[Eeglablist] Hardware recommendations

[Makoto Miyakoshi]

Dear Malte,

> I actually wanna see ERSP between 1 Hz and 80 Hz.

This is how you calculate the sliding window length.
If you want to decompose 1Hz with 3 cycles, then your sliding window length
is 3 second. You have only -1 to 2 second, so the sliding window cannot
slide.

See more info in this section (again).

https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline#A_tip_to_compute_time-frequency_transform_.28i.e._ERSP_.26_ITC.29_.2801.2F11.2F2017_updated.29

The shortening of ERSP/ITC data also happens because of this sliding window
nature of this analysis. Basically, the half length of the sliding window
will be lost--but 'lost' in terms of apparent data, and in reality the
'lost' part of the data was still used to generate the very first/last data
points.

Makoto

On Wed, Jan 29, 2020 at 4:12 AM Malte Anders <malteanders at gmail.com> wrote:

> Dear Makoto and Dear EEGLAB List,
>
> I am still having trouble finding the right settings for ERSP.
> My test data set is - for preprocessing - only highpass filtered at 1 Hz
> (and a small amount of ASR).
>
> I now call pop_newtimef() with the following settings:
> Epoch -1 to 2 seconds as recommended
> Cycles 3 0.5
>
> In the first test, I set Frequency limits to [3 40] (or 'freqs', [3 40].
> This gives me a good plot that shows ERSP roughly between 3 Hz and 40 Hz:
> https://imgur.com/ldviMsZ
> In the second test, the only thing I change is Frequency limits to [3 80]
> (or 'freqs', [3 80]). The resulting plot cuts off quite a few amount of
> data and only plots ERSP roughly between 6 Hz and 80 Hz:
> https://imgur.com/sQnAqvD
>
> I actually wanna see ERSP between 1 Hz and 80 Hz. However, the more I
> increase the number of the highest frequency (e.g. 40 Hz to 80 Hz), the
> more is cut off in the lower frequency range. I don't even get down to 1 Hz
> with 'freqs', [3 40]. Setting  'freqs', [1 40] results in an error: Not
> enough data points, reduce the window size or lowest frequency.
> The bad thing is that I do need especially the lower frequencies (1 - 12
> Hz would be perfect) and the higher gamma frequencies between 40 and 80 Hz
>
> I have dug into the available documentation, from my understanding setting
> freqs to [3 40] and wavelet cycles to [3 0.5] should result in 1 Hz being
> the lowest frequency analyzed, but in this case the plot is cut off at 3
> Hz. Those are the settings: https://imgur.com/q8jcAss, resulting in the
> picture above ( https://imgur.com/sQnAqvD ) where crucial info below 5-6
> Hz is simply cut off in the plot...
>
> Anybody can shed some light onto what I need to change?
>
> Thanks!
>
> Malte
>
> Am Mi., 15. Jan. 2020 um 20:14 Uhr schrieb Makoto Miyakoshi <
> mmiyakoshi at ucsd.edu>:
>
>> Dear Malte,
>>
>> Let me summarize our communication outside the list for others.
>>
>> %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>>
>> > CPU utilization is ~40% during the process, RAM utilization is 100% (so
>> 100% of the 32 Gb are used)
>>
>> I suggest you try the following solution.
>>
>>    - Use the following optional inputs: 'cycles', [3 10], 'freqs', [1 45],
>>    'nfreqs', [50], 'timesout', 6849 This should set your sliding window
>> size
>>    to 3380 ms or something like that, which should produce 6849 time point
>>    i.e., about 1 time point per second.
>>
>> The most alarming sign was 'Using 3 cycles at lowest freq to 25600 at
>> highest' 25600 cycles sounds crazy. If you use the above parameters, it
>> will be up to 10.
>>
>> Makoto
>>
>> On Mon, Jan 13, 2020 at 11:20 PM Malte Anders <malteanders at gmail.com>
>> wrote:
>>
>> > Hi everybody,
>> >
>> > I have collected huge amounts of EEG data with 512 Hz sampling frequency
>> > and 34 electrodes.
>> >
>> > For one subject (~2-3 hours of data), a simple Time/Frequency
>> decomposition
>> > between 1 and 40 Hz with standard settings calling the newtimef function
>> > takes approx. 8 hours for 1 hour of recorded data on my Computer
>> (4790k, 32
>> > Gb of DDR3 Ram, Intel HD GPU), so 16-24 hours for one dataset in total.
>> > This is simply too long as I am currently working alone on this. CPU
>> > utilization is ~40% during the process, RAM utilization is 100% (so
>> 100% of
>> > the 32 Gb are used).
>> >
>> > I would like to invest in new hardware (I am also considering used
>> > hardware) with a price point around 1500-3000€. Which would make more
>> > sense:
>> > -A new Ryzen 3950x, 16 cores with 128 Gb of DDR4 RAM (2666 Mhz) ~ 1500€
>> > -A used workstation, e.g. with Dual CPU: 2 x 12 core Xeon E52670v3 and
>> 256
>> > Gb RAM ~ used approx. 2200€
>> >
>> > The Xeon has more cores, but way lower clock speed. It does have double
>> the
>> > amount of RAM though.
>> >
>> > Would you recommend a dedicated GPU (Nvidia?) for this?
>> >
>> > Any recommendations are greatly appreciated. Thanks!
>> > Malte
>> > _______________________________________________
>> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> > To unsubscribe, send an empty email to
>> > eeglablist-unsubscribe at sccn.ucsd.edu
>> > For digest mode, send an email with the subject "set digest mime" to
>> > eeglablist-request at sccn.ucsd.edu
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu
>> For digest mode, send an email with the subject "set digest mime" to
>> eeglablist-request at sccn.ucsd.edu
>
>
>
> --
> Mit freundlichen Grüßen,
>
> Malte Anders
>