Intel Grantley-EP platform, comprised of a "Wellsburg" Platform Controller Hub and a large number of "Haswell-EP" Xeon E5-2600 v3 microprocessors, is expected to launch in the third quarter of this year. We already reported on new features of Haswell-EP core, and last month we published several Xeon E5-2600 v3 model numbers, spotted in one of Intel's Product Change Notifications. Since then we located several other sources (5 in total), that listed some characteristics of forthcoming processors. Merging all specs together, we created a list of all possible third generation E5 models and their features, including operating frequency, the number of cores, the size of L3 cache and TDP. [...]
Thats an odd change to previous. So I would doubt that Haswell-EP gets 18 cores.
But you are quoting a reference source of wccftech, which I find to be a VERY unreliable source of information.
Did you miss the vr-zone watermark?
Original source:
http://vr-zone.com/articles/broadwe...s-reliability-per-core-performance/66845.html
Also more to add:
This platform is interesting for another reason. As the Haswell-EP Xeon range (from which Haswell-E desktop variant is derived) has three main die flavours 10-core, 14-core and 18-core
Sorry my mistake then
This platform is interesting for another reason. As the Haswell-EP Xeon range (from which Haswell-E desktop variant is derived) has three main die flavours 10-core, 14-core and 18-core
Uh, no. The flavors are 8, 12, and 18.
See this thread for a 12c chip delid.
Apparently you can buy a Haswell-EP Xeon E5 V3 CPU along with server grade Intel LGA 2011-3 socketed motherboard on eBay:
http://www.overclock.net/t/1504701/ebay-haswell-e-motherboard-and-cpu-appear
http://wccftech.com/server-grade-in...rd-haswell-ep-xeon-e5-v3-processor-sale-ebay/
How's that motherboard for pure tech porn... 24 DIMM slots...
I'm confused by the point you are trying to make.
The previous generation, has the following diagram:
So the die configuration will give a maximum number of cpu cores, available on the die(s).
But a particular SKU, can have a number of cores LESS than that, because some of the (possibly faulty) cores can be disabled (fused off, or whatever).
Are you trying to say that because the ES (Engineering Sample) is 12 cores, ALL SKUs have to have 12 cores ?
There are going to be something like 3 different die configurations, so I would not "read" too much into the thread that you linked, as it is just ONE example of a die configuration. As I understand it.
I don't think VR-zone is necessarily trustworthy or not... It's more likely that Intel didn't want to release 18-core EP chips, so the slide deck presented earlier has them saying it'll only go to 14 cores on EP. Which is why we have the PCN adding the 16 and 18 core variants.
On the other hand, the article quoted said there would be 10, 14, and 18 core die variants... afaik that is not the case (it's 8, 12, and 18), which is what I wanted to point out. The author on VR-Zone is making a rather interesting guess on the die variants. We've seen the 12-core variant via the other thread, so yeah...
Also, Ivy Bridge EP chips come in 6, 10, and 15 core variants. I don't honestly believe that there is truly a 12-core variant as suggested in Intel's slide deck. EX and EP share a lot in common that makes me suspect 12-core EP chips are simply cut down EX chips with the 3rd QPI link reallocated (used as PCIe x8 on EP) and no external memory buffer support (EP uses direct-to-RAM, EX uses memory buffers). The stepping family (they are both in the ABCD stepping family) suggests that IVB-EP-12c and IVB-EX chips share the same die.
Note: Server CPU stepping family
ABCD - high core
_LM - medium core
_RS - low core
Thanks for the interesting and informative reply.
My original post, was only suppose to be accurate, as regards the claim that the new Haswell-EP, would go to 18-cores.
The other information (core sizes) were just a copy from the source, and had NOT been researched by me, as to their accuracy and/or agreement with other sources.
It's possible that Intel on purpose has an excessive number of (manufactured) cores, available on the die(s), to maximize the overall yield, of usefully saleable highish core count cpus.
E.g. The Playstation 3, had a spare core, so that one core could fail (manufacturing defect/yield), yet the cpu would still have the correct number of usable cpu cores.
It's possible that they originally intended for Haswell-EP to max out at 14 cores, so that the Broadwell-EP, would look impressive with 18 cores.
But they may have changed their minds, because of upcoming Arm competition, and stuff like that.
I have heard (rumours) about a new, 8 core Broadwell-DE series, so they may have increased the core counts, to help with the market segmentation, or something.
Sooner or later, one or more of the Arm suppliers may be able to come out with a high core count (e.g. 32 core) A57-Cortex cpu, which would provide interesting competition, to the Intel server chips.
One thing I have been wondering about, recently is if Intel produced 2 very similar cpu chips, would there be any/much performance between them ?
E.g. (Hypothetical)
Clock speed and Cache sizes are identical, between these 2 chips.
A one die, 12 Core, Haswell-EP, 2.1 GHz vs a (Single cpu chip package) two die, 2x6 = Still 12 cores, 2.1 GHz ?
I.e. Would the one die (12 core) chip be faster, as there would be no potentially slower communications (QPI link, or whatever?), between the 2 dies ?
The two die solution connected via QPI could be considered a sort of dual-processing on one chip. This is always slower than having the cores on die, especially with coherency traffic going over the relatively slow QPI bus (instead of the internal processor ring bus).
While it's true the Ivy 12/15c chips have two memory controllers and all smaller die versions only one memory controller, meaning the 2x6c idea is rather clean, NUMA issues would lower performance on the two die solution. Since the IMC isn't shared, some memory requests now have to go through QPI, a latency and bw penalty.
The reason you go with two 6 cores over one 12 core is pretty much because you need connectivity: two chips means 80 PCIe lanes.
Thanks, for such a good answer!
I see, so it would indeed affect the performance, which was what I was worried about. (I thought that it probably would slow it down, but hoped that it wouldn't).
So as a result of the QPI link, trying to use a pair of 6 or 8 cored Haswell-EP xeons (because the lowish frequency ones are much cheaper, than the rest of the range), in a dual processor (socket) motherboard, in order to achieve the performance of a (single socket) 12 core Haswell-EP, but at less cost.
Would NOT work out brilliantly, due to the performance delays/loss in the QPI link.
Not necessarily - in the case of two CPUs in separate sockets, you can use more memory channels (4 per chip, instead of 2 per chip in the single-chip dual-die hypothetical).
As applied to the real world where 1x12c vs. 2x6c, I'd have to change the line in my previous post about why you pick 2 6c chips to include if you want more memory (since you get 8 channels total) but not necessarily bandwidth, which does not necessarily add up because of NUMA.