256 Threads - 1 CPU? Yes please!

[Patrick]

Here is the single CPU Intel Xeon Phi x200 (7210) system with 256 threads. You can also see the 6x 16GB DIMMs as well as the 8x 2GB MCDRAM packages:


I will post a short video this week on some of the initial setup of the machine.

 

[OBasel]

h-o-l-y _______________________ and add a few more.

 

[RlngT]

Good Gracious Father .___________________________>

 

[T_Minus]

wowzaaaaaa

use cases?

 

[Patrick]

wowzaaaaaa

use cases?

I have a feeling come Jan 1 many more! Intel MKL for machine learning will be out.

The big use case is AVX512 which right now more or less requires icc and folks working in HPC applications.

 

[Hank C]

password cracking? lol

 

[Patrick]

Ok so getting slightly better results now - Monero Mining Performance

Single chip outpaces dual E5-2698 V4 and there is a ton more room for optimization compiler wise.

 

[gbeirn]

Any memory segmentation? How does the hierarchy with the MCDRAM work?

 

[Patrick]

Any memory segmentation? How does the hierarchy with the MCDRAM work?

There is a cache mode (think giant L3 cache of 16GB), Flat Mode where the 16GB is added to system memory that is addressable which you can see in the above screenshot. You can also do a hybrid mode which works as a combination.

The fun part is that there are 5 cluster modes available plus an auto setting.

 

[gigatexal]

Silvermont cores right? They could in theory be used for virtualization? Maybe docker containers?

 

[Patrick]

Silvermont cores right? They could in theory be used for virtualization? Maybe docker containers?

Docker does work. I fired up the Monero mining container on the KNL. Easy.

Virtualization is not supposed to work actually (and I did not see the flags for it.) The BIOS does have a VT-d setting set to on which is strange.

 

[gigatexal]

Damn that's nice. I think you could get a ton of docker centric use-cases from this then.

 

[Jon Massey]

Why would you want to play games on such a massively parallel system?

 

[Patrick]

@Nanotech single threaded general purpose compute on these things is not great. After yesterday I think games will "run" but will choke on low single core throughput.

Even compile benchmarks are choking at some point.

 

[i386]

Why would you want to play games on such a massively parallel system?

[url=http://vizdoom.cs.put.edu.pl/competition-cig-2016]"Ai/machine learning" [/url]

 

[Maritime]

How will it deal with 'standard' RAM when program needs more memory than 7210 has? This situation might be a bottleneck for system?

Any ideas where/how to test performance on specific problem?

 

[MiniKnight]

How will it deal with 'standard' RAM when program needs more memory than 7210 has? This situation might be a bottleneck for system?

Any ideas where/how to test performance on specific problem?

The various MCDRAM modes allow a few different options. You can use standard RAM but then use the MCDRAM as a cache just an example here.

I was reading up on this last week. Maybe we can get some benchmarks of the modes?

 

[Maritime]

I am interested in testing KNL on my test files. It could be much better idea then 4 processor machine for almost same investment.

 

[Patrick]

Likely in the first week or two of Jan we will get an opportunity to run select other workloads.

 

[gigatexal]

@Nanotech for things like that I like hardwarecanucks and thetechreport and guru3d to name a few. They often do CPU benchmarks with games with different GPU If a game performs about the same with a GTX1060 as it does a GTX 1080 then it is CPU limited. That said most games, as far as I know, don't scale past about 4 cores.