JYR wrote:

>> Oh, ok, thanks for making this point clear. It will perhaps prevent
>> everybody posting here about "this bug is still not fixed" and the like.
>> Well, 1st impression here: 3.7.0.beta6 seems halfway between 3.6.0a and
>> 3.7.0.beta5a as far as speed is concerned. I'll try to run a few tests on
>> my scenes now.


BTW some hints for running speed comparison tests: the most 'pure' comparison
is to run the new code with '/threads 1', since that most closely
approximates the way the old code worked. that is what I do if I want to work
out where the new code is slower than the old.

from a practical standpoint, however, if you have hyperthreading available
you ought to run at least four and perhaps more threads. if you have a
dual-core CPU then you ought to run four or perhaps eight if you have a
dual-core HT box (though the CPU time reports will be misleading in the
latter case).

if you have an SMP box with discrete CPU's you will suffer a performance
penalty on some scenes as against a dual-core due to the need for write-
combines on the CPU caches. this will show itself as increased CPU time
needed for a render when you switch from one thread to more than one.

by way of example, I have a dual Xeon and a dual-core AMD machine here. the
render times on (for example) mediasky.pov are very similar on both when run
with one thread, but when I start using both CPU's the AMD thrashes the Xeon
because it has both CPU's on-chip and they share a cache. (this would
actually be a disadvantage if you were running multiple applications, but in
the case of a single app it can be very useful).

I haven't tested on an Intel dual-core yet so can't comment on that one.

users of discrete SMP machines will however probably notice a more
significant improvement in this current beta than users of other systems
since some of the changes I made affect that class of system more strongly.

the other thing to take into account is that if you render with AA the new
code will be at a disadvantage since the current default AA code actually
renders more pixels; to do a side-by-side comparison of speed it's best to
leave AA off.

-- Chris

Post a reply to this message