For me on all of my computers (various Windows versions, and various Intel CPU's), VMethod = NFCV task has always run much faster than VMethod = LOO tasks.

I've seen before on the board that people say what you say, but for me, it's always been the other way around.

Thanks Joe!

As always then, YMMV (all depending on Computer/Memory/OS/CPU). In my case NFCV always beats LOO

Not a clue. I'm not a programmer. But that does of course not stop me from being able to observe how the two different types behave on my computers,
Edit: Also that my CPU's runs hotter from crunching NFCV tasks,compared to LOO tasks.

From some CPU performance testing I did on various machines during the second half of 2020, I offer the following about MCM1 running on Linux...

Firstly, MCM1 programs execute a higher proportion of branch instructions than any other WCG application I looked at! For a typical NFCV task it's over 19.5%; for a typical LOO task it's just over 17.8%. For comparison, MIP1 was about 15%, OPN1 is about 11.2%..11.5%, FAH2 was about 8.7%, ARP1 is just under 7.5% and HST1 is under 3% :-) -- unfortunately, I don't have these numbers for SCC1.

Next, if the only BOINC task running on one of my Intel machines is an MCM1 task, a LOO task executes around twice as many instructions per second as an NFCV task would. That tends to confirm that there's more compute-intensive instructions in the main NFCV loop(s) - data processing versus computing :-). (All my Intel machines are 7th Generation; newer Intel CPUs might give different numbers if my Ryzen is an indicator...)

On my Ryzen 3700X a solo MCM1 LOO task only manages about 20% more instructions a second, so the better pipelines on the Ryzen help out NFCV code! This might also suggest [in part?] the apparent inversion Grumpy Swede was talking about -- the compiler used for Windows might have made a better job of code generation for some parts of LOO! And again, there's no real fall-off in throughput as the workload increases... All MCM1 jobs run faster on the Ryzen anyway, but the difference in NFCV tasks is very noticeable!

One really nice thing about MCM1 (and SCC1!) is that there's not much fall-off in performance as the work load increases (and the performance ratios are still very similar); memory seems to be used in such a way as to need minimal L3 cache access!

And one final observation - I ran a 32-bit Linux VM for a while and let it run one BOINC task at a time. On that machine, MCM1 tasks often had 10% to 20% lower run time than tasks running on the host! I seem to recall other cases where folks have said that 32-bit versions of WCG apps ran better than their 64-bit equivalents, but I may be misremembering... :-)

Cheers - Al.

P.S. performance figures acquired using perf stat.

Thanks for your explanation alanb1951. I am running older Xeons and there is a stark difference in the run times for each. It also appears the Windows compiler is more efficient in this case than the Linux compiler for this application.
Cheers