Hello, With current setup of All2All/Flat, we seem to have a stable 410 GB/s from MCDRAM and 78 GB/s from DDR. For STREAMing from MCDRAM, Intel people suggested to try SNC-4/Flat promising 430 GB/s. Our previous trials with other combinations of NOC settings lead to unstable results. Do we want to try it now? It will probably require hard reboot. Vitali On 04/13/2016 01:06 PM, Raman, Karthik wrote:

Ok I read this as 418 GB/s J

And ok, just realized you are running in All2All mode, so yes this looks good. For Quad/Flat you should see ~440 GB/s with the same run.

*From:*Vitali A. Morozov [mailto:[email protected]] *Sent:* Wednesday, April 13, 2016 11:02 AM *To:* Raman, Karthik <[email protected]>; Chinthamani, Sundaram <[email protected]>; Kumaran, Kalyan <[email protected]> *Cc:* Sodani, Avinash <[email protected]>; Fromkin, Russ <[email protected]>; Harms, Kevin <[email protected]>; Morozov, Vitali A. <[email protected]>; Allen, Benjamin S. <[email protected]> *Subject:* Re: [JLSE-Intel-NDA] KNL Status

Hi Karthik,

Repeating the same options you have:

mpiicc -O3 -xMIC-AVX512 -opt-report=5 -Wall -g -fopenmp -mcmodel medium -shared-intel -opt-streaming-stores always stream.c -o stream

------------------------------------------------------------- STREAM version $Revision: 5.9 $ ------------------------------------------------------------- This system uses 8 bytes per DOUBLE PRECISION word. ------------------------------------------------------------- Array size = 134217728, Offset = 0 Total memory required = 3072.0 MB. Each test is run 100 times, but only the *best* time for each is used. ------------------------------------------------------------- Number of Threads requested = 64 -------------------------------------------------------------

< skiped >

Function Rate (MB/s) Avg time Min time Max time Copy: 362331.5200 0.0060 0.0059 0.0061 Scale: 359281.9806 0.0060 0.0060 0.0062 Add: 418122.7024 0.0078 0.0077 0.0081 Triad: 418122.7024 0.0078 0.0077 0.0081 Store: 224685.6729 0.0048 0.0048 0.0049

408GB/s

Vitali