The manuscript describes an effort to refactor the MITGCM sea ice model from Fortran to Python. This is based on previous work with an ocean model (Veros) written in Python. This is a nice effort.

Unfortunately the manuscript itself is on the weak side and it seems as if the authors dives into too many things without really doing more than scratching the surface. It compares a demonstration case by Mehlmann et al (2021) with a stationary wind field and states that narrow features are observed.  Are these right or wrong? Then it simulates Antarctic in order to show that the coupled system works. But again no real validation or discussion. The largest focus is on optimization and they show scaling based on increased domain size. This Scaling is normally shown with increasing number of processors (including the timing of running on 1 processor). For climate models the limitation should be usage of the full bandwidth, which is not necessarily reached. 

Specific comments:

Line 15: This is a subjective thing. I don’t think that Python is easier to read or maintain. I think that this depends on the programmer. I agree that Python is easier accessible as it does not require a compiler and that more people have used this.

Line 77: I assume this is theoretical speaking and if all resources are used.

Section 2.2 Validation:  What is the setup for the dynamic test? 1000 iterations is a few days Is this long enough?

2.3 Parallelization

EVP cannot use JAX. How does the model speed up the dynamics in the JAX parallelization cases in the figure if it only do dynamics?

 This is normally the most expensive part.

Function fill overlap. This is the communication part and the part that adds a mpi synchronization point in addition to the compilation just in time.

Section 3.1 

Is this the same test as the dynamics? Described in section 2.2?

The manuscript "Veris: Fast and Efficient Sea-Ice Modelling in Python with GPU acceleration" describes the translation of MITgcm sea-ice component using JAX. The authors have demonstrated that Veris can be run on single GPU with low power use and a performance comparable to the original Fortran code run on several CPUs. 

The manuscript is concise and well written. The results presented under benchmarking section makes sense and authors arguments are also sensible. However, I do not find it of much relevance for the community who have been porting the code on modern architecture for more than a decade now and is aware of the advantages and disadvantages of the accelerators. Demonstration of JAX as a potential DSL is possibly new and my opinion would have been different had the code been run on multiple GPUs demonstrating usefulness of JAX on large-scale HPC. I therefore, unfortunately, recommend rejection but encourage authors to resubmit when Veris is able to run on multiple GPUs. 

I have following suggestions on the paper, which I hope authors find useful:

1. Sec 2.2: if I understood right, the section describes regression testing using MITgcm as the reference. This is code verification not validation. I am sorry for being strict here, but for a community that is trying to modernize itself, it is timely to use appropriate terminology. 
2. sec 2.2: authors indicate relative differences of 10^-7 and 10^-3 for the quantities under testing and conveniently mention that these differences are acceptable without any justification. Typically, one perturbs the initial condition to generate ensembles to get an estimate of expected errors. The observed error is then subjected to these expected errors . I suggest authors to do a literature review on the subject of testing. One good example from the community is "THE ACCUMULATION OF ROUNDING ERRORS AND PORT VALIDATION FOR GLOBAL ATMOSPHERIC MODELS∗  JAMES M. ROSINSKI† AND DAVID L. WILLIAMSON†" and the from outside the community: "Verification of Codes and Calculations  Patrick J. Roache*"
3. Sec 3.1: The results section is does not have any reference. For a non-expert in sea-ice modelling, it is difficult to evaluate correctness of the results.