Name of Icelake successor leaks out

[Arachnotronic]

Also, regarding Intel's new Tick, Tock, Tock cadence. Isn't the "new release every year", largely driven by PC OEM demands, to have new models for every back-to-school and Holiday season?

Just like AMD's and NVidia's GPU re-brands for OEMs. Even when they didn't have new models on 28nm node, to release... they "did" for OEMs.

It is absolutely driven by OEMs' desires to have new products out every year. I am sure more than one OEM was furious when they had to deal with "Haswell Refresh" in 2014.

Intel will also likely face more potent competition than it has in a while once AMD rolls out Zen, so it can't afford to not make yearly architectural enhancements even if their process guys still can't get it together.

 

[ElFenix]

I am not familiar with what happened back then with Dell and all of that, but I do know Intel's products were being smacked around silly by AMD's.

http://www.wsj.com/articles/SB10001424052748704067504575305003810938096


http://archive.fortune.com/2010/07/27/news/companies/dell_settlement_intel.fortune/index.htm

 

[Arachnotronic]

http://www.wsj.com/articles/SB10001424052748704067504575305003810938096


http://archive.fortune.com/2010/07/27/news/companies/dell_settlement_intel.fortune/index.htm

Much obliged!

 

[superstition]

By implementing this CRAP methodology, AMD hopes that it can take full advantage of buzz-phrases such as "Clean Sheet Design" and the ever popular, "It Can't Get Much Worse than This" sentiment among enthusiasts.

"We'll get a really nice revenue boost in years when we drop a Transformative Unique Re-Design, or TURD, in the marketplace. This is great for investors because it means that our stock price will predictably plunge in the year following the TURD launch as the product gets a bit dry and frankly stale," Oron said with a grin, struggling to hold back laughter.

There is a lot of reason to be "pro-Intel" when we are talking about CPUs. They make truly excellent products ...

There is a difference between liking a product and posting potty humor about a competitor's.

AMD on the other hand does not make great processors for the vast majority of enthusiast desktop PC buyers right now and hasn't for years. So it should probably be more puzzling to you to see "enthusiasts" who are rabidly pro-AMD even though AMD has done a poor job serving them over the last five years or so.

I think some people are still pleased that AMD makes x86 CPUs at all (following Intel's bribing of OEMs and such which cost AMD a lot of money), particularly the FX line which does provide adequate performance, considering the pricing, for some workloads. The existence of these CPUs provides some competition and also, when talking about the APUs, has made it easier to port and develop console games. The $100 8320E with $50 board combo from Microcenter still gets a better score in Cinebench multi, POV-ray, and so on than a more expensive Skylake i5. All ones needs to do is use a better cooler and overclock to reasonable levels. The chips are old and power hungry but since AMD used solder and gave them unlocked multipliers (something you can't get from Intel for $100) they still hold their own on a value basis for some workloads.

As for the lake naming, I like icelake as a code name.

 

[Arachnotronic]

There is a difference between liking a product and posting potty humor about a competitor's.

All intended to be in good fun! Hey, maybe you didn't find it funny, and that's cool, different strokes for different folks, humor is subjective, and so on

 

[ShintaiDK]

Stupid question but is power consumption under legacy loads half of IB with Skylake with all the instruction sets enabled or do you have to disable them in the bios (I remember reading somewhere that you disabled them and saw a large power drop) to see the power savings?

The reason I ask is that most of the reviews I've read have painted the picture that skylake isn't all that energy efficient. I'm a 3770k @4.3ghz owner and am looking to upgrade sometime in the next year or two. This machine is on 24/7 so idle or close to idle power savings are important for me from a long term cost perspective. I'm curious what kind of power reduction I'd see going to skylake and if there would be any power reduction at all going to Broadwell-e.

Everything is enabled, its just regular loads like gaming for example with little to no usage of these instructions. Nothing disabled. The only thing that can make A 6700K reach its TDP is AVX2/FMA3 very heavy loads. But performance is also very different under these loads. While its much easier to get SB/IB to reach its TDP in regular loads.

This is also why only servers get AVX512. Its too big a TDP burden for end user products. Without having any usage to offset the TDP penalty.

For you, idle usage wont change much. So if that's the primary reason then you wont see any worthwhile benefit.

 

[superstition]

Everything is enabled, its just regular loads like gaming for example with little to no usage of these instructions. Nothing disabled. The only thing that can make A 6700K reach its TDP is AVX2/FMA3 very heavy loads. But performance is also very different under these loads. While its much easier to get SB/IB to reach its TDP in regular loads.

This is also why only servers get AVX512. Its too big a TDP burden for end user products. Without having any usage to offset the TDP penalty.

For you, idle usage wont change much. So if that's the primary reason then you wont see any worthwhile benefit.

Perhaps things like AVX512 might be spun off to a support chip to make cooling easier. Having everything creating heat in such a small space causes problems, especially with the polymer TIM.

 

[ShintaiDK]

Perhaps things like AVX512 might be spun off to a support chip to make cooling easier. Having everything creating heat in such a small space causes problems, especially with the polymer TIM.

I doubt the changes in Cores between server and desktop/mobile have anything to do with the TIM only desktop uses. But rather the TDP needed.

Servers can also lower the clock easier as some already do for AVX2/FMA3. Imagine a 4Ghz desktop CPU only running AVX512 at 3Ghz. The drama, despite being much faster.

 

[witeken]

Hell of an expensive back-up plan.

Okay, so here's the thing.

What makes you think Tigerlake is 10nm?

Look for a pattern here:

32nm: Sandy Bridge
22nm: Ivy Bridge

22nm: Haswell
14nm: Broadwell

14nm: Skylake
10nm: Cannonlake

10nm: Icelake
?nm: Tigerlake

 

[ShintaiDK]

You forgot the 14nm Kaby Lake.

I agree Tigerlake looks to be 10nm.

 

[NTMBK]

Perhaps things like AVX512 might be spun off to a support chip to make cooling easier. Having everything creating heat in such a small space causes problems, especially with the polymer TIM.

You could add some texture units and call it a GPU

 

[Fjodor2001]

Let me guess,

You want power consumption tested with AVX2/FMA3. But you only want performance directly tested in regular/legacy, while ignoring power consumption.

Did I miss any?

So you were wrong as usual, 6700K is not twice as fast as 3770K as you claimed. And it was exposed immediately when you were asked for for data to back up your claims which you of course did not have. Instead of admitting that you try to shift goalposts. Why am I not surprised.

 

[ShintaiDK]

So you were wrong as usual, 6700K is not twice as fast as 3770K as you claimed. And it was exposed immediately when you were asked for for data to back up your claims which you of course did not have. Instead of admitting that you try to shift goalposts. Why am I not surprised.

I think you should check your own posts and see what you claimed. Specially in the performance/watt metric.

How much perf/watt have we seen on desktop? Regressing from 77W to 95W for the top end SKUs from IB to SKL. That while only seeing minuscule performance improvement.

The only way you can reach your flawed position using IB as baseline is to ignore actual performance at the power consumption.

Lets take CB11.5 as an example.
Unless proven otherwise we can assume that the stock 3770K uses 77W.
http://forums.anandtech.com/showthread.php?t=2443441

At 77W it scores 7.61. My 6700K at ~58W scores 9.66. 6700K in this example is 68.5% better in performance/watt. And your miniscule performance improvement at that lower power consumption reached 27%.

Now I could show you some AVX2 vs the IB that would beat the IB plain silly. I assume this was common knowledge due to the move to 256bit paths with Haswell and Skylake and the performance boost it gave.

 

[Arachnotronic]

Okay, so here's the thing.

What makes you think Tigerlake is 10nm?

Look for a pattern here:

32nm: Sandy Bridge
22nm: Ivy Bridge

22nm: Haswell
14nm: Broadwell

14nm: Skylake
10nm: Cannonlake

10nm: Icelake
?nm: Tigerlake

It's 10nm.

 

[dave_the_nerd]

Last big IPC jump was PPro.

Really? No love for Core/Core2 and/or Nehalem?

 

[ShintaiDK]

Really? No love for Core/Core2 and/or Nehalem?

Compared to 60%+ those was minor

 

[IntelUser2000]

Compared to 60%+ those was minor

CPU power consumption of Pentium Pro was nearly double, made worse by the on package L2 cache. Also, it was a server chip and its 16-bit performance sucked. Perhaps its better comparing to Itanium, which for its native performance was pretty decent too.

Also process gains back then were much better. But more recently, 20% for a new architecture was pretty average. Pentium III to Pentium M was 30%, Core Duo to Core 2 Duo was 20%, Core 2 Duo to Sandy Bridge was 20%.

Nowadays, we get nothing in process, nothing in architecture, no room for power increases, no room to make it cheaper, while making the chips get more expensive, and the future process technologies keep getting delayed(EUV for one) and with less improvement but with more effort. I highly doubt those exotic CNT, "silicon replacement", germanium, III-IV transistors bring more than Tri-Gate did. It'll probably take heroic efforts to implement them to get less improvements than ever(maybe they'll claim 50% at 0.6V which is useless for high performance chips).

Let's analyze this:

32nm: Sandy Bridge
22nm: Ivy Bridge - 5% gain

22nm: Haswell - 10% gain
14nm: Broadwell - 5% gain

14nm: Skylake -7% gain
14nn: Kabylake - 5% gain?
10nm: Cannonlake - 5% gain?
10nm: Icelake - 5% gain?
?nm: Tigerlake - 5% gain?

Perhaps an identical latter half naming suggests performance improvement will be quite minimal.

 

[ShintaiDK]

Also process gains back then were much better. But more recently, 20% for a new architecture was pretty average. Pentium III to Pentium M was 30%, Core Duo to Core 2 Duo was 20%, Core 2 Duo to Sandy Bridge was 20%.

IPC? No....

Core Duo to Core 2 Duo was mainly single cycle SSE. And even then you dont get 20% average. The actual real life average was more around 5-10% back then.

Core 2 Duo to SB is Conroe->Penryn->Nehalem->Westmere->Sandy bridge.

Pentium-M had multiple uarchs. So did Pentium-3 that would affect performance.

T2400 vs T5600.

 

[ehume]

Speaking of patterns:

45nm: Penryn/Wolfdale (tick)
45nm: Nehalem (tock)
45nm: Lynnfield (tock)
45nm: Lynnfield uncapped multiplier (tock)

32nm: Westmere (tick)
32nm: Sandy Bridge (tock)

22nm: Ivy Bridge (tick)
22nm: Haswell (tock)
22nm: Devil's Canyon (tock)

14nm: Broadwell (tick)
14nm: Skylake (tock)
14nm: Kaby Lake (tock)

10nm: Cannonlake (tick)
10nm: Icelake (tock)
10nm? Tigerlake (if 10nm, would it be another tock? or a new tick?)

It seems that only in Westmere-Sandy did we have a single tick-tock sequence. The rest have been tick-tock-tock or worse. The 2-year tick-tock seems never to have been realized.

 

[dark zero]

I am wondering... There will be Kabylake-E? Yeah since those situaciones of Tick-Tocks are like that.. And knowing that Kabylake is just a little more like a Skylake rebrand...

 

[tenks]

Speaking of patterns:

45nm: Penryn/Wolfdale (tick)
45nm: Nehalem (tock)
45nm: Lynnfield (tock)
45nm: Lynnfield uncapped multiplier (tock)

32nm: Westmere (tick)
32nm: Sandy Bridge (tock)

22nm: Ivy Bridge (tick)
22nm: Haswell (tock)
22nm: Devil's Canyon (tock)

14nm: Broadwell (tick)
14nm: Skylake (tock)
14nm: Kaby Lake (tock)

10nm: Cannonlake (tick)
10nm: Icelake (tock)
10nm? Tigerlake (if 10nm, would it be another tock? or a new tick?)

It seems that only in Westmere-Sandy did we have a single tick-tock sequence. The rest have been tick-tock-tock or worse. The 2-year tick-tock seems never to have been realized.

This is wrong. I don't think you understand what tick-tock is. A tock is a new uARCH, a tick is a shrink+enhancements to the tock core.

Conroe (tock)
Penryn (tick)

Nehalem (tock) Lynnfield is nots its own tock, its apart of the nehalem family. I won't even comment on Lynnfield unlocked being it's own tock, lol?
Westmere (tick)

SandBridge (tock)
IvyBridge (tick)

Haswell (tock) Haswell Refresh isn't its own tock, as it was just a higher binned speed of an existing chip/core.
Broadwell (tick)

Skylake (tock)
Kaby Lake (tick on same node) This is different then Haswell refresh as there will be core changes and just just a higher bin
Cannonlake (tick)

I am wondering... There will be Kabylake-E? Yeah since those situaciones of Tick-Tocks are like that.. And knowing that Kabylake is just a little more like a Skylake rebrand...

if I were to guess, no. Because of long development times of HEDT, they dont seem to be as affected as much as the DT skus. We didn't get a haswell-e refresh, just a delay in broadwell-e so haswell-e was in market longer. I think the same thing will happen when skylake-e launches. Broadwell-E will be around into Kabylake DT. And then we will get Skylake-E

 

[CHADBOGA]

So if Intel has been struggling to provide enough 14nm processors because they take longer to produce because of the extra patterning involved, how does Intel address this problem?

Are they bringing another 14nm factory online?

 

[Arachnotronic]

So if Intel has been struggling to provide enough 14nm processors because they take longer to produce because of the extra patterning involved, how does Intel address this problem?

Are they bringing another 14nm factory online?

The problem isn't cycle time, which is what you are referring to, but actual yields. They have capacity in place to build X chips at Y yields. If yields are significantly smaller than Y, then you wind up with a shortage.

A brute force way to cope with poor yields is to simply put in more capacity and run more wafers, but this is a problem for 2 reasons:

1. Putting in more capacity simply costs more. It also takes a long time to build out more capacity so you need to be aware of the problems well in advance.
2. When your yields come up, you end up with far more capacity than you need, meaning excess capacity. This is bad because those factories/equipment need to be depreciated and if they are not actually being used to produce anything they just hurt the bottom line.

Also those factories are staffed with people, you need to keep the lights on, etc so it's just a huge expense.

The way to handle it is exactly what Intel seems to have done, deal with the shortages and try to satisfy demand for lower priority customers with 22nm silicon until 14nm is in good shape.

 

[ShintaiDK]

So if Intel has been struggling to provide enough 14nm processors because they take longer to produce because of the extra patterning involved, how does Intel address this problem?

Are they bringing another 14nm factory online?

They already did it seems.

So D1X, D1D, D1C, Fab24 and Fab32 is making 14nm chips now. Fab42 was originally to open and make 14nm chips too.

 

[PPB]

Talking about future launches, bits and chips confirmed 100% FIVR is back with cannonlake. Yay for the return of cheap OCable motherboards!