Just realised this sounds incredibely sarcastic, its a genuine nice article. Hopefully we will see some more sales of these parts like the one that happened last year. Would love to play with this beast. (and AVX512 is also nice to have)
Bleh, mental mixup. Thanks for the heads up :) I'm wondering how easy it would be to run our Linux testing suite on the processor part and if it's just as plug as play as we do now. I suspect not, but I can hope.
Well the PCIe ones need extra libraries for openMP IIRC, but one of the major points of the socketed one is that you should in theory not need anything, as its just an "ordinary" CPU, with a DMI link to the chipset, so (again in theory) any x86 OS should work on it.
It would be very interesting seeing benchmarks over some vector data for AVX2 vs 512 too understand more about its implementation.
Should be plug and play as long as you don't use an ISA subset that isn't supported.
And FYI, I think you have the configurations a bit off. Pretty sure the OPI variant of KNL can be socketed. In fact I'm pretty sure that's the only option for that one as the on package KNL/OPI link takes up the majority of the PCIe lanes. Its also fairly obvious that the socket you are showing has the specific keepaways for the OPI variant as well.
You are right, I didn't think of that. Seems like it requires an internal ribbon connector as a result though - I don't think they'd have a physical place to slot it in as well as dealing with a socket.
yeah, it is a ribbon connector into the phy cage. AFAICT, they are still using the 100G/4x25G signalling/phy standard (just like Bull is with their new 100G interconnect as well and really for that matter just like EDR IB).
When ever I read something about KNL I get the feeling of wanting a pack of those cores as co-processors in my i7 instead of a beefier GPU. My Skylake GPU is doing 24/7 number crunching, but it's a lot more complicated than with even AMD GPUs, because - frankly - nobody cares about writing and maintaining that openCL software. Let alone dealing with the bugs in Intels driver (e.g. Einstein@Home still doesn't work on Skylake at all). If KNL cores were really as simple to use they'd be better suited than a GPU to what I want to do with them.
We've had a few images end up without attribution this past week. Given the lack of (tech) media here at SC15, I just want to make sure we can be attributed if they get shared around. Trying to make it both visible but not overpowering, but it's a balancing act. Thoughts welcome :)
Perhaps just a bump effect, with a little bit of blur would look better, that way it fits well in most settings, and isnt restrictive, while also being hard to remove, and noticable
I don't think it detracts at all, and whatever you need to do to avoid being ripped off, is just fine in my book. There is a reason Anandtech is still the best tech site (no matter what the clowns that comment on every Apple or Samsung thread say about being bought and paid for).
I mean I know it, obviously, but I'm not good at explaining stuff, so maybe somebody who is good at explaining bleeding edge geekery could Ehm... Let other people know?
You know, just in case somebody is lurking and doesn't know, but don't want to ask?
WOW 683mm^2 14nm die, amazing. Still not seen ANY >200mm dies on the latest gen processes from anyone else. Nobody else is shipping sub 20nm stuff this big, or even close. NV/AMD have never even done GPU's this big before, on any process. Only really seeing the IBM POWER and Oracle Sparc chips this big, but of course they are still all on 22/28nm fabs.
My DSLR broke a couple of weeks back, and I've been on the road since. So in retrospect, not bad for a $160 smartphone camera shot :) But I know what you mean.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
26 Comments
Back to Article
LukaP - Thursday, November 19, 2015 - link
Just noticed lots of mixups of KNL with KNC in the third paragraph. Nice article.LukaP - Thursday, November 19, 2015 - link
Just realised this sounds incredibely sarcastic, its a genuine nice article. Hopefully we will see some more sales of these parts like the one that happened last year. Would love to play with this beast. (and AVX512 is also nice to have)Ian Cutress - Thursday, November 19, 2015 - link
Bleh, mental mixup. Thanks for the heads up :) I'm wondering how easy it would be to run our Linux testing suite on the processor part and if it's just as plug as play as we do now. I suspect not, but I can hope.LukaP - Thursday, November 19, 2015 - link
Well the PCIe ones need extra libraries for openMP IIRC, but one of the major points of the socketed one is that you should in theory not need anything, as its just an "ordinary" CPU, with a DMI link to the chipset, so (again in theory) any x86 OS should work on it.It would be very interesting seeing benchmarks over some vector data for AVX2 vs 512 too understand more about its implementation.
ats - Friday, November 20, 2015 - link
Should be plug and play as long as you don't use an ISA subset that isn't supported.And FYI, I think you have the configurations a bit off. Pretty sure the OPI variant of KNL can be socketed. In fact I'm pretty sure that's the only option for that one as the on package KNL/OPI link takes up the majority of the PCIe lanes. Its also fairly obvious that the socket you are showing has the specific keepaways for the OPI variant as well.
Ian Cutress - Friday, November 20, 2015 - link
You are right, I didn't think of that. Seems like it requires an internal ribbon connector as a result though - I don't think they'd have a physical place to slot it in as well as dealing with a socket.ats - Friday, November 20, 2015 - link
yeah, it is a ribbon connector into the phy cage. AFAICT, they are still using the 100G/4x25G signalling/phy standard (just like Bull is with their new 100G interconnect as well and really for that matter just like EDR IB).Ian Cutress - Saturday, November 21, 2015 - link
http://www.anandtech.com/show/9561/exploring-intel...Might answer some of the signalling questions.
MrSpadge - Thursday, November 19, 2015 - link
Interesting stuff and nice article, indeed!When ever I read something about KNL I get the feeling of wanting a pack of those cores as co-processors in my i7 instead of a beefier GPU. My Skylake GPU is doing 24/7 number crunching, but it's a lot more complicated than with even AMD GPUs, because - frankly - nobody cares about writing and maintaining that openCL software. Let alone dealing with the bugs in Intels driver (e.g. Einstein@Home still doesn't work on Skylake at all). If KNL cores were really as simple to use they'd be better suited than a GPU to what I want to do with them.
MrSpadge - Thursday, November 19, 2015 - link
-> "... as simple to use as Intel claims..."ahcox - Friday, February 26, 2016 - link
Yep, it might be interesting to know how much of a Xeon Phi could be packed into the space taken up by integrated graphics currently: http://www.techpowerup.com/img/15-08-18/77a.jpgImSpartacus - Thursday, November 19, 2015 - link
What's up with all of the watermarks? Has Anandtech always done that?Ian Cutress - Thursday, November 19, 2015 - link
We've had a few images end up without attribution this past week. Given the lack of (tech) media here at SC15, I just want to make sure we can be attributed if they get shared around. Trying to make it both visible but not overpowering, but it's a balancing act. Thoughts welcome :)LukaP - Thursday, November 19, 2015 - link
Perhaps just a bump effect, with a little bit of blur would look better, that way it fits well in most settings, and isnt restrictive, while also being hard to remove, and noticablefanofanand - Thursday, November 19, 2015 - link
I don't think it detracts at all, and whatever you need to do to avoid being ripped off, is just fine in my book. There is a reason Anandtech is still the best tech site (no matter what the clowns that comment on every Apple or Samsung thread say about being bought and paid for).ImSpartacus - Thursday, November 19, 2015 - link
It's not bad - you're putting them towards the edges and stuff.I was just wondering if this will be a regular thing.
V900 - Thursday, November 19, 2015 - link
Does anybody know what MCDRAM is?I mean I know it, obviously, but I'm not good at explaining stuff, so maybe somebody who is good at explaining bleeding edge geekery could Ehm... Let other people know?
You know, just in case somebody is lurking and doesn't know, but don't want to ask?
Like that guy over there?
McHandle - Thursday, November 19, 2015 - link
Clearly McDonald's has gotten into the DRAM business.PEJUman - Thursday, November 19, 2015 - link
^^^Lol^^^PEJUman - Thursday, November 19, 2015 - link
All kidding aside, I am guessing MC.... = Memory Cube DRAM.Joseph Luppens - Thursday, November 19, 2015 - link
MCDRAM = "Multi-Channel DRAM"Source:
http://www.anandtech.com/show/9794/a-few-notes-on-...
extide - Thursday, November 19, 2015 - link
WOW 683mm^2 14nm die, amazing. Still not seen ANY >200mm dies on the latest gen processes from anyone else. Nobody else is shipping sub 20nm stuff this big, or even close. NV/AMD have never even done GPU's this big before, on any process. Only really seeing the IBM POWER and Oracle Sparc chips this big, but of course they are still all on 22/28nm fabs.postpwnit - Thursday, November 19, 2015 - link
You should invest in a polarizing filter to get rid of the reflection in the last shot. It would also be useful in similar shooting situations.Ian Cutress - Thursday, November 19, 2015 - link
My DSLR broke a couple of weeks back, and I've been on the road since. So in retrospect, not bad for a $160 smartphone camera shot :) But I know what you mean.FunBunny2 - Thursday, November 19, 2015 - link
How about interest from the database crowd (not Hadope and such)???III-V - Friday, November 20, 2015 - link
This is a beautiful piece of engineering. I hope the things that Intel's learned in designing this get passed down to consumers later down the road.