard with overclocking features that you cant use on the cpu just in order to utilize ram fast enough to utilize the true potential of the chip. This is more than segmentation, it is gimping the whole line-up below the K models or forcing the consumer to buy a needlessly expensive motherboard.

 

[Space69]

Not having AVX on a vast majority of the CPU means developers don't even try to use it efficiently.

And most developers doesn't even use SSE efficiently - 98% rely on compilers and couldn't care less. - they won't even look at ISPC. I do agree that AVX (and specifically AVX512) would have been great on all CPUs.

 

[IntelUser2000]

Skylake-U GT3e (15W) shouldn't be far from Geforce GT940M level of performance. Instead of Skylake-U GT2 (15W) + Low-end dGPU I can see some OEMs going this route.

Don't be so sure. It's mostly a mobile benchmark which likely inflates some strength of the GT3e configuration versus GT2. Now that's a speculation on my part, but on the 45/47W Iris Pro's the average gain is 40%. 40% on a 47W part with 3W eDRAM. What do you think will happen when 2-3W eDRAM goes on a 15W part? I'd say 40% will be the high range. Wait for more games to be tested.

Also, practically no OEMs went Iris Pro 5200 or Iris Pro 6200. When you fail twice, you stop believing their claims.

Less power, cheaper cooling/battery, thinner/lighter and probably less expensive as well? As a whole this will probably cost more than bread and butter GT2 models but less than CPU + dGPU solutions.

Let's see, how many design wins on a Iris Pro 5200/6200? Almost nothing.

Haswell Iris Pro 5200:
-Battery life on par with discrete parts, when the discrete GPU was enabled. Unfortunately for Intel, discrete parts had decent switching system which allowed integrated part to run, so a iGPU + discrete combo beat Iris Pro standalone parts in battery life!
-Anandtech review of Iris Pro 5200 stated that OEMs didn't want to use it because it was not perf/$ competitive against discrete competition! Clearly Intel positions Iris parts to make money
-On Maxwell parts the Nvidia chips were so competitive that it had better perf/watt than Iris integrated parts
-And one can never forget the always worse than competition drivers

On the basis of the facts above, I am not so hopeful about Iris the 3rd. Stop promising and start delivering Intel!

 

[Sweepr]

Don't be so sure. It's mostly a mobile benchmark which likely inflates some strength of the GT3e configuration versus GT2.
Now that's a speculation on my part, but on the 45/47W Iris Pro's the average gain is 40%. 40% on a 47W part with 3W eDRAM. What do you think will happen when 2-3W eDRAM goes on a 15W part? I'd say 40% will be the high range. Wait for more games to be tested.

Do you have a source for Skylake-U 64MB eDRAM power consumption? Can't simply extrapolate from different products.
Also 60% faster than the top HD Graphics 520 score is nothing to cough at given that these are the very first submissions so far. Honestly much better than your 30-40% faster prediction if they can sustain that kind of performance with proper cooling in Ultrabook-style devices.

Also, practically no OEMs went Iris Pro 5200 or Iris Pro 6200. When you fail twice, you stop believing their claims.

We all know those parts were targeted at Apple, PC OEMs were better served by regular Haswell/Broadwell + dGPUs. And they succdeded, Apple chose them instead of Skylake-S for entry-level iMacs due to much better graphics performance, good enough to dicth a dGPUs in their view (not saying I agree with this).

-Anandtech review of Iris Pro 5200 stated that OEMs didn't want to use it because it was not perf/$ competitive against discrete competition! Clearly Intel positions Iris parts to make money
-On Maxwell parts the Nvidia chips were so competitive that it had better perf/watt than Iris integrated parts
-And one can never forget the always worse than competition drivers

Totally different comparison, these are not 45-47W 22nm Gen 7.5 parts but 15W 14nm chips with Gen 9 graphics. And power consumption is much more critical here. If Intel can deliver anywhere close to Skylake-U GT2 + Geforce GT940M gaming performance with Skylake-U GT3e then they will have a very interesting product in their hands. It could very well enable low-end dGPU performance inside very thin, light mobile devices like the Surface Pro 4 (where you won't find dGPUs).

On the basis of the facts above, I am not so hopeful about Iris the 3rd. Stop promising and start delivering Intel!

I'm personally more excited about this chip than even Skylake GT4e. To me, having top notch iGPUs makes more sense in low-power mobile chips than places where you could fit a dGPU anyway.

 

[knutinh]

And most developers doesn't even use SSE efficiently - 98% rely on compilers and couldn't care less. - they won't even look at ISPC. I do agree that AVX (and specifically AVX512) would have been great on all CPUs.

Having a "unified" instruction set would benefit purchasers of the expensive units as well. More software would be optimized for the fancy instructions, and performance would probably be better on a high end cpu for generic software.

This all assumes that Intel can somehow practically/economically shave off performance in its low-end products while keeping the same relative performance for different code variants ("optimizing code for i3 will automatically optimize code for i7"). A simple way to achieve approximately this would be lower clocks on low-end products.

They could also have "AVX512" support in low-end cpus that behind the scenes was implemented on top of a narrower 256, 128 or even 64-bit wide vector machine. This might give AVX/Intel a bad name, but would allow them to put different (cheaper) hw into low-end offerings while keeping basic instruction level parity. Looking at ARM Neon, it seems that they did something similar (offer wider vector instructions that offer negligible speedup compared to narrower instructions). By using the wide version you get a small decrease in code size, and the possibility of speedup if/when hardware choose to actually implement all of those adders and mults etc in parallell.

-k

 

[khon]

The first benchmark results for the m3-6y30 are starting to show up now.

https://browser.primatelabs.com/geekbench3/3978280

Looks like a 10-20% gain in Geekbench compared to the m5y10. Will be interesting to see results for other benchmarks as well, but I haven't been able to find any yet.

 

[Nothingness]

And most developers doesn't even use SSE efficiently - 98% rely on compilers and couldn't care less. - they won't even look at ISPC. I do agree that AVX (and specifically AVX512) would have been great on all CPUs.

Many developers don't even know what options their compilers have, frightening...

Having a "unified" instruction set would benefit purchasers of the expensive units as well. More software would be optimized for the fancy instructions, and performance would probably be better on a high end cpu for generic software.

This all assumes that Intel can somehow practically/economically shave off performance in its low-end products while keeping the same relative performance for different code variants ("optimizing code for i3 will automatically optimize code for i7"). A simple way to achieve approximately this would be lower clocks on low-end products.

They could also have "AVX512" support in low-end cpus that behind the scenes was implemented on top of a narrower 256, 128 or even 64-bit wide vector machine. This might give AVX/Intel a bad name, but would allow them to put different (cheaper) hw into low-end offerings while keeping basic instruction level parity. Looking at ARM Neon, it seems that they did something similar (offer wider vector instructions that offer negligible speedup compared to narrower instructions). By using the wide version you get a small decrease in code size, and the possibility of speedup if/when hardware choose to actually implement all of those adders and mults etc in parallell.

AVX-512 could open more widespread use of vectorization, but Intel is definitely not willing to see more use of their extensions

OTOH I wonder what kind of end user task would significantly benefit from AVX-512.

 

[Nothingness]

The first benchmark results for the m3-6y30 are starting to show up now.

https://browser.primatelabs.com/geekbench3/3978280

Looks like a 10-20% gain in Geekbench compared to the m5y10. Will be interesting to see results for other benchmarks as well, but I haven't been able to find any yet.

Nice, the Surface Pro 4 has almost caught up with the iPhone 6s :ninja: