sse2 vs avx
log in

Advanced search

Message boards : Number crunching : sse2 vs avx

Previous · 1 · 2
Author Message
Profile Beyond
Avatar
Send message
Joined: 2 Nov 16
Posts: 50
Credit: 44,372,499
RAC: 0
United States
Message 1429 - Posted: 11 Dec 2018, 16:50:08 UTC - in response to Message 1428.

Is fma faster than avx on the Ryzen 1700?

It seems to be just slightly, though they are so close that it would take longer-term testing to be sure. I think it would be easier for the project to find the best extension for a given processor type, and just use it.

I'll drink to that. None of my machines ever get the fma version so I don't even have the latest executable to test. Looking at the current applications it doesn't even show an fma version for Windows. Maybe v1.10 was the last fma version? I wish we could just pick the app(s) we want to run through project preferences.

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 0
United States
Message 1430 - Posted: 11 Dec 2018, 18:40:59 UTC - in response to Message 1429.

Looking at the current applications it doesn't even show an fma version for Windows. Maybe v1.10 was the last fma version? I wish we could just pick the app(s) we want to run through project preferences.

My Ryzen 1700 and 2700 machines are on Ubuntu 18.04, and yes the fma is version 1.10. But I used Windows earlier on an i7-3770, and it did quite well compared to Linux. But that one got only SSE2 and AVX. I think the latter was slightly faster.

I agree; let us select them and be done with it.

Profile JStateson
Send message
Joined: 7 May 18
Posts: 2
Credit: 3,577,000
RAC: 0
United States
Message 1431 - Posted: 12 Dec 2018, 14:04:30 UTC
Last modified: 12 Dec 2018, 14:12:40 UTC

had to abort sse2 after 10 hours but with 442 days remaining. there were no other cpu tasks running other then this project. Looking HERE sse2 and avx had same problem but fma and an anon succeeded. wonder what the "anon" was.

Profile valterc
Project administrator
Project tester
Send message
Joined: 30 Oct 13
Posts: 623
Credit: 34,677,535
RAC: 12
Italy
Message 1433 - Posted: 13 Dec 2018, 15:12:58 UTC - in response to Message 1429.

Is fma faster than avx on the Ryzen 1700?

It seems to be just slightly, though they are so close that it would take longer-term testing to be sure. I think it would be easier for the project to find the best extension for a given processor type, and just use it.

I'll drink to that. None of my machines ever get the fma version so I don't even have the latest executable to test. Looking at the current applications it doesn't even show an fma version for Windows. Maybe v1.10 was the last fma version? I wish we could just pick the app(s) we want to run through project preferences.

The latest fma version for Windows (v1.10) was available only as a 'test' app, mainly because problems (crashes) with the earliest bios for Ryzen cpus. The difference in speed between sse, avx and fma are not so huge so we decided not to build the v1.11 fma version for Windows.

Profile valterc
Project administrator
Project tester
Send message
Joined: 30 Oct 13
Posts: 623
Credit: 34,677,535
RAC: 12
Italy
Message 1434 - Posted: 13 Dec 2018, 15:21:11 UTC - in response to Message 1431.
Last modified: 13 Dec 2018, 15:39:22 UTC

had to abort sse2 after 10 hours but with 442 days remaining. there were no other cpu tasks running other then this project. Looking HERE sse2 and avx had same problem but fma and an anon succeeded. wonder what the "anon" was.

"aborted by user" is not 'technically' an error, it's a user's choice. I agree that if I saw a workunit stuck at 5% with an estimated time for completion of days (even if the estimate were completely wrong) I would also be tempted to abort it. The 'problematic' behavior of the TCGA workunits doesn't depend on the version of the application.

The sse2 problem of some Ryzen cpu with the current application is a 'real' problem: the app crashes with an 'illegal instruction' error.

I just sent an e-mail to Daniel (the user who actually wrote the sse2 code) asking for hints.

Profile [B@P] Daniel
Volunteer developer
Send message
Joined: 19 Oct 16
Posts: 90
Credit: 2,205,103
RAC: 0
Poland
Message 1445 - Posted: 17 Dec 2018, 22:28:21 UTC - in response to Message 1434.

had to abort sse2 after 10 hours but with 442 days remaining. there were no other cpu tasks running other then this project. Looking HERE sse2 and avx had same problem but fma and an anon succeeded. wonder what the "anon" was.

"aborted by user" is not 'technically' an error, it's a user's choice. I agree that if I saw a workunit stuck at 5% with an estimated time for completion of days (even if the estimate were completely wrong) I would also be tempted to abort it. The 'problematic' behavior of the TCGA workunits doesn't depend on the version of the application.

The sse2 problem of some Ryzen cpu with the current application is a 'real' problem: the app crashes with an 'illegal instruction' error.

I just sent an e-mail to Daniel (the user who actually wrote the sse2 code) asking for hints.

Hello again, I was not here for long long time :)

This looks like another Ryzen bug. This time CPU also jumps to address in middle of instruction, what must end in crash sooner or later.

I have reported this bug on AMD forum. Here is link to my post, it should be visible soon when moderator approves it:
https://community.amd.com/message/2890585
____________

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 0
United States
Message 1447 - Posted: 18 Dec 2018, 18:33:39 UTC - in response to Message 1445.


This looks like another Ryzen bug. This time CPU also jumps to address in middle of instruction, what must end in crash sooner or later.

I have reported this bug on AMD forum. Here is link to my post, it should be visible soon when moderator approves it:
https://community.amd.com/message/2890585

The SSE2 problem for me (and Beyond) was only on the Ryzen 1700. The Ryzen 2700 is OK. Maybe you should amend your report?

Thanks for looking into this.

Profile [B@P] Daniel
Volunteer developer
Send message
Joined: 19 Oct 16
Posts: 90
Credit: 2,205,103
RAC: 0
Poland
Message 1448 - Posted: 18 Dec 2018, 21:16:33 UTC - in response to Message 1447.


This looks like another Ryzen bug. This time CPU also jumps to address in middle of instruction, what must end in crash sooner or later.

I have reported this bug on AMD forum. Here is link to my post, it should be visible soon when moderator approves it:
https://community.amd.com/message/2890585

The SSE2 problem for me (and Beyond) was only on the Ryzen 1700. The Ryzen 2700 is OK. Maybe you should amend your report?

Thanks for looking into this.

Interesting. I have followed link on task info page to get info about CPU and OS, so looks that sometimes crashes occur on 2700 too. Unfortunately today this task page is deleted so I cannot add link here. Anyway, I will update my report that bug happens mostly on 1700.
____________

Profile Beyond
Avatar
Send message
Joined: 2 Nov 16
Posts: 50
Credit: 44,372,499
RAC: 0
United States
Message 1449 - Posted: 19 Dec 2018, 3:12:59 UTC - in response to Message 1448.

Interesting. I have followed link on task info page to get info about CPU and OS, so looks that sometimes crashes occur on 2700 too. Unfortunately today this task page is deleted so I cannot add link here. Anyway, I will update my report that bug happens mostly on 1700.

Never had SSE failures on my 2700, only on the 2 1700 Ryzens. AVX and FMA are bulletproof on all of them.

Thanks/Ed

noxcivi
Send message
Joined: 19 Aug 17
Posts: 3
Credit: 2,564,781
RAC: 0
Germany
Message 1626 - Posted: 30 Oct 2019, 15:04:26 UTC
Last modified: 30 Oct 2019, 15:05:49 UTC

I have to dig up this topic. This project is very limited in allowing the users to decice which app will run. It is all decided by whatever algorithm.

Now, I have a new CPU that would be faster with the avx version but my machine will not get that version, because the old CPU was slower on avx. The avx-FLOPS for my machine are outdated, but cannot become refreshed as I will not get avx tasks. This is a loop I cannot escape on my own.

Using an app_info.xml is a shitty workaround because it creates a new "generic local app"-profile on the servers statistics. I'd like to use the avx version the official way.

Jim1348
Send message
Joined: 29 Dec 16
Posts: 87
Credit: 21,013,002
RAC: 0
United States
Message 1627 - Posted: 30 Oct 2019, 15:29:49 UTC - in response to Message 1626.
Last modified: 30 Oct 2019, 15:31:32 UTC

Now, I have a new CPU that would be faster with the avx version but my machine will not get that version, because the old CPU was slower on avx.

If you are referring to your Ryzen 3600, it will probably be faster with the fma version. That is what both my Ryzen 2600s end up with, though they are on Ubuntu.

PS - The difference is very small anyway, almost not worth the selection process.

noxcivi
Send message
Joined: 19 Aug 17
Posts: 3
Credit: 2,564,781
RAC: 0
Germany
Message 1631 - Posted: 1 Nov 2019, 10:20:13 UTC - in response to Message 1627.

If you show me an example that a Ryzen 3000's advantage is less than 3% for using avx/fma over sse2, I will be quiet on this topic forever. My guess was that it would be much more and that is why I wanted to mention the missing options for the users.

Previous · 1 · 2
Post to thread

Message boards : Number crunching : sse2 vs avx


Main page · Your account · Message boards


Copyright © 2024 CNR-TN & UniTN