Intel launches Haswell Xeon E5

[Arachnotronic]

What a solid piece of engineering.

 

[Idontcare]

Does thermal expansion of such a large die also have the potential to limit sizes?

Yes and no ( did you honestly expect any less of an answer? )

Thermal expansion plays havoc with many different aspects of the engineering and electrical performance (including reliability over time) of the chip.

And the larger the chip, the more of an issue that becomes from a "pin shear stress" perspective. (see slide 16)

However, in my efforts to be brief for the sake of maintaining a legible and digestible response, my concern here is that my words will leave you with the impression that it is more of a concern than it really is.

Suffice to say, yes it is something that is of large enough consequence that it requires the engineers to factor it in, but at the same time it is small enough of an absolute impact that I can't think of a specific example where it has ever limited the die size of an IC or circuit.

Not to date myself, but eons ago when I worked with SUN Sparc chips the die sizes were considerable and film delamination from thermal mismatch really was a problem. But, in restrospect, the truth of the matter was that at the time a lot of us engineers (myself included!) on that project were just ignorant (or naive) of the issues and it was simply a matter of gaining experience (which we did!) in handling silly things such as dummy fill patterns and metal density requirements to design-in some counter stresses for the whole thermal mismatch situation.

What it really boils down to is that, as a materials science engineer, you have to approach an integrated circuit as if it were a composite material for which you must balance thermal mismatch as a straightforward linear process of adjusting the relative weighting ratios of your admixtures of each component present in the composite. X amount of copper, Y amount of methylated silicon oxide, Z amount of silicon substrate, etc., etc.

Not to understate or overstate the severity of the situation, but suffice to say it is not a problem until such time that it becomes one

That on die fabric is getting seriously complex! The "cluster on die" mode is pretty interesting- just treat the two rings as separate NUMA nodes, except with extremely high speed on-chip switches between them instead of off-die QPI links.

I wonder where the fabric design will go next? Will they go back to a single, larger bidirectional ring? More "clusters" of cores on their own rings, connected in a network by switches? Or something more like Knights Landing's 2D-mesh fabric?

It really is, and yet in some ways it seems needlessly so. This is a 3D problem that has been resolved via a 2D engineering solution which results in some quirky one-off solutions IMO.

Not that I am questioning Intel's engineering decisions, they have surely ran the design through an order of 10^6 more simulated performance scenarios than I have in my wee little mind, so all I can pontificate on are the academic solutions that can be concocted without the restraint of the chip being economically viable

Which leads to Intel17's excellent observation and point:

What a solid piece of engineering.

^ to which I say "+1"

 

[zir_blazer]

I usually have a feud with other enthusiasts because most of them are mere gamers/overclockers that don't know how to truly appreciate Server parts just because they are inferior for these purposes. I was already impressed by 15 Core Ivy Bridge-EX, this 18 Core Haswell-E is even more amazing. I could fall in love with such silicon. Makes you think what Intel could do at 14nm...

 

[RU482]

Dayy-um. I'm almost scared to think how many cores Broadwell-EP is likely to be packing. 24, anyone?

What's worse is seeing the Opterons just get pulverized in this test. I mean I wouldn't expect it to even come close to beating Haswell-EP anyway, given that at the high-end Intel now has more cores than AMD has modules... but it looks like AMD's best Opterons are still only about competitive with Westmere, to say nothing of the three Intel cores that have come and gone since then.

EDIT: Someone should install Windows 7 on a system with two E5-2699 v3 chips, just to see what its task manager looks like on a system with 72 logical cores!

I've done it with a dual-deca core E5-v2 setup. It was bananas. I can't imagine how many pages of CPUs there would be with almost 2x that!

 

[Arachnotronic]

I usually have a feud with other enthusiasts because most of them are mere gamers/overclockers that don't know how to truly appreciate Server parts just because they are inferior for these purposes. I was already impressed by 15 Core Ivy Bridge-EX, this 18 Core Haswell-E is even more amazing. I could fall in love with such silicon. Makes you think what Intel could do at 14nm...

I wonder if Haswell-EX will have an even higher core count or if 18 is as far as Intel's going to push it on 22nm.

 

[zir_blazer]

I wonder if Haswell-EX will have an even higher core count or if 18 is as far as Intel's going to push it on 22nm.

There will be no Haswell-EX. Xeon E7 platform skips a generation, so it will be Broadwell-EX.

 

[kimmel]

There will be no Haswell-EX. Xeon E7 platform skips a generation, so it will be Broadwell-EX.

I wouldn't be so sure.

 

[Techhog]

The thing I found impressive is the market that buys these things essentially move 100% into new generation 6-9 months from product introduction. That's big contrast to consumer market where even 3 years after people might still be buying Sandy Bridge based products.

You also have to keep in mind that time is money, so every little boost helps.

 

[Burpo]

http://www.extremetech.com/computin...-will-try-to-preempt-the-arm-server-onslaught

"With its huge set of new SKUs and capabilities, Intel has launched a series of parts that dares up-and-coming players to try and match them."

 

[IntelUser2000]

I wouldn't be so sure.

Supposedly its still coming. The rumors put it at 20-22 cores. It seems the EX's are back at 1 year cadence.

 

[kimmel]

Supposedly its still coming. The rumors put it at 20-22 cores. It seems the EX's are back at 1 year cadence.

With the 18C die that big do you really think they are going to do a 20+ core sku? Also how would they do layout of the extra cores?

 

[IntelUser2000]

With the 18C die that big do you really think they are going to do a 20+ core sku? Also how would they do layout of the extra cores?

http://vr-zone.com/uploads/17046/INTC_Brickland.jpg

http://wccftech.com/intel-brickland...nerations-ivy-bridgeex-haswellex-broadwellex/

http://techreport.com/news/26911/errata-prompts-intel-to-disable-tsx-in-haswell-early-broadwell-cpus

It won't be a big deal to add 2-4 more cores considering how small they are. Haswell EP is at 662mm2. The biggest Intel chip ever was Poulson at 699mm2.

 

[ashetos]

So, does anybody know how this compares to IBM POWER8?