Intel Unleashes Haswell-EX Xeon E7 V3 Processors

[zir_blazer]

Haswell chips for 2011v3 have a dual imc. But only ddr4 is used.

False. Check these: 1, 2, 3

The point is that only very few models seems to use it...

This motherboard only supports the DDR3 compatible CPU (E5-2669 v3, E5-2649 v3, E5-2629 v3). You cannot power on the system if you use a DDR4 compatible CPU.

...which is rather sad. Heck, these parts seems to be OEM only and not even released to the public, they even recently deleted them from Intel Ark as I recall seeing them. Its like they are actively hidding that DDR3 support is there, since they want to push DDR4 instead. This is rather similar to what happened at some point with the Pentium 3 i820 Chipset and RAMBus, just that they didn't had a true backup plan ready when they figured out than RAMBus was going nowhere and screwed it when they had to fall back, while they're currently prepared to do so anytime now.

 

[Idontcare]

Lets not turn this into a dinosaur thread, but I got to be a master at taking the the RLL controllers off bad 30mb Seagates and swapping them to the 20mb mfm Seagate drives & low leveling it for 30megs. Must have done 50 or 60 of those. Yeah, things have come quite a ways since -debug days. 5.6 billion transistors on chip is hella impressive to an old timer like me
https://books.google.com/books?id=x...mands for 30mb seagate rll controller&f=false

Don't kid yourself, it is hella impressive even to the folks who designed the damned thing that consumes 5.6 billion transistors.

The only sentient species not impressed by such pinnacles of humankind's CMOS achievements are the aliens that gifted us said alien technology in the first place, and the random forum goer who truly seems incapable of drawing even a single NOR logic circuit let alone enough of them to consume a 5.6 billion xtor project budget

 

[Arachnotronic]

Don't kid yourself, it is hella impressive even to the folks who designed the damned thing that consumes 5.6 billion transistors.

The only sentient species not impressed by such pinnacles of humankind's CMOS achievements are the aliens that gifted us said alien technology in the first place, and the random forum goer who truly seems incapable of drawing even a single NOR logic circuit let alone enough of them to consume a 5.6 billion xtor project budget

I wonder what kind of yields Intel is getting on these bad boys. Intel is selling SKUs with all sorts of different core counts...from 4 all the way to 18...based on the same die.

 

[ShintaiDK]

I wonder what kind of yields Intel is getting on these suckers. Intel is selling SKUs with all sorts of different core counts...from 4 all the way to 18...based on the same die.

The 4 is the exception, due to high ST performance requirement. Else the next one starts at 10 cores that also looks binned. Its first at 14 it really says defects.

But it looks more like binning than yield in most cases.

 

[mrmt]

I wonder what kind of yields Intel is getting on these bad boys. Intel is selling SKUs with all sorts of different core counts...from 4 all the way to 18...based on the same die.

4 cores with 140W at only 3.2GHz base clock is way bad. I think it's a very niche SKU, maybe to build some validation boxes for testing RAS features?

 

[Arachnotronic]

4 cores with 140W at only 3.2GHz base clock is way bad. I think it's a very niche SKU, maybe to build some validation boxes for testing RAS features?

Pretty expensive chips for validation boxes

 

[Idontcare]

The 4 is the exception, due to high ST performance requirement. Else the next one starts at 10 cores.

But it looks more like binning than yield in most cases.

It's all yield, parametric versus functional, but I know what you meant. Just wanted to clarify for the rest of the audience.

I doubt Intel has much of a D0 (functional yield) issue, even with their 600+ mm^2 dies, but the parametric yield is probably the dominating issue for them. Particularly in terms of clockspeed versus power-consumption (itself dominated by the clockspeed vs. Vcc shmoo distribution).

 

[mrmt]

Pretty expensive chips for validation boxes

If you need to test RAS features, what choice do you have?

 

[imported_ats]

4 cores with 140W at only 3.2GHz base clock is way bad. I think it's a very niche SKU, maybe to build some validation boxes for testing RAS features?

4 cores at 140W but 6-12TB memory capable, which matters for lots of DB workloads. Low cost licensing for large scale in memory DBs.

 

[DrMrLordX]

Still well behind Power8 - at least at sap

http://download.sap.com/download.ep...B9777C2576CCCD47210DABDC85175C7935E1E53D78A37

http://download.sap.com/download.ep...A8415F3B32C9914852A34134AB2756D168C18C90EE4DB

Xeon had 80% more cores

Power8 still 36% faster

Pardon my ignorance, but I can't get either one of those URLs to work (tried Waterfox and Spartan/Edge). How does one access such content?

 

[Burpo]

It's not your ignorance. Amazing how people can speculate when NOBODY has had access to this new chip.

 

[TuxDave]

Pardon my ignorance, but I can't get either one of those URLs to work (tried Waterfox and Spartan/Edge). How does one access such content?

Not sure why it opens for me but:

IBM Power Enterprise System E870, 8 processors / 80 cores / 640 threads,
POWER8, 4.19 GHz, 32 KB (I) and 64 KB (D) L1 cache and 512 KB L2
cache per core, 8 MB L3 cache per core, 2048 GB main memory

vs

Fujitsu PRIMEQUEST 2800E2, 8 processors / 144 cores / 288 threads,
Intel Xeon Processor E7-8890 v3, 2.50 GHz, 64 KB L1 cache and
256 KB L2 cache per core, 45 MB L3 cache per processor,
2048 GB main memory

And then a bunch of numbers but I think IBM is winning.

 

[DrMrLordX]

It's not your ignorance. Amazing how people can speculate when NOBODY has had access to this new chip.

Apparently, Fujitsu has . . .?

Not sure why it opens for me but:

IBM Power Enterprise System E870, 8 processors / 80 cores / 640 threads,
POWER8, 4.19 GHz, 32 KB (I) and 64 KB (D) L1 cache and 512 KB L2
cache per core, 8 MB L3 cache per core, 2048 GB main memory

vs

Fujitsu PRIMEQUEST 2800E2, 8 processors / 144 cores / 288 threads,
Intel Xeon Processor E7-8890 v3, 2.50 GHz, 64 KB L1 cache and
256 KB L2 cache per core, 45 MB L3 cache per processor,
2048 GB main memory

And then a bunch of numbers but I think IBM is winning.

Hmm okay. I would like to know what the net power draw is for both of those machines during testing.

 

[thunng8]

Not sure why it opens for me but:

IBM Power Enterprise System E870, 8 processors / 80 cores / 640 threads,
POWER8, 4.19 GHz, 32 KB (I) and 64 KB (D) L1 cache and 512 KB L2
cache per core, 8 MB L3 cache per core, 2048 GB main memory

vs

Fujitsu PRIMEQUEST 2800E2, 8 processors / 144 cores / 288 threads,
Intel Xeon Processor E7-8890 v3, 2.50 GHz, 64 KB L1 cache and
256 KB L2 cache per core, 45 MB L3 cache per processor,
2048 GB main memory

And then a bunch of numbers but I think IBM is winning.

You can browse all results here:

http://global.sap.com/solutions/benchmark/sd2tier.epx
The power8 results are from last year so you have to go to 2nd or 3rd page.

It is basically 79750 vs 58626 users

POWER8 36% higher on per chip basis and an amazing 2.45x faster on per core basis. Those Power8 cores make the Haswell core look very slow on this workload.

 

[Burpo]

I don't see much difference especially since one is using Windows Server 2012 & the other Red Hat Enterprise Linux.

 

[imported_ats]

You can browse all results here:

http://global.sap.com/solutions/benchmark/sd2tier.epx
The power8 results are from last year so you have to go to 2nd or 3rd page.

It is basically 79750 vs 58626 users

POWER8 36% higher on per chip basis and an amazing 2.45x faster on per core basis. Those Power8 cores make the Haswell core look very slow on this workload.

Or 40% slower per thread. All depends on how you want to look at it. SAP 2D tends to have a lot of idle time on a per context basis due to latency of storage and cache misses. Xeon would likely show near linear gains from SMT2 -> SMT4.

 

[thunng8]

That's statement is laughable. Xeon was not designed for 4 threads and trying to shoehorn something that it wasn't designed for would likely decrease performance.

Ibm has given estimated scaling from 1 thread to 8 threads per core in their performance reports. I am taking the 192 core e880 as an example

1: 1936
2: 2808
4: 3650
8: 3905

http://public.dhe.ibm.com/common/ssi/ecm/po/en/poo03017usen/POO03017USEN.PDF

Going from 4 threads to 8 threads give only a small increase

Also going by those numbers, power8 running in single thread would still be faster than Xeon

 

[imported_ats]

That's statement is laughable. Xeon was not designed for 4 threads and trying to shoehorn something that it wasn't designed for would likely decrease performance.

Ibm has given estimated scaling from 1 thread to 8 threads per core in their performance reports. I am taking the 192 core e880 as an example

1: 1936
2: 2808
4: 3650
8: 3905

http://public.dhe.ibm.com/common/ssi/ecm/po/en/poo03017usen/POO03017USEN.PDF

Going from 4 threads to 8 threads give only a small increase

Also going by those numbers, power8 running in single thread would still be faster than Xeon

all depends on the workload and what the limiters are. Some workloads scale amazingly well with SMT contexts, others not so well.

 

[DrMrLordX]

Okay, it is interesting to see how Xeon is performing here, at least in this one benchmark. The systems, as configured, show that POWER8 wins with an even socket count. I still would like to know, what are the individual power draw numbers for both of these machines?