Hi everyone!
I’ve recently had the chance to compare a MacBook Air M4 and a MacBook Air M5, both with 24 GB of RAM, focusing on the kind of work I actually do: prototyping and relatively light Data Science / Machine Learning workloads. I thought it might be useful to share some real-world numbers for anyone wondering what kind of performance uplift to expect between these two machines.
My tests were not aimed at benchmarking the machines in an artificial way, but rather at seeing how they behave in practical workflows.
The first test was an LSTM training run using GPU acceleration with TensorFlow and tensorflow-metal. In this case, the MacBook Air M5 completed the training in 10 minutes 59 seconds, while the MacBook Air M4 took 13 minutes 47 seconds.
That puts the M5 at roughly 25% faster for this workload, which is a meaningful gain for this kind of GPU-assisted prototyping. It is not the kind of jump that completely changes what you can do, but it is enough to make iteration feel noticeably quicker, especially when you repeat this kind of experiment many times.
The second test was much more CPU-oriented. I ran a training workflow involving several classical ML algorithms such as Random Forest, XGBoost, kNN, etc., together with cross-validation and random search. In other words, this is the sort of job that leans heavily on CPU rather than GPU.
Here, the MacBook Air M4 took 15 minutes 17 seconds, while the MacBook Air M5 finished in 12 minutes 57 seconds.
That works out to about a 18% improvement for the M5. Again, this feels like a solid generational gain: not revolutionary, but clearly noticeable in real usage.
So, based on my tests, the takeaway is that the M5 seems to offer something in the range of 15–25% better performance than the M4 for prototyping and light Data Science / ML workloads, depending on whether the task is more GPU-leaning or CPU-leaning.
What I found especially interesting was the thermal behavior. I monitored both systems using Stats. On the M5, maximum CPU/GPU temperature did not go beyond 80 ºC, whereas on the M4 it went above 90 ºC. That is quite a large difference for fanless machines running sustained workloads.
I have to say that this result surprised me. Is it possible that the M5 is genuinely running cooler under these loads? Or Stats still needs to be updated for the new M5 model and may not yet be reporting temperatures perfectly?
The improvement of M5 is not dramatic enough that I would call the M4 obsolete or insufficient, because the M4 is already very capable for this class of work. But the M5 does seem to deliver a clear generational refinement: shorter runtimes, somewhat better sustained behavior, and potentially lower thermal stress.
If someone is coming from an M4 Air, I do not think this is an essential upgrade unless they run these workloads frequently enough for those time savings to matter. But for anyone choosing between the two, or buying new, the M5 looks like the stronger option for DS prototyping and light ML work.
I’ve recently had the chance to compare a MacBook Air M4 and a MacBook Air M5, both with 24 GB of RAM, focusing on the kind of work I actually do: prototyping and relatively light Data Science / Machine Learning workloads. I thought it might be useful to share some real-world numbers for anyone wondering what kind of performance uplift to expect between these two machines.
My tests were not aimed at benchmarking the machines in an artificial way, but rather at seeing how they behave in practical workflows.
The first test was an LSTM training run using GPU acceleration with TensorFlow and tensorflow-metal. In this case, the MacBook Air M5 completed the training in 10 minutes 59 seconds, while the MacBook Air M4 took 13 minutes 47 seconds.
That puts the M5 at roughly 25% faster for this workload, which is a meaningful gain for this kind of GPU-assisted prototyping. It is not the kind of jump that completely changes what you can do, but it is enough to make iteration feel noticeably quicker, especially when you repeat this kind of experiment many times.
The second test was much more CPU-oriented. I ran a training workflow involving several classical ML algorithms such as Random Forest, XGBoost, kNN, etc., together with cross-validation and random search. In other words, this is the sort of job that leans heavily on CPU rather than GPU.
Here, the MacBook Air M4 took 15 minutes 17 seconds, while the MacBook Air M5 finished in 12 minutes 57 seconds.
That works out to about a 18% improvement for the M5. Again, this feels like a solid generational gain: not revolutionary, but clearly noticeable in real usage.
So, based on my tests, the takeaway is that the M5 seems to offer something in the range of 15–25% better performance than the M4 for prototyping and light Data Science / ML workloads, depending on whether the task is more GPU-leaning or CPU-leaning.
What I found especially interesting was the thermal behavior. I monitored both systems using Stats. On the M5, maximum CPU/GPU temperature did not go beyond 80 ºC, whereas on the M4 it went above 90 ºC. That is quite a large difference for fanless machines running sustained workloads.
I have to say that this result surprised me. Is it possible that the M5 is genuinely running cooler under these loads? Or Stats still needs to be updated for the new M5 model and may not yet be reporting temperatures perfectly?
The improvement of M5 is not dramatic enough that I would call the M4 obsolete or insufficient, because the M4 is already very capable for this class of work. But the M5 does seem to deliver a clear generational refinement: shorter runtimes, somewhat better sustained behavior, and potentially lower thermal stress.
If someone is coming from an M4 Air, I do not think this is an essential upgrade unless they run these workloads frequently enough for those time savings to matter. But for anyone choosing between the two, or buying new, the M5 looks like the stronger option for DS prototyping and light ML work.
