Skylake/Broadwell Roadmap Update @Vr-zone

[oobydoobydoo]

wccftech...
Intel Preparing Something ‘Big’ and ‘Disruptive’ with Skylake Microarchitecture

Haha.



I hope they are right, I really do. All this secrecy though... it worries me.

 

[Arachnotronic]

Haha.



I hope they are right, I really do. All this secrecy though... it worries me.

They're most likely going to tell you everything you want to know about Skylake and more at IDF in August. Relax.

 

[witeken]

wccftech...
Intel Preparing Something ‘Big’ and ‘Disruptive’ with Skylake Microarchitecture

Of course they do... (irony)

Probably another 20X reduction in idle power with a new S0_Oix sleep state. We'll see.

 

[Fjodor2001]

The reason you are seeing a relatively small advancement in ST perf/clock is because Intel has spent the last few generations re-targeting the power profile of its chips. Also, in a design like Haswell, you can get a significant boost in performance if you utilize the new instructions such as AVX2.

I think with Core at a sufficiently low power level, future improvements can target raw performance increases.

If Intel could have improved IPC more with their current uArch as base they would have. That would have benefited their low power CPUs too.

 

[Arachnotronic]

If Intel could have improved IPC more with their current uArch as base they would have. That would have benefited their low power CPUs too.

They need to improve that IPC while at the same time keeping "C" constant and power consumption flat-to-down.

Also, don't forget that Haswell's advancements focused a lot on power management, which is very tricky to get right. Even a company like Intel can't make too many changes generation/generation without risking the project's schedule.

 

[Fjodor2001]

They need to improve that IPC while at the same time keeping "C" constant and power consumption flat-to-down.

Yes, and they have not succeeded very well looking at the performance improvements of the latest generations. That's why they'll need a completely new uArch to make some real advancements going forward.

 

[SAAA]

That WCCF rumor sounds so bombastic...

Here have my rumor:

http://browser.primatelabs.com/geekbench3/1753502

This looks like it's a Skylake early sample (GenuineIntel Family 6 Model 94 Stepping 1) and it' the same as this one probably (22 January 2015):

http://www.sisoftware.eu/rank2011d/...d4e3d2e3d5e5d0f684b989afcaaf92a284f7caf2&l=en

The scores aren't half bad if the cpu has that exact speed, it's close to mobile chips with 3.2-3.4 GHz turbo, in that case we could see a decent 20/30% jump over Haswell.

Take this with a huge grain of salt, beside there are still 6 months for any sofware and hardware optimization. Example is the SHA score with no improvements when Skylake should have a new extension for it.

 

[mikk]

Example is the SHA score with no improvements when Skylake should have a new extension for it.

I thought CannonLake gets some SHA extensions: http://forums.anandtech.com/showpost.php?p=36982604&postcount=1202


Geekbench score looks very good if this is a SKL for this low clock especially MT score. Only 15% slower than a much higher clocked i7-4770K.

 

[mikk]

That WCCF rumor sounds so bombastic...

Here have my rumor:

http://browser.primatelabs.com/geekbench3/1753502

This looks like it's a Skylake early sample (GenuineIntel Family 6 Model 94 Stepping 1) and it' the same as this one probably (22 January 2015):

Yes this seems to be a Skylake engineering model. Both Device IDs are from Skylake.


Northbridge 191F --> Sky Lake Host Bridge
Southbridge A141 --> Sunrise Point-H

https://pci-ids.ucw.cz/read/PC/8086

 

[Arachnotronic]

I thought CannonLake gets some SHA extensions: http://forums.anandtech.com/showpost.php?p=36982604&postcount=1202


Geekbench score looks very good if this is a SKL for this low clock especially MT score. Only 15% slower than a much higher clocked i7-4770K.

It's gotta be that MorphCore

 

[ThatBuzzkiller]

wccftech...
Intel Preparing Something ‘Big’ and ‘Disruptive’ with Skylake Microarchitecture

I hope this isn't true and it's just another one of their garbage articles ...

AVX-512 will be just enough to last us a long time so Intel doesn't need any of the "MorphCore" crap ...

 

[mikk]

What exactly is crap? Do you have more infos about "MorphCore"? I don't understand your AVX-512 comparison (not to mention that Client Skylake most likely won't support AVX-512)

 

[PPB]

For anyone interested in what the WCCF article is talking about, here is a article about MorphCore:

PDF ARTICLE

Sounds like a solution that is intended to make CPU more apt at multithreaded, throughput-oriented loads while retaining the high IPC and low latency capabilities at running single threaded, serialized code.

EDIT 2: What I liked about the part of the article I have read is that they are finally realizing that SMT has the most gains when it is used in in-order cores. The first itinerations of Atom were a testament to this

 

[SAAA]

MorphCore is not a bad idea necessarily, just something really complicated and different from usual, at least for Intel's current path of incremental upgrades (say since Core and tick tock introduction).

Still it could explain the large improvements we see in the benches I posted, especially in multithreaded scores.
Single thread has limitations that even a completely overhauled architecture would hardly overcome and Haswell is a really high bar to surpass.

 

[ThatBuzzkiller]

What exactly is crap? Do you have more infos about "MorphCore"? I don't understand your AVX-512 comparison (not to mention that Client Skylake most likely won't support AVX-512)

I don't see any reason why the desktop version of Skylake shouldn't support AVX-512 since it has many performance benefits for a lot of consumer applications seeing as how Intel went there way to push TSX on some Haswell processors and soon for almost every Broadwell Processor plus that only benefits lock-free data structures ...

AVX-512 + TSX is an extremely effective combination for aggressively exploiting parallelism ...

If I wanted TLP I would get a GPU or a xeon phi, MorphCore would be the last thing I would think about ...

 

[Arachnotronic]

I don't see any reason why the desktop version of Skylake shouldn't support AVX-512 since it has many performance benefits for a lot of consumer applications seeing as how Intel went there way to push TSX on some Haswell processors and soon for almost every Broadwell Processor plus that only benefits lock-free data structures ...

AVX-512 + TSX is an extremely effective combination for aggressively exploiting parallelism ...

If I wanted TLP I would get a GPU or a xeon phi, MorphCore would be the last thing I would think about ...

Just a guess, but notice that CNL is expected to feature AVX-512. This implies to me that there is either a power, area, or time-to-market reason that AVX-512 was disabled on the consumer Skylake since consumer Cannonlake will have it. I don't think it's for artificial segmentation purposes.

My bet would be on it being to speed up time to market so that they don't end up with another TSX-like fiasco.

 

[krumme]

It's gotta be that MorphCore

I like that thought! That would be worth something.

 

[mikk]

I don't see any reason why the desktop version of Skylake shouldn't support AVX-512 since it has many performance benefits for a lot of consumer applications seeing as how Intel went there way to push TSX on some Haswell processors and soon for almost every Broadwell Processor plus that only benefits lo

Like AVX2 there is almost no productive application available for consumers. It mainly helped stress tester like Prime95, for Intel this was a disadvantage for Haswell in many temperature and power tests. And by the way it is Intel who confirmed that AVX-512 is Xeon only.

 

[ThatBuzzkiller]

Just a guess, but notice that CNL is expected to feature AVX-512. This implies to me that there is either a power, area, or time-to-market reason that AVX-512 was disabled on the consumer Skylake since consumer Cannonlake will have it. I don't think it's for artificial segmentation purposes.

My bet would be on it being to speed up time to market so that they don't end up with another TSX-like fiasco.

Extending support for a wider SIMD instructions isn't exactly hard since Intel has done it twice already ...

I highly doubt that AVX-512 wouldn't be included in the consumer versions since the server versions are using the exact same core micro architecture as the consumer ones so pushing it fast to the market without the consumer versions doesn't make much sense when their implementations are identical ...

 

[ThatBuzzkiller]

Like AVX2 there is almost no productive application available for consumers. It mainly helped stress tester like Prime95, for Intel this was a disadvantage for Haswell in many temperature and power tests. And by the way it is Intel who confirmed that AVX-512 is Xeon only.

The only reason why we aren't reaping the benefits of AVX2 yet is because it takes time to update the applications to support it ...

 

[Idontcare]

It's gotta be that MorphCore

It'll help the itanium oems transition to a post-itanium world

 

[oobydoobydoo]

I bet its HBM.


Wouldn't it be possible to move all the system memory into RAM and just have no storage at all? Like what the PS4 and Xbone have.

 

[ShintaiDK]

I bet its HBM.


Wouldn't it be possible to move all the system memory into RAM and just have no storage at all? Like what the PS4 and Xbone have.

HBM is not suited for it. HMC is if you think main memory replacement.

Also systsm memory is RAM. And PS4 and Xbox One certainly got storage else they wouldnt function.

 

[kimmel]

Extending support for a wider SIMD instructions isn't exactly hard since Intel has done it twice already ...

I highly doubt that AVX-512 wouldn't be included in the consumer versions since the server versions are using the exact same core micro architecture as the consumer ones so pushing it fast to the market without the consumer versions doesn't make much sense when their implementations are identical ...

Would you pay more money for a laptop chip with AVX-512? Probably not. Would you pay more money for a server chip with AVX-512? Probably.

That's really all that needs to be said about why it looks like AVX isn't on the consumer parts. Intel has gotten pretty good about doing multiple die splits to keep die size down to drive down cost. Why not the cores too?

 

[ThatBuzzkiller]

Would you pay more money for a laptop chip with AVX-512? Probably not. Would you pay more money for a server chip with AVX-512? Probably.

That's really all that needs to be said about why it looks like AVX isn't on the consumer parts. Intel has gotten pretty good about doing multiple die splits to keep die size down to drive down cost. Why not the cores too?

Because they both SHARE the same micro architecture ?

It makes no sense to disable a perfectly functional part of a chip ...

There's more to gain by keeping that extension in the consumer versions than disabling it ...

Consumers having AVX 512 capable CPUs translates to programmers actively seeking out to support it since it is a key enabler to unlocking more performance and not only that but it can also increase revenue too because there would be more incentives to buy a CPU that supports that extension so it's a win-win scenario all around ...