31 May
2012
31 May
'12
1:14 p.m.
On 05/31/2012 07:16 AM, Gabriel M. Beddingfield wrote:
On 05/30/2012 09:21 AM, Sciss wrote:
Interesting! Thanks for sharing.
thanks for the link and the info. so you think
atom processors are
fine enough? the latency actually doesn't matter in my case. i'm more
worried that i'm going to through a lot of CPU heavy stuff on it, as
this will run experimental software I wrote myself (and I won't have
any time for performance tuning of the software itself).
I have an Atom N450 netbook. In practical terms, its limits for a
single task are:
* Time-stretching a clip in realtime to 50% to original
time span.
* A set with 4-5 monophonic synths, and 4-5 pure audio
tracks.
* All of the Renoise sample programs
* Most anything you can throw at Mixxx. While it has very good audio performance -- it has
significantly less
headroom than a Core2 or i-series processor (feels like a factor of 2 or
4).
Other non-audio tasks:
* Large compiles take 4-6x longer (e.g. kernel, Qt)
* Number crunching tasks are very slow. It's like the
floating point stuff is driving drunk.
LOL.
* Processor has a high performance hit for
inefficient
memory access (compared to Core2, i-series).
Clearly. i5 has 8192KB L1 cache.
I have not seen an ATOM CPU with more than 512KB.
* Processor doesn't benefit as much from SIMD
(SSE)
optimizations. E.g. you're lucky to get a 2x performance
boost using SIMD instructions... whereas a Core2 or i-series
will see at least a 2x performance boost.
* Most Atom devices have only 1GB RAM (2GB if you're lucky).
I've not seen an Atom device with more then 2GB.
Only a few Atom models support >2GB. The D525 for instance does.
Finally, all this experience is in 32-bit mode.
I've been running in
64-bit mode recently, but haven't done much audio with it. Overall, it
feels about the same.
For audio (read: jackd) 64 bit should make no difference. It's all 32
bit floats, anyway.
The Atom D525 that Egor and i recommended is not that bad. I don't have
access to an Atom N450, but the N270 is waaaay /slower/:
Here's a quick jconvolver benchmark:
A - Intel(R) Core(TM) Duo CPU @ 1.66GHz (32 bit - 2 cores, 2 CPUs)
B - Intel(R) Core(TM) i5 CPU @ 2.80GHz (64 bit - 4 cores, 4 CPUs)
C - Intel(R) Atom(TM) CPU D525 @ 1.80GHz (64 bit - 2 cores, 4 CPUs)
D - Intel(R) Atom(TM) CPU N270 @ 1.60GHz (32 bit - 1 core, 2 CPUs)
1) `jackd -d alsa -p1024` ;
4 instatances of `jconvolver weird.conf` in series
first and last instance connected to system:*
A: CPU-load: 16% (all CPUs) jack-DSP-load: 7.1%
B: CPU-load: 6% (all CPUs) jack-DSP-load: 6.2%
C: CPU-load: 13% (all CPUs) jack-DSP-load: 10.6%
D: CPU-load: 56% (all CPUs) jack-DSP-load: 24.9%
2) `jackd -d alsa -p64`
4 instatances of `jconvolver weird.conf` in series
first and last instance connected to system:*
A: CPU-load; 27% (all CPUs) jack-DSP-load: 13.0%
B: CPU-load: 8% (all CPUs) jack-DSP-load: 7.3%
C: CPU-load: 18% (all CPUs) jack-DSP-load: 20.3%
D: -- jconv exit with 2nd instance: processor can't keep up ---
ciao,
robin