6th Generation Intel Core i7/i5/i3 Mobile Processors support up to 64GB of memory

[SergeyKostrov]

*** 6th Generation Intel Core i7/i5/i3 Mobile Processors support up to 64GB of memory *** Do you know that some 6th Generation Intel Core i7/i5/i3 Mobile Processors support up to 64GB of physical memory instead of 32GB? Here are some technical details: Code Name: Skylake Lithography: 14nm Platform: Mobile ** Intel Core i7 Processors ** http://ark.intel.com/compare/88972,88970,88969,88967 Intel Core i7-6920HQ Processor ( 8M Cache, up to 3.80 GHz ) Intel Core i7-6820HQ Processor ( 8M Cache, up to 3.60 GHz ) Intel Core i7-6820HK Processor ( 8M Cache, up to 3.60 GHz ) Intel Core i7-6700HQ Processor ( 6M Cache, up to 3.50 GHz ) ** Intel Core i5 Processors ** http://ark.intel.com/compare/88959,88962 ** Intel Core i3 Processor ** http://ark.intel.com/compare/89063

[Bernard]

Do you plan upgrading your platform to newer one which supports Skylake CPU if so will you be willing to buy 64GB of RAM?

[Bernard]

Actually I bought SSD drive with the capacity of 120 GB. Page file consumes exactly 100GB and I started to test computation of  large data sets up to 10e+9 elements of type double. For that purpose I base my code on STXXL library.

[SergeyKostrov]

>>Do you plan upgrading your platform to newer one which supports Skylake CPU if so will you be willing to buy 64GB of RAM? It doesn't make sence for me to upgrade my top-end hardware from 32GB to 64GB. On IDF 2015 Intel's CEO promised 128GB physical RAM capable mobile systems and I've expected to see it. In reality Intel announced just 64GB mobile systems. It looks like something went wrong ( power problems? manufactoring problems? ).

[SergeyKostrov]

>>...I bought SSD drive with the capacity of 120 GB. Page file consumes exactly 100GB... That's a different thing and with storage there are No any problems. I need a 128GB physical RAM capable mobile workstation for in-memory processing of big data sets, like large matrices greater than 64K by 64K of elements ( that is 64K rows and 64K columns ). As soon as Virtual Memory is used performance of computations degrades significantly that is why all data needs to be in memory.

[Bernard]

Yes , it is understood pretty well that the performance will be worse when the virtual memory is in use. 

For my own purpose as I physically cannot upgrade to more than 32GB I decided to at least partially compensate over low HDD R/W speed by using 120 GB SSD. 

Why are you locked so on mobile workstations? Can not you invest in second hand WS platform where you can easily upgrade to >128 GB of RAM?

[SergeyKostrov]

>>...Why are you locked so on mobile workstations? Because of their Mobility and Portability.

[SergeyKostrov]

>>...For that purpose I base my code on STXXL library... I always test some codes against some High-Performance library. In case of linear algebra algorithms Intel MKL gives me base performance numbers and then I compare results for my algorithms with these numbers. However, when an algorithm, or processing, is very simple a verification against of pure C code could also work. Try to test STXXL vector against pure C-array. ... vector< int > vData; // set size for example 128MB int iData[ 128*1024*1024*1024 ]; ... and then do some tests...

[Bernard]

This is a good approach.

In my case usually a reference solution is provided by the Mathematica 8 routines to which my implementation or in some case a tiny C++ wrapper around Fortran 77 code is compared. Of course this is in the case of numerical quadrature, ODE solvers and special functions evaluations. 

Yes I plan to test stxxl::vector against pure C-style array and I do not expect from stxxl to be faster.

P.s

Was your intention to allocate 128GB or 128MB in static array declaration?

[SergeyKostrov]

Modern computers are so fast that in case of performance tests input data sets need to be as large as possible. Every company has different test cases and methods but they share the same: Codes need to be Tested. A size as 128M I've mentioned is just an example and larger sets need to be created dynamically on the Heap. Usually performance and stress tests are done for different sizes of input data sets, like 64MB, 128MB, 256MB, 512MB, 1GB, 2GB, etc. In case of very large data sets, greater than 4GB, negative effects of Virtual Memory always clearly seen. Why wouldn't you create a different thread in the Watercooler Forum dedicated to STXXL?

[Bernard]

I actually plan to create a thread for the Quadrature tests which I conduct now.

STXXL is based on the External Memory concept so it can   be slower than std::vector or C-style array test when working on data set lying in physical memory. The more interested part will be when you are crossing into the realm of Virtual Memory.

[SergeyKostrov]

>>...The more interested part will be when you are crossing into the realm of Virtual Memory... Honestly, there is nothing interesting here because processing will get slower and I guarantee it for 100%.

[SergeyKostrov]

>>>>...The more interested part will be when you are crossing into the realm of Virtual Memory... >> >>Honestly, there is nothing interesting here because processing will get slower and I guarantee it for 100%. I could submit some performance numbers which demostrate how processing times are increasing as soon as VM is used.

[Bernard]

 >>>I could submit some performance numbers which demostrate how processing times are increasing as soon as VM is used.>>>

I would like to see it.

[Bernard]

Honestly, there is nothing interesting here because processing will get slower and I guarantee it for 100%.

  Yes that is expected.