Intel Skylake / Kaby Lake

[LTC8K6]

I don't believe Intel could have launched a 6C/12T Skylake chip a couple years ago.

Or a 6C/12T Kabylake chip one year ago.

Intel was unable to produce enough i7 Skylake chips at release as it was. It was several months after launch before supplies were reasonable to where you could expect to be able to find a 6700K when you went shopping. Skylake is barely two years old.

The power was a problem for the desktop socket at the time.

I think Intel could not reasonably do the 6C chips for desktop until 14nm++ came along.

14nm+ was too small an improvement to change the 6C scenario, imo.

Plus, Intel knew that 14nm++ was coming and that it would be better to try a 6C chip then.

 

[TahoeDust]

More benchmarks means more data.

If somebody would volunteer to compile a table of this latest benchmark comparison run, that would be very convenient while trying to follow what's going on with all the scores.

https://docs.google.com/spreadsheets/d/19hyFLgt1KIgBTxUEzac4BbXKtqUniDH9zBO4kGwMCp4/edit?usp=sharing

 

[raghu78]

They stayed at 4 cores for mainstream, but they kept throwing in larger iGPUs, integrated additional technologies, improved the media engine significantly, and so on.

These were all features that directly improved the user experience for mainstream PC users far more than throwing in a couple of extra cores would've.

Anyway, I'm glad Intel is going to 6+2 for mainstream, and I hope they will push to even higher core counts and throw in more features too, that's all great stuff for the consumer.

If I can get a 10 core mainstream Ice Lake or Tiger Lake, I'd be really happy

This die size table summarizes what happened over the last 8 years

http://www.anandtech.com/show/9505/skylake-cpu-package-analysis

45nm Lynnfield 4C - 296 sq mm (Graphics on motherboard chipset)
32nm Sandybridge 4C (GT2) - 216 sq mm (First on die graphics Intel CPU)
22nm Ivybridge 4C(GT2) - 160 sq mm
22nm Haswell 4C(GT2) - 177 sq mm
14nm Skylake 4C(GT2) - 122 sq mm

So even though Intel added graphics and video decode die size shrunk at every process node significantly. Intel had ample opportunity to provide more cores. The fact that they did not came down to 2 things - 1.) Lack of competition 2.) Intel's focus on gross margins. Now that competition is back core counts will go up and Intel's gross margins will go down.

 

[Arachnotronic]

Now that competition is back core counts will go up and Intel's gross margins will go down.

Thermal considerations will put a lid on how quickly mainstream core counts will go up. There are simply laws of physics that chipmakers have to contend with.

Why do you think AMD isn't rushing to stick two CCXes onto Raven Ridge?

 

[raghu78]

Thermal considerations will put a lid on how quickly mainstream core counts will go up. There are simply laws of physics that chipmakers have to contend with.

Why do you think AMD isn't rushing to stick two CCXes onto Raven Ridge?

I understand all cpu designs are governed by thermal considerations. But the fact that die size of Intel Skylake 4C is 56% that of Sandy bridge 4C shows how much Intel has used the lack of competition to improve margins. As for AMD Raven Ridge they are governed by die size . AMD would ideally prefer a die size around 200 sq mm and with a 11CU Vega GPU a 4 core CCX is what they could fit.

 

[Arachnotronic]

I understand all cpu designs are governed by thermal considerations. But the fact that die size of Intel Skylake 4C is 56% that of Sandy bridge 4C shows how much Intel has used the lack of competition to improve margins. As for AMD Raven Ridge they are governed by die size . AMD would ideally prefer a die size around 200 sq mm and with a 11CU Vega GPU a 4 core CCX is what they could fit.

Intel's margins haven't improved at a corporate level since the release of Sandy Bridge.

In fact, Intel's gross margin percentage is down since 2011.

Tell me again how AMD's absence from this market has allowed Intel to increase its margins?

 

[Zucker2k]

I understand all cpu designs are governed by thermal considerations. But the fact that die size of Intel Skylake 4C is 56% that of Sandy bridge 4C shows how much Intel has used the lack of competition to improve margins. As for AMD Raven Ridge they are governed by die size . AMD would ideally prefer a die size around 200 sq mm and with a 11CU Vega GPU a 4 core CCX is what they could fit.

So 91watts for the upper tier mainstream is not high enough of a tdp?

 

[Charlie22911]

It wasn't about "quad core milk strategy" at all.

These chips are designed first and foremost for notebooks, and it wasn't really feasible until 14nm++ to increase the core count while staying within a reasonable power envelope.

I agree to an extent, but I'd counter with thuban.

 

[raghu78]

Intel's margins haven't improved at a corporate level since the release of Sandy Bridge.

In fact, Intel's gross margin percentage is down since 2011.

Tell me again how AMD's absence from this market has allowed Intel to increase its margins?

That chart is inspite of Intel wasting billions of dollars on things like Mcafee acquisition, mobile contra revenue.

 

[Zucker2k]

I agree to an extent, but I'd counter with thuban.

125watts vs 95watts.

 

[Arachnotronic]

That chart is inspite of Intel wasting billions of dollars on things like Mcafee acquisition, mobile contra revenue.

Money spent on acquisitions doesn't impact the gross profit margin line -- acquisitions are one-time uses of cash and don't usually impact the P&L (aside from amortization of acquired intangibles or, if those acquisitions are busts, write-downs, which are usually one-offs in nature).

Contra-revenue had an impact on gross profit margin (since it was a direct deduction from revenue even though the associated costs were in the COGS line), but that's pretty much done and yet Intel still hasn't seen its margins hit the peak levels that it did in 2010.

There's a lot that goes into driving the gross margin of a product line/overall corporate gross profit margins beyond chip die size.

 

[dullard]

Intel is having an Optane question and answer session over on Reddit:
https://www.reddit.com/r/intel/comments/6tryjq/ama_with_intel_optane_memory/

So, I asked the main question that peaked my interest regarding "Next Gen Optane Memory support" on the Z370 boards. I didn't get a direct answer (because companies never like to disclose new information outside of their sanctioned launches), but at least I didn't get a complete dismissal of potential new Optane features either. Intel's answer: "The Z370 chipset launch and disclosures will introduce the new features which will be supported with Optane Memory."
https://www.reddit.com/r/intel/comments/6tryjq/ama_with_intel_optane_memory/dlpn5zp/

 

[Markfw]

It seems that the CPU speed and SMT mode had nothing to do with the score. CPU@3400 5700 something score, CPU@4100, exact same score. I tried everything. But when I messed with the mhz of the video card I got up to 2000 and got this:

 

[IntelUser2000]

That chart is inspite of Intel wasting billions of dollars on things like Mcafee acquisition, mobile contra revenue.

I think margins are not gained because the PC market has declined.

Anyway I believe they should sacrifice margins to improve products. They focus too much on margins. 50% margin would be fine, if the revenue is 20% greater than today. Intel's Net Profit sucks even with those amazing margin numbers.

They are #1 by far in R&D spend. Many more companies do more than them in actual products and money generated. What's wrong with them?

 

[PeterScott]

Money spent on acquisitions doesn't impact the gross profit margin line -- acquisitions are one-time uses of cash and don't usually impact the P&L (aside from amortization of acquired intangibles or, if those acquisitions are busts, write-downs, which are usually one-offs in nature).

Contra-revenue had an impact on gross profit margin (since it was a direct deduction from revenue even though the associated costs were in the COGS line), but that's pretty much done and yet Intel still hasn't seen its margins hit the peak levels that it did in 2010.

There's a lot that goes into driving the gross margin of a product line/overall corporate gross profit margins beyond chip die size.

Lots of reasonable answers in this thread. I have one more.

It is perfectly reasonable for companies to design products to maximize their profits. Intels strategy was to target the mainstream with 4 cores + iGPU, and AMD's strategy with Zeppelin, was to leave out the APU and go all cores.

Both are geared toward maximizing profits, and both are reasonable.

 

[deathBOB]

This die size table summarizes what happened over the last 8 years

http://www.anandtech.com/show/9505/skylake-cpu-package-analysis

45nm Lynnfield 4C - 296 sq mm (Graphics on motherboard chipset)
32nm Sandybridge 4C (GT2) - 216 sq mm (First on die graphics Intel CPU)
22nm Ivybridge 4C(GT2) - 160 sq mm
22nm Haswell 4C(GT2) - 177 sq mm
14nm Skylake 4C(GT2) - 122 sq mm

So even though Intel added graphics and video decode die size shrunk at every process node significantly. Intel had ample opportunity to provide more cores. The fact that they did not came down to 2 things - 1.) Lack of competition 2.) Intel's focus on gross margins. Now that competition is back core counts will go up and Intel's gross margins will go down.

Each process is more expensive than the last. Shrinking size doesn't necessarily mean shrinking cost.

 

[IntelUser2000]

but that's pretty much done and yet Intel still hasn't seen its margins hit the peak levels that it did in 2010.

Margins were high in 2010 because 2010 and 2011 were the years when their revenue increased by 20-30% each year. When the market starts shrinking the decisions made to reduce the effect likely impacts it in terms of margins.

Each process is more expensive than the last. Shrinking size doesn't necessarily mean shrinking cost.

Considering the raw cost of the die is a fraction of the selling price, it can easily be offset by better market positioning.

If in the Haswell generation they had dies that cost $15 to make but were all made into Celerons it would be a loss compared to say a Skylake generation that had a cost structure of $35(2.3x the cost) but were made into Core i3s.

Of course not everything is that clear cut. But as they moved through generations they were able to make them into more dies. In the Pentium III generation they literally had one die to scale from a $100 Celeron chip to a $1000 enthusiast one. Features were merely disabled, whether due to yield or market demand.

Westmere generation had:
-2C + memory controller and GPU die

Sandy Bridge generation had:
-2C + GT1 die
-2C + GT2 die
-4C + GT2 die

Skylake generation has:
-2C + GT2 die
-2C + GT3 die
-4C + GT2 die
-4C + GT4 die

(They claim there's 5 dies total. If there is, I assume the Y-chips have its own die, omitting the space required for eDRAM support)

I'm just pointing out the non-HEDT client segment. Having different die means they can optimize the cost for the segment needed to reduce the impact of increased cost due to a more expensive process.

 

[TahoeDust]

It seems that the CPU speed and SMT mode had nothing to do with the score. CPU@3400 5700 something score, CPU@4100, exact same score. I tried everything. But when I messed with the mhz of the video card I got up to 2000 and got this:

Interesting that it did not respond to CPU core speed changes. If you feel like running it at the 1080p High setting or any of the other benchmarks, I would add them to the chart. The more data, the clearer the picture should become.

https://docs.google.com/spreadsheets/d/19hyFLgt1KIgBTxUEzac4BbXKtqUniDH9zBO4kGwMCp4/edit?usp=sharing

 

[Phynaz]

I understand all cpu designs are governed by thermal considerations. But the fact that die size of Intel Skylake 4C is 56% that of Sandy bridge 4C shows how much Intel has used the lack of competition to improve margins.

It may not show what you think it does, but it does show that thermal density has nearly doubled.

 

[Markfw]

Interesting that it did not respond to CPU core speed changes. If you feel like running it at the 1080p High setting or any of the other benchmarks, I would add them to the chart. The more data, the clearer the picture should become.

https://docs.google.com/spreadsheets/d/19hyFLgt1KIgBTxUEzac4BbXKtqUniDH9zBO4kGwMCp4/edit?usp=sharing

Based on my results, I think you should add video card mhz to the chart, as I am now convinced for this benchmark, thats all that matters. My CPU was also virtually not running during the test, it only got to 53c. When straining at 4.1, it goes over 90c

 

[Java Sunday]

After looking at the slides from the ice lake preview, it looks to me like 14++ will be the highest performing process until 10++. Maybe I'm taking the graph too literally, obviously there are other factors in play (performance is not really a scalar metric).

http://images.anandtech.com/doci/11722/kaizad-mistry-2017-manufacturing(1)_29.png

 

[nvgpu]

https://www.fool.com/investing/2017/08/16/intel-corporation-is-probably-going-to-kill-this-p.aspx

 

[TahoeDust]

Based on my results, I think you should add video card mhz to the chart, as I am now convinced for this benchmark, thats all that matters. My CPU was also virtually not running during the test, it only got to 53c. When straining at 4.1, it goes over 90c

All the other were run at 2025-2038. Will your 1080 ti run that?

 

[Markfw]

All the other were run at 2025-2038. Will your 1080 ti run that?

No, I saw somebody with 2012 got 6100 or something, but 2000 was the best it would do, and I think that was throttling. I saw 2038, and after a minute, it went down to 2000 (XOC precision was running)

 

[IntelUser2000]

After looking at the slides from the ice lake preview, it looks to me like 14++ will be the highest performing process until 10++. Maybe I'm taking the graph too literally, obviously there are other factors in play (performance is not really a scalar metric).

http://images.anandtech.com/doci/11722/kaizad-mistry-2017-manufacturing(1)_29.png

They did say both 10nm and 10nm+ would be outperformed by 14nm++. It is a graph, it is literal.