[LAU] Bye Bye 32 bit

[Len Ovens]

On Wed, 27 Dec 2017, Manuel Haible wrote:

> I assume: for      high samplerates  + many tracks  + many fx  + very small
> buffer size  + ultra low latency singlecore performance is equally important as
> the number of cores?

Any one track with all its effects and buses it feeds must be in the same 
thread. That is, any audio operation that requires another's result to 
begin it's own processing must be run by the same thread. So yes, single 
core performance is important. For example, a guiter effects chain would 
be run as a single thread for audio (the GUI would be separate and 
probably of low priority) on one core.

That may be part of the problem with multi-core effects that sound better 
with multi-core use turned off. An effect is generally a serial chain of 
audio processes that should not be split up.

Secondary cores that borrow 
resources from another core like hyperthreading, may cause trouble for 
very low latency use. In the days of the single core P4, I found that with 
HT enabled I had the odd xrun with jack set 64/2, but with HT off I could 
get latency of 16/2 with no xruns using an ice1712 interface. I have not 
tested things on newer i3/7 cpus as I have side stepped the whole issue by 
using a 4 core i5.

CPU speed still affects things though. Anything that changes the cpu speed 
seems to affect audio. I have had better results forcing a cpu to only 
800Mhz than allowing that same cpu to run at 1.6Ghz with speed changes. My 
i5 runs up to 3.2Ghz and just putting it into performance mode helps but 
"Boost" can cause xruns too I have found. I think I could set performance 
to 3.4ghz (effectively over clocking) and get things to work as the 
temperature stays within limits even with all 4 cores at 100% (building 
Ardour takes 15 minutes or so and makes quite a good cpu "burn" test) and 
audio can not run at 100% anyway. But I have not tried that.

--
Len Ovens
www.ovenwerks.net