After receiving a work unit, the client performs a baseline smoothing on the data to remove any wideband (Du > 2 kHz) features. This prevents the client from confusing fluctuations in broadband noise (due in part to variations in the hydrogen line emission as the field of view transits the sky) with intelligent signals. The client then begins the main data analysis loop, shown schematically below:
for Doppler drift rates from -50 Hz/s to +50 Hz {
for bandwidths from 0.075 to 1220 Hz in 2X steps {
Generate time ordered power spectra.
Search for short duration signals above a
constant threshold (spikes) for each frequency {
Search for faint signals matching beam
parameters (Gaussians)
Search for groups of three evenly spaced
signals (triplets)
Search for faint repeating pulses (pulses)
}
}
}
At the start of each passage through the loop, the data is transformed into an accelerated frame of a given Doppler drift rate. The drift rates at which the client searches the data for signals vary from ?10 Hz/sec to +10 Hz/sec (accelerations expected on a rapidly rotating planet) in steps of 0.0018 Hz/sec. The client also examines the data at Doppler drift rates out to ±50 Hz/sec (accelerations of the magnitude that would arise from a satellite in low orbit about an earth-like planet), but at a more coarse step of 0.029 Hz/sec. A signal from an alien world would most likely have a negative drift rate (as the accelerations involved would be away from the observer). Despite this, we examine both positive and negative drift rates for the purpose of statistical comparison and to leave open the possibility of detecting a deliberately chirped extraterrestrial signal.
At each drift rate, the client searches for signals at one or more bandwidths between 0.075 and 1,221 Hz. This is accomplished by using FFTs of length 2n (n = 3, 4, ..., 17) to transform the data into a number of time-ordered power spectra. To avoid repeating work, not all bandwidths are examined at every Doppler drift rate. Only when the change in drift rate becomes significant compared to (1/Du2) does the program compute another FFT of a given length. Therefore, 32k-point transforms are performed one quarter as often at 64k-point transforms.
The transformed data is examined for signals that exceed 22 times the mean noise power. This threshold corresponds to 7.2 × 1025 W/m2 at our finest frequency resolutions, or the equivalent of detecting a cell phone on one of the moons of Saturn. The SETI@home client reports any such spike signals in the resulting transmission.
If there is sufficient time resolution in the transformed data (n < 15) and the SETI receiver is not tracking an object on the sky, the client examines it for signals that match the telescope beam's parameters. As a radio source drifts through the field of view, the measured power will vary depending on the telescope's beam profile, which is approximately Gaussian. The SETI@home client performs a weighted c2 fit on any signals that exceed 3.2 times the mean noise power and reports those for which the goodness of fit exceeds a certain level. This power level typically corresponds to 8.4 × 10?25 W/m2.
The client then divides transformed data at each frequency into chunks with duration equal to the time required for an object to transit the telescope field of view. Two algorithms serve to analyze these chunks for pulsed signals. The first algorithm, the triplet finder, searches each chunk for three evenly spaced signals that each exceed 7.75 times the mean noise power (as little as 5.3 × 10?25 W/m2) and report any detected signals.
The second algorithm is a modified fast-folding algorithm. A folding algorithm divides the data into chunks of duration equal to the period being searched and co-adds them to improve signal-to-noise ratio. An FFA performs this function on a large number of periods without duplicating additions. The SETI@home folding algorithm searches roughly N log N pulse periods, where N is the length of the input array. This corresponds to periods between two samples and N/3 samples. During a typical run of the client, this typically means half a million periods between 2 ms and 10 s. The program computes the threshold for detecting a pulsed signal dynamically to match the number of co-added samples. This threshold can be as low as 0.04 times the mean noise power for pulses with periods less than 10 ms. This corresponds to pulse energies of about 1.8 × 10?26 J/m2.
Depending on the individual work unit's parameters, this processing loop requires 2.4 to 3.8 trillion floating-point operations (Tflop). It takes a typical (500 MHz) home computer 10 to 12 hours to complete a work unit. For an average work unit, the SETI@home client would report eight signals?four spike signals, one Gaussian, one pulsed signal, and one triplet signal.
more info here
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)
Discussion Intel current and future Lakes & Rapids thread
