SSE2 with app_info?
log in

Advanced search

Message boards : Number crunching : SSE2 with app_info?

Author Message
Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2734 - Posted: 22 Jun 2022, 18:32:13 UTC

I would like to run the SSE2 version rather than the fma version on a Linux (Ubuntu 20.04.4) machine for reduced heat production.

Is there an app_info that can accomplish that? I only see one for avx.

Profile valterc
Project administrator
Project tester
Send message
Joined: 30 Oct 13
Posts: 560
Credit: 31,454,864
RAC: 13,726
Italy
Message 2736 - Posted: 24 Jun 2022, 13:00:14 UTC - in response to Message 2734.

Just substitute avx with sse2 everywhere inside the app_info you got. Just remember to download the right executable for your platform, also do a "chmod a+x"

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2737 - Posted: 24 Jun 2022, 14:29:29 UTC - in response to Message 2736.

... also do a "chmod a+x"

Thanks. That is what I was missing. I had done the other stuff.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2738 - Posted: 24 Jun 2022, 16:16:12 UTC - in response to Message 2734.
Last modified: 24 Jun 2022, 16:16:47 UTC

I would like to run the SSE2 version rather than the fma version on a Linux (Ubuntu 20.04.4) machine for reduced heat production.

We just started another heat wave here. Do you have a feel for the percent Wattage reduction?

Warning: For an app_info.xml to take effect you must restart BOINC. When you do it will delete every TN-Grid WU you have and DL only SSE2 WUs and start fresh.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2739 - Posted: 24 Jun 2022, 16:18:00 UTC - in response to Message 2737.

... also do a "chmod a+x"

Thanks. That is what I was missing. I had done the other stuff.

I'm lazy, I make a copy of my app_config and rename it, comes with permissions preloaded :-)

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2741 - Posted: 24 Jun 2022, 17:21:04 UTC - in response to Message 2738.

We just started another heat wave here. Do you have a feel for the percent Wattage reduction?

No, I just look at the temp. I was running on all cores of a Ryzen 3950X, and it was getting up to 84C, which is near the danger zone when I pick up errors. I like to keep it below 80C.

But for the moment, I have achieved that by under-volting and setting the clock speed to what it can handle. I am now good at 77C even with the fma when running at 3900 MHz and 1.25000 volts on the CPU (each chip is different as you know). However, I know from past experience that SSE2 will run several degrees cooler with only a small loss in performance, if I need it. It probably is only a 5 watt difference in power on that chip though.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2744 - Posted: 24 Jun 2022, 23:08:06 UTC

5 Watts on a 16c/32t CPU is well worth the effort. Think I'll get out the watt meter and run a test. Thanks for the tip.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2745 - Posted: 27 Jun 2022, 14:42:28 UTC

I'm not sure I can tell the difference. I have an older Ensupra Energy Monitor. The newer ones have a graph. I thought it had an averaging feature but I can't find the instructions. The range I saw has a big overlap and may be confounded by GPU or RAM fluctuations.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2763 - Posted: 8 Jul 2022, 19:04:56 UTC

Still curious about the topic of which instruction set is the most energy efficient or at least runs CPU at coolest temperature under full work load.
Found this interesting looking paper:
Thermal design power and vectorized instructions behavior, Amina Guermouche & Anne-Cécile Orgerie, CONCURRENCY & COMPUTATION: PRACTICE & EXPERIENCE, Feb 2021.
https://hal.archives-ouvertes.fr/hal-03185821/document

I haven't read the whole thing yet but it seems to imply SSE has the lowest power ratio for both CPU operation and DRAM. They also test AVX512.

"AVX2 extension adds fused multiply add instructions (FMA)."
Does this mean what we label as FMA is also AVX2?
They use the term memory-bound which I confess I don't understand. E.g., https://www.intel.com/content/www/us/en/develop/documentation/vtune-help/top/reference/cpu-metrics-reference/memory-bound.html
Are TN-Grid WUs memory-bound?
Do TN-Grid WUs ever trigger Intel CPU Turbo Boost?

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2779 - Posted: 22 Jul 2022, 12:13:16 UTC
Last modified: 22 Jul 2022, 12:13:35 UTC

With the Northern Hemisphere on fire I'm surprised no one has taken any interest in this paper I posted about reducing CPU power consumption.

I switched all computers to SSE2.

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2780 - Posted: 22 Jul 2022, 14:38:26 UTC - in response to Message 2779.

A good point, but FMA is working for me well enough.
It is mainly west of the Mississippi River that they have the problems.

We may need to give up the Louisiana Purchase.

Bryn Mawr
Send message
Joined: 23 Jun 20
Posts: 40
Credit: 11,269,665
RAC: 9,867
United Kingdom
Message 2782 - Posted: 22 Jul 2022, 17:44:50 UTC - in response to Message 2780.

A good point, but FMA is working for me well enough.
It is mainly west of the Mississippi River that they have the problems.

We may need to give up the Louisiana Purchase.


And France and Spain and Greece and England and …

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2783 - Posted: 22 Jul 2022, 20:01:56 UTC - in response to Message 2782.

And France and Spain and Greece and England and …

I know. It is a big problem.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2792 - Posted: 30 Jul 2022, 19:17:38 UTC - in response to Message 2780.
Last modified: 30 Jul 2022, 19:24:21 UTC

We may need to give up the Louisiana Purchase.

Well there goes the farm belt :-)

But does any one agree with the authors that SSE uses less CPU Wattage than other instruction sets?

Speedy
Send message
Joined: 13 Nov 21
Posts: 17
Credit: 421,300
RAC: 3,814
New Zealand
Message 2813 - Posted: 7 Aug 2022, 4:40:30 UTC

Can you run SSE2 on a Windows AMD PC and does anybody know if it brings the runtime down from around 3 hours? On a AMD 5900 X CPU

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2814 - Posted: 7 Aug 2022, 11:48:55 UTC - in response to Message 2813.

Can you run SSE2 on a Windows AMD PC and does anybody know if it brings the runtime down from around 3 hours? On a AMD 5900 X CPU

The best I can tell you is that running the fma on a Ryzen 5950X under Ubuntu 20.04.4 takes 2 hours 27 minutes. And SSE2 is a little slower than that, while as I recall Windows is a little slower than Linux, but not by too much.

I think 3 hours is about what you should expect.

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2815 - Posted: 7 Aug 2022, 14:21:15 UTC

My point is energy efficiency, not run time.

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 2,285
United States
Message 2816 - Posted: 7 Aug 2022, 15:52:20 UTC - in response to Message 2815.
Last modified: 7 Aug 2022, 16:47:15 UTC

My point is energy efficiency, not run time.

It won't change it that much, a few watts. Anything in your system will use more. Do you have high-efficiency power supplies?

(It will take more than that to save the Louisiana Purchase anyway.)

Aurum
Send message
Joined: 18 Jul 18
Posts: 91
Credit: 288,121,456
RAC: 5,794
United States
Message 2862 - Posted: 2 Sep 2022, 13:21:56 UTC

It's astounding that no one has even read this paper. I think TN-GRID should stop issuing AVX and FMA WUs.
Depending on the program SSE could use as little as 60% of the power used by other intsruction sets. For an i9-7980XE with a 165 W TDP that's a savings of up to 66 W per computer.


Post to thread

Message boards : Number crunching : SSE2 with app_info?


Main page · Your account · Message boards


Copyright © 2022 CNR-TN & UniTN