Index  | Recent Threads  | Unanswered Threads  | Who's Active  | Guidelines  | Search
 

Quick Go ยป
No member browsing this thread
Thread Status: Active
Total posts in this thread: 12
Posts: 12   Pages: 2   [ Previous Page | 1 2 ]
[ Jump to Last Post ]
Post new Thread
Author
Previous Thread This topic has been viewed 3801 times and has 11 replies Next Thread
alanb1951
Veteran Cruncher
Joined: Jan 20, 2006
Post Count: 929
Status: Recently Active
Project Badges:
Reply to this Post  Reply with Quote 
Re: What happened to MCM on Arm?

Thanks for the numbers! -- it looks like a native ARM app would be a definite advantage for most tasks - see below...

There seem to be two different types of MCM1 jobs, one showing VMethod=NFCV in its parameters, the other showing VMethod=LOO. On my Intel and Ryzen machines, NFCV tasks seem to take from 25% longer (Ryzen, with extra FP pathways!) up to over twice as long to run (Intel Kaby Lake), and something equivalent may be playing out with your numbers on the Pi4 (though it's a much bigger difference than the ones I've seen on native applications.)

When they ran the last MCM1 beta, I snagged a few on my Pi4, and the LOO ones typically took about 4.25 to 4.5 hours (3 at a time, 2GHz clock.) Given the lack of L3 cache and the relatively slow clock compared to my "traditional" machines, I reckon comparing that against around 1.75 hours on an i5-7600 or i7-7700K or around 1.6 hours on a Ryzen 3700X isn't too bad!!! Unfortunately, I have no recorded times for NFCV tasks on the Pi4 :-(

By the way, the beta more or less equalized the run-times of NFCV and LOO tasks on Intel and Ryzen - if anything, the NFCV tasks were now a bit quicker than the LOO tasks. Whether this would also apply on the Pi remains to be seen, as [according to perf stat anallysis done on the production application on various platforms] NFCV appears to make more calls on L3 cache per instruction executed than LOO does, and the lack of such cache may affect timings. ,

Given the above, I'd repeat my plea to the WCC technicians - please can we have an ARM version of MCM1? If that Beta test was anything to go by, it would make far better use of my Pi4 than does OPN1 (with its much higher amount of L2 cache misses!)

Cheers - Al.
[Aug 28, 2021 1:14:30 AM]   Link   Report threatening or abusive post: please login first  Go to top 
mgpointner
Advanced Cruncher
Argentina
Joined: Nov 16, 2009
Post Count: 55
Status: Offline
Project Badges:
Reply to this Post  Reply with Quote 
Re: What happened to MCM on Arm?

Raspberry Pi 4 with Android 11
----------------------------------
Name CPU Time Credit
MCM1_0180026_9501_0 04.67 / 04.83 128.2 / 0.0 VMethod = LOO
MCM1_0179994_4882_0 04.68 / 04.84 025.0 / 0.0 VMethod = LOO
MCM1_0179927_4524_0 18.68 / 19.26 091.6 / 0.0 VMethod = NFCV
MCM1_0179809_5346_0 04.65 / 04.79 014.3 / 0.0 VMethod = LOO
MCM1_0179806_0132_0 17.97 / 18.48 065.4 / 0.0 VMethod = NFCV
MCM1_0179806_6979_0 17.97 / 18.48 065.4 / 0.0 VMethod = NFCV
----------------------------------------

[Aug 28, 2021 2:51:00 AM]   Link   Report threatening or abusive post: please login first  Go to top 
Posts: 12   Pages: 2   [ Previous Page | 1 2 ]
[ Jump to Last Post ]
Post new Thread