Clarify "heavy computational demand" review note

paper/paper.md

@@ -65,14 +65,11 @@ While these tools represent significant progress, there remained a need for soft
 - The `shp` subpackage estimates the SAR backscatter distribution to find neighborhoods of statistically homogeneous pixels (SHPs) using the generalized likelihood ratio test from @Parizzi2011AdaptiveInSARStack or the Kolmogorov-Smirnov test from @Ferretti2011NewAlgorithmProcessing.
 - The `phase_link` subpackage processes the complex SAR covariance matrix into a time series of wrapped phase using the CAESAR algorithm [@Fornaro2015CAESARApproachBased], the eigenvalue-based maximum likelihood estimator of interferometric phase (EMI) [@Ansari2018EfficientPhaseEstimation], or the combined phase linking (CPL) approach from @Mirzaee2023NonlinearPhaseLinking.
 - The `unwrap` subpackage exposes multiple phase unwrapping algorithms, including the Statistical-cost, Network-flow Algorithm for Phase Unwrapping (SNAPHU) [@Chen2001TwodimensionalPhaseUnwrapping] and the PHASS algorithm (available in the InSAR Scientific Computing Environment [@Rosen2018InSARScientificComputing]).
-- The `timeseries` module contains basic functionality to invert an overdetermined network of unwrapped interferograms into a time series and estimate the average surface velocity.
+- The `timeseries` module contains basic functionality to invert an overdetermined network of unwrapped interferograms into a time series and estimate the average surface velocity. The outputs of `dolphin` are also compatible with the Miami INsar Time-series software for users who are already comfortable with MintPy [@Yunjun2019SmallBaselineInSAR].
 
-The outputs of `dolphin` are also compatible with the Miami INsar Time-series software for users who are already comfortable with MintPy [@Yunjun2019SmallBaselineInSAR].
+To meet the computational demands of large-scale InSAR processing, `dolphin` leverages Just-in-time (JIT) compilation, maintaining the readability of Python while matching the speed of compiled languages. The software's compute-intensive routines use the XLA compiler within JAX [@Bradbury2018JAXComposableTransformations] for efficient CPU or GPU processing. Users with compatible GPUs can see 5-20x speedups by simply installing additional packages. `dolphin` manages memory efficiently through batch processing and multi-threaded I/O, allowing it to handle datasets larger than available memory while typically using a few gigabytes for most processing stages. These optimizations enable dolphin to process hundreds of full-frame Sentinel-1 images with minimal configuration, making it well-suited for large-scale projects such as OPERA.
 
-While `dolphin` has been primarily tested on CPU-based workstations and cloud environments, the compute-intensive routines leverage the accelerated linear algebra (XLA) compiler within `jax` library  [@Bradbury2018JAXComposableTransformations]. This means that with only an extra installation, users can accelerate their processing by 5-20x on machines with a compatible GPU.
-
-The `dolphin` command line tool provides an interface for running the end-to-end displacement workflow on large datasets (e.g. hundreds of full-frame Sentinel-1 images) with minimal required configuration.
-For example, if a user has created a stack of coregistered SLCs in a `data/` directory, they only need to follow two steps to run the full workflow with all default parameters:
+The `dolphin` command line tool provides an interface for running the end-to-end displacement workflow. To illustrate, if a user has created a stack of coregistered SLCs in a `data/` directory, they only need to follow two steps to run the full workflow with all default parameters:
 
 1. Configure the workflow with the `config` command, indicating the location of the SLCs, which dumps the output to a YAML file: