Intel Broadwell-K & Skylake (non-K) desktop CPUs to launch in Q2-2015

[SAAA]

Are broadwell-based Xeons out of the picture?

Quad cores yes, we should have them now instead there's a Haswell refresh, but the EP/EX Xeons should be released with no delays. They come one year after the consumer parts so the 14nm delay doesn't touch them, probably.

 

[SAAA]

In short, we should expect Broadwell to overclock very close to 5GHz on average. Extrapolations are fun.

This was enough convincing as extrapolation, it makes sense too if they are upping base clocks on unlocked CPUs, otherwise there wouldn't be any margin left for overclocking. So 4.7-4.8GHz average overclock for Devils Canyon from 4.0GHz base, if Broadwell really reaches 5GHz they can increase of another 10% up to 4.4GHz base clock and still have an upgrade in performance without touching the architecture. Plus that would be a 4.4/3.6=22% advantage over any locked chip, if Skylake doesn't increase clock (maybe it will decrease instead?).

 

[Fjodor2001]

90nm -> 65nm: +20.5% (Pentium D 820 vs. 955 EE*)
65nm -> 45nm: +17.0% (QX6700 vs. QX9650**)
45nm -> 32nm: + 13.7% (i7 975 vs. i7-2600K***)

Good work! But I'm missing the numbers from 32nm -> 22 nm. Do you have any data on that?

 

[witeken]

Aha, I see. Then you misinterpreted it. Here is what he actually said:

In other words it totally contradicts what you stated, since it mentions Skylake production being ramped up in 2015H2. That implies an actually availability of Skylake in 2016Q2/Q3.

According to this link, it means increasing the production, so it could start earlier for desktop i5 and i7, and in H2 production will be increased for high volume parts.

Secondly, it doesn't imply availability in Q2 or Q3, but much earlier. From the start of volume production of 14nm, in March, to availability in September/October is 6-7 months.

But I actually think that he meant availability in H2, because Skylake isn't dependent on a new node, so Intel can start production whenever it wants, so it doesn't really make sense to launch Skylake (much) more than 1 year after Broadwell.

 

[witeken]

Except that OC scaling fell negative from SandyBridge 32nm to 22nm IvyBridge (even removing the IHS) and Haswell didnt improve over Ivy if im not mistaken.

And i dont want to sound like a pessimist but since Intel is pursuing the Mobile market, im expecting less and less OC scaling generation to generation.

Homeles already explained that the curve of 22nm changed because of FinFET (positive for mobile, slightly negative for high-end), but 14nm should see a regular improvement over 22nm as other nodes.

 

[AtenRa]

Homeles already explained that the curve of 22nm changed because of FinFET (positive for mobile, slightly negative for high-end), but 14nm should see a regular improvement over 22nm as other nodes.

Manufacturing process is not the only metric that affects frequencies.

 

[SAAA]

Manufacturing process is not the only metric that affects frequencies.

Yes with the other things architecture matters a lot, same 65nm:

4477MHz average Pentium 955 EE (source HWbot)

3597MHz average QX 6700

So clocks can go down and they did with Haswell over Ivy on the same process, about 6% using i7 values. On top of that they decreased of the same 6% from Sandy to Ivy, but this should be an exception due to 22nm and its FinFET.

 

[Homeles]

Manufacturing process is not the only metric that affects frequencies.

There are two primary reasons 22nm regressed:
Switch to FinFETs.
Switch to thermal paste exacerbated preexisting gap between IHS and die.

At this point I've done enough research on the subject to write a master's thesis on it...

Good work! But I'm missing the numbers from 32nm -> 22 nm. Do you have any data on that?

Hwbot numbers: 5085MHz 2600K, 4725MHz 3770K
32nm -> 22nm: -7.1% (2600K vs 3770K)

Those numbers weren't included because the change in transistor shape, in addition to the departure from soldering, masked any benefits of an improved process; they're not relevant. The FinFET penalty is only "felt" once, so 14nm will continue to benefit from traditional scaling (or more accurately, "equivalent scaling").

 

[witeken]

At this point I've done enough research on the subject to write a master's thesis on it...

Maybe you should do that 😎.

 

[Fjodor2001]

The FinFET penalty is only "felt" once, so 14nm will continue to benefit from traditional scaling (or more accurately, "equivalent scaling").

So since we'll get Haswell Devil's Canyon at 4.0 GHz / 4.4 Ghz (Turbo) on 22 nm, we should eventually be able to see ~4.5 GHz / ~5 GHz (Turbo) CPUs based on Broadwell 14 nm? If we're assuming the frequency increase due to node shrink trend would continue, and we ignore the 22 nm FinFET penalty.

And they should be able to OC to ~6 GHz?

 

[Homeles]

So since we'll get Haswell Devil's Canyon at 4.0 GHz / 4.4 Ghz (Turbo) on 22 nm, we should eventually be able to see ~4.5 GHz / ~5 GHz (Turbo) CPUs based on Broadwell 14 nm? If we're assuming the frequency increase due to node shrink trend would continue, and we ignore the 22 nm FinFET penalty.

And they should be able to OC to ~6 GHz?

I'm highly suspicious of the claimed 4.0 base / 4.4 turbo from the rumors. There's not enough information to suggest it'd clock that high. Going back to solder would allow Haswell to OC a couple hundred MHz... there's no way that change alone would catapult Haswell up to 4.4 GHz turbo.

There's mention of "updated packaging materials," and the only thing that comes to mind that could allow that kind of performance boost is a huge overhaul... like when the industry switched from ceramic substrates to organic substrates. The only thing I can think of is Intel's BBUL that they were touting back in the P4 era, but shelved for a later date (still being worked on, if you check patent filings). It's apparently compatible with LGA sockets... which is interesting.

Another theory is that they could optimize the packaging specifically for running at high voltages and clocks... but that's a shot in the dark.

Broadwell should OC to ~5GHz, on the assumption that it's soldered and isn't combined with any crazy packaging technology (i.e., assuming the 4.4 turbo rumor is false). I think 6GHz would be somewhere around a 25% increase in clock speed? Boy, wouldn't that be something? 😛

 

[Idontcare]

Except that OC scaling fell negative from SandyBridge 32nm to 22nm IvyBridge (even removing the IHS) and Haswell didnt improve over Ivy if im not mistaken.

And i dont want to sound like a pessimist but since Intel is pursuing the Mobile market, im expecting less and less OC scaling generation to generation.

Remember the switching speed vs. voltage curves cross-over at higher voltages for 32nm and 22nm because of the change in xtor design.

The occurrence of that cross-over point is a one-time deal and won't be repeated with 14nm vs 22nm.

I'm highly suspicious of the claimed 4.0 base / 4.4 turbo from the rumors. There's not enough information to suggest it'd clock that high. Going back to solder would allow Haswell to OC a couple hundred MHz... there's no way that change alone would catapult Haswell up to 4.4 GHz turbo.

The numbers are no less credible than when we first got speculative leaks on AMD's 4.7GHz FX-9590.

And what made the FX-9590 possible? An increased TDP, higher voltages, and binning.

IMO the rumors for Devil's Canyon are credible if we allow for the possibility that Intel is going to raise TDP for the SKU (possibly up to 140W).

 

[TreVader]

It will not be a 4.4 turbo. How many haswell Ks can even clock 4.4 consistently? Zero chance of that kind of boost in clockspeed.


Very interested in surface pro 3. I will probably get it despite windows and horrible DPI scaling.

Really hoping apple finally builds something similar. They could do it easily, they just like making a ton of money on $700 iPads and would rather charge people for more memory on an ARM device vs having to engineer something like SP3.

 

[Techhog]

So since we'll get Haswell Devil's Canyon at 4.0 GHz / 4.4 Ghz (Turbo) on 22 nm, we should eventually be able to see ~4.5 GHz / ~5 GHz (Turbo) CPUs based on Broadwell 14 nm? If we're assuming the frequency increase due to node shrink trend would continue, and we ignore the 22 nm FinFET penalty.

And they should be able to OC to ~6 GHz?

That was debunked. It's 3.7/4.1

 

[Idontcare]

It will not be a 4.4 turbo. How many haswell Ks can even clock 4.4 consistently? Zero chance of that kind of boost in clockspeed.

Prior to the existence of the FX-9590 the same argument could have been (and was) made that there was little chance of a stock FX piledriver chip being able to turbo to 5GHz based on the lack of consistent overclocks above 4.8GHz with the 8350 at that time.

But we came to know differently.

And so, discounting Intel's ability to do similar seems premature.

 

[jpiniero]

Right, but the rumored TDP is 88 W. That just seems unrealistic with 4.0/4.4.

 

[jj109]

Right, but the rumored TDP is 88 W. That just seems unrealistic with 4.0/4.4.

Asus seems pretty convinced though.

 

[Alatar]

4.4GHz turbo has already been confirmed multiple times:

 

[TreVader]

How is that possible when people often can't hit 4.2 on haswell?

Didn't think they could get that big of an improvement in their process in such a short time.

 

[Alatar]

People who can't hit 4.4GHz let alone 4.2 on haswell are incompetent. The OCing problems have been exaggerated by people who were used to bumping up vcore and raising multi on SB and IB. HSW requires more fine tuning due to the IVR.

Also max turbo volts have been bumped by 150mV or so over the 4770K.

I wouldn't be surprised if these ship without coolers just like the E chips.

 

[Idontcare]

How is that possible when people often can't hit 4.2 on haswell?

Didn't think they could get that big of an improvement in their process in such a short time.

The difference between knowing how something like this is possible versus knowing how something like this is impossible is the basis of many a highly compensated engineering positions existing within the industry.

Some would say this is comparable to rocket science, others would (rightly) say IC design and process node development is a step beyond rocket science.

Don't let yourself become convinced that such things are impossible just because you might not have the education or experience to differentiate between magic and science when it comes to IC's.

In any event, as a spectator to this industry you really should dial down your own expectations of what can or cannot be done based on your personal estimations and leave the "that's impossible" exclamations about rocket science to the rocket scientists IMHO.

 

[StinkyPinky]

Man I really hope the 4790k comes out in June. Even at stock that's a nice performing CPU.

 

[AtenRa]

According to HWbot.org

Core i7 2600K Average OC (air) = 5085MHz
Core i7 3770K Average OC (air) = 4725MHz
Core i7 4770K Average OC (air) = 4497MHz

Both 3770K and 4770K are made by the same 22nm process and so the design dictates the OC.
Since Broadwell is targeting the mobile and this time the very LOW POWER tablets, i dont expect it will OC more than IB/HW.

edit: as for Devils canyon, they could make it with 4GHz base and 4.4GHz turbo. They raised the TDP to 88W from 84W, most HW CPUs will run at 4GHz with default voltage. Also, 4.4GHz will only be with a single thread, so no problem there.
One more thing, the second CPUz picture could be fake, CPUz 1.68.0 shouldn't be able to read 88W on that CPU. CPUz Version 1.68.0 was released 3-4 months ago.

 

[cbn]

How is that possible when people often can't hit 4.2 on haswell?

Didn't think they could get that big of an improvement in their process in such a short time.

Maybe this new Devil's Canyon is not using the "mobile optimized 22nm xtor" from the initial Haswell batch and instead is using a more mature and perfected 22nm xtor that was originally used in Ivy Bridge?

 

[Fjodor2001]

I have to say that Homles's theory about the one time FinFET frequency penalty appears sound.

However that is just one aspect of it. As we all know there are other metrics that determine what clock speed is achievable too, for example the uArch design chosen and what types of chips the process tech is optimized for. As AtenRa pointed out, Intel's 14 nm process could be optimized more for mobile than desktop if anything.

Then we have the interesting observation that Intel will be releasing the 14 nm Broadwell chips in strict order of increasing frequency and TDP, starting with Broadwell-Y, then Broadwell-U, and then finally ending with Broadwell-H/K around 9 months later. That is an unusually long release period. It's also unusual to release all the low frequency parts first. Possibly this can be an indication that Intel is still perfecting the 14 nm process to allow for higher clock speeds.

Summing all of this up, I think it's quite hard to predict what clock speeds we can expect from Broadwell. It's also frustrating that Intel hasn't given any indication despite that we are only months away from release of Broadwell-Y. I wouldn't be surprised if Intel will not disclose any info until IDF in September, i.e. just before Broadwell will actually be released.