#1

[kruoli]

Originally Posted by Xyzzy

Note that our BIOS offers the following memory interleaving options:

I have it set to "none", but I'm not sure how this will affect msieve's performance.

[axn]

Originally Posted by Xyzzy

We think that our CCX layout is:

I find lscpu -p to be a compact representation of the situation.

[Xyzzy]

Here are all 5 possibilities:

AUTO:

Logical Processor to NUMA Node Map:
************************  NUMA Node 0
-  NUMA Node 1
Calculating Cross-NUMA Node Access Cost...
                                          
Approximate Cross-NUMA Node Access Cost (relative to fastest):
     00  01
00: 1.2 1.0
01: 0.0 0.0

CHANNEL:

Logical Processor to NUMA Node Map:
************------------  NUMA Node 0
------------************  NUMA Node 1
Calculating Cross-NUMA Node Access Cost...
                                          
Approximate Cross-NUMA Node Access Cost (relative to fastest):
     00  01
00: 1.0 1.2
01: 1.4 1.3

NONE:

Logical Processor to NUMA Node Map:
************------------  NUMA Node 0
------------************  NUMA Node 1
Calculating Cross-NUMA Node Access Cost...
                                          
Approximate Cross-NUMA Node Access Cost (relative to fastest):
     00  01
00: 1.0 1.6
01: 1.4 1.4

DIE:

Logical Processor to NUMA Node Map:
************************  NUMA Node 0
-  NUMA Node 1
Calculating Cross-NUMA Node Access Cost...
                                          
Approximate Cross-NUMA Node Access Cost (relative to fastest):
     00  01
00: 1.3 1.0
01: 0.0 0.0

SOCKET:

Logical Processor to NUMA Node Map:
************************  NUMA Node 0
-  NUMA Node 1
Calculating Cross-NUMA Node Access Cost...
                                          
Approximate Cross-NUMA Node Access Cost (relative to fastest):
     00  01
00: 1.3 1.0
01: 0.0 0.0

We have no idea which is best.

[Xyzzy]

With the memory interleave setting on "channel" and with 12 cores running msieve, the CPU graph looks very different.

Before:

OXXOXO
XOXOXO
XOXOXO
XOXOXO

After:

OOOOOO
OOOOOO
XXXXXX
XXXXXX

[VBCurtis]

But is it faster?

[Xyzzy]

We changed the video card to a Radeon. We have never seen GPU-Z show memory and memory VRM temperatures before. That is pretty handy!

[Xyzzy]

Here is the Windows task manager in "NUMA" mode:

[PhilF]

Originally Posted by Xyzzy

We changed the video card to a Radeon. We have never seen GPU-Z show memory and memory VRM temperatures before. That is pretty handy!

You went from Nvidia to Radeon??? Man, you're brave!

[Xyzzy]

[ewmayer]

Originally Posted by PhilF

You went from Nvidia to Radeon??? Man, you're brave!

I did the same thing when I upgraded my old Haswell system to use a Radeon VII earlier this year ... had a really old nVidia gtx430 card in there which Xyzzy had gifted me years ago so I could play with CUDA coding on a card that supported doubles (compute level 2.0, or whatever it's called). That needed me to rip out all the nVidia-driver stuff in order to get the R7 working, and was still borking me i another way until recently, in that I was unable to do the manual mem-clock fiddling R7 users do to optimize their FLOPS/watt ... in the end it turned out the ghost of the old gtx430 was still present by way of occupying the 'card0' entries in /sys/class/drm/, meaning my attempts to fiddle the R7 mclk via the usual sequence

echo "manual" >/sys/class/drm/card0/device/power_dpm_force_performance_level
#Undervolt by setting max voltage
#               V Set this to 50mV less than the max stock voltage of your card (which varies from card to card), then optionally tune it down
echo "vc 2 1801 1010" >/sys/class/drm/card0/device/pp_od_clk_voltage
#Overclock mclk to 1150
echo "m 1 1150" >/sys/class/drm/card0/device/pp_od_clk_voltage
#Push a dummy sclk change for the undervolt to stick
echo "s 1 1801" >/sys/class/drm/card0/device/pp_od_clk_voltage
#Push everything to the card
echo "c" >/sys/class/drm/card0/device/pp_od_clk_voltage

...were failing because the R7 was in fact at card1.

[moebius]

Originally Posted by PhilF

You went from Nvidia to Radeon??? Man, you're brave!

F*** off current Navi, even old Vega10 chip , is 4 times faster than an 8-core Ryzen 3700X, and also at least 50 Euros cheaper than a AMD RX 5700XT....