From next year’s (2027) release onwards, MacOS will no longer include Rosetta 2 to run Intel-based software on Macs with Apple Silicon. All Mac software needs to be natively written for the Apple ARM-based chips. In June, PixInsight came out with a complete overhaul of its software. It now runs natively on Apple Silicon Macs. This started with version 1.9.4. The better fit for Apple Silicon meant substantial performance benefits. But along with this overhaul, a number of processes were completely rewritten as well, such as image registration, local normalisation and several others. Three months after 1.9.4, an even faster version 1.9.5 was released. And just a few days ago version 1.9.6 was released. The performance gains could be well felt across the application.
The current release of macOS Golden Gate is the first version that cannot be installed on Intel-based Macs. So it was about time to upgrade my main desktop computer, a trusty old Intel-based Mac Pro from 2019. Apple recently released the long-awaited Mac Studio with M5 Ultra chip, the most powerful Mac in the line-up. The version I chose was the base model running 30 CPU cores, 64 GPU cores and 96GB of unified memory. With today’s memory prices, the next step up in internal memory (256GB) would be prohibitively expensive and not worth it for my use. I did treat myself to a 2GB internal drive, even though most of my day to day work is done on removable drives.
With so many changes in hardware and software, it felt right to do some objective testing and benchmarking, to see where improvements were made and how big these improvements actually were.
Methodology
Besides the two desktop computers, Included in the performance tests was also my MacBook Pro running an M2 Max chip. As for software, only PixInsight 1.9.3 and 1.9.6 were used, in the version (x64 and ARM64) as applicable to the respective computers.
The computers were as follows:
Mac Pro 2019 - processor 3.2 GHz 16-Core Intel Xeon W - GPU AMD Radeon Pro W5700X 16 GB - Memory 160GB
MacBook Pro - processor Apple M2 Max 12 core - GPU 30 core, unified RAM - Memory 64GB
Mac Studio - processor Apple M5 Ultra 30 core - GPU 64 core, unified RAM - Memory 96GB
The software versions were as follows:
PixInsight for macOS (x64), version 1.9.3 Lockhart, build 1646
PixInsight for macOS (x64), version 1.9.6 Lockhart, build 1713
PixInsight for macOS (ARM64), version 1.9.6 Lockhart, build 1713
On Apple Silicon Macs, version 1.9.6 is available in both x64 and ARM64. For this testing, any reference to 1.9.6 on Apple Silicon means the ARM64 build.
The basis for testing is the preprocessing using WeightedBatchPreProcessing (WBPP) of a dataset of 240 full-frame LRGB images (122MB each). This included 150 Luminance images and 30 images for each of the colour channels. Master Darks and Flats were used for calibration. Subsequent process testing was done on the masters created by WBPP. Timings were read from the Process Console when available. Processes that did not register total duration times were timed by hand using a stopwatch. When tests were running, no other tasks were performed on the same computer. Processes and scripts were run in their default settings unless specified otherwise.
PixInsight Benchmark
The simplest benchmark is the one that comes with PixInsight itself in the form of the script ‘PixInsight Benchmark’. Results from this test can sometimes vary, so four benchmarks were run, and the three best scores were averaged.
The Mac Pro, at the time a real speed monster, really shows its age here with scores just around the 20,000 mark. But the M2 Max did not much better with version 1.9.3. Processing on the Mac Pro was faster, but disk access was better on the M2 Max. And on average the score was very similar. It was with version 1.9.6 that both machines showed very different responses. The Mac Pro improved by about 21% but the M2 Max did almost twice as well with this Apple Silicon optimised version. Much better handling of CPU and GPU related tasks more than doubled the score for the Apple Silicon.
The M5 Ultra was in a league of its own. The total score of 106,997 was almost three times as high as the M2 Max. And this score was ‘held back’ by a swap disk score that was ‘only’ twice as fast as the M2 Max. The internal drive on the M5 Ultra has read/write speeds of over 12,000 Mbps. That is insanely fast. But the internal drive of the M2 Max has read/write speeds of 5,000-6,000 Mbps, not a slouch either. A two-fold increase in swap disk speed makes sense when compared to the underlying disk read/write speeds. The CPU score was 3.5x higher than the M2 Max, which is just an insanely high number.
Benchmark results of PixInsight can also be uploaded to the PixInsight website and compared with other systems. As it turns out, the M5 Ultra Mac is the first computer system to ever score a result of more than 100,000. For many years, Linux systems were considered the golden standard for PixInsight when raw speed was important. That apparently is no longer the case. The fastest Linux systems score just over 90,000. The fastest Windows machines remain behind at around 65,000. So current generation Macs, in combination with PixInsight written natively for the Apple Silicon processors, are very compelling choices for PixInsight.
WBPP
The task that takes the most time by far is preprocessing using WBPP. It is not unusual for this task to run for hours. But both the new versions of PI and the Apple Silicon computational power completely reset expectations here, as became clear from this test. First let’s look at the Intel performance on the Mac Pro. The full execution of the preprocessing took 1h 49m on version 1.9.3 and this is very typical for how I have experienced it over the past years. But the many improvements in core processes that were introduced in version 1.9.4 through to 1.9.6, add to a significant speed boost. The same set can now be done in only 19 minutes on the same machine.
The M2 Max took 3h 33m for the full set to be preprocessed under version 1.9.3. This is more than twice as long as on the Mac Pro. This fits with my experience that in the past the M2 Max was generally slower than the Mac Pro with WBPP. This may also be attributable to the difference in internal memory, which is 160GB for the Mac Pro versus 64GB for the M2 Max. Less memory means more disk swapping. A smaller dataset, or data from a smaller than full-frame sensor might show less dramatic difference. But whatever holds the M2 Max back under version 1.9.3, any limitation seems to be completely gone under version 1.9.6. It processed the full set in a very respectable 22m. That is in line with the Mac Pro with a much larger internal memory. And it represents a 90% speed improvement compared to version 1.9.3 on the same machine.
The M5 Ultra was done with the whole set in just over 8 minutes. That is 2.5-3x faster than either the M2 Max or the Mac Pro running the same version. Clearly its higher compute power, bigger memory and massive memory bandwidth make it a fun experience to just sit there and watch the pipeline being processed in seconds per line-item, rather than minutes. About half the time was spent on image integration. Note that by default I turn off fast-integration. The quality improvement is minimal, so there are good arguments to be made to turn it on and save a minute or two because of it. Also I have large scale pixel rejection turned on, to improve removal of plane and satellite trails. Turning that off will reduce total processing time by roughly another minute.
Of all the tests presented in this blog, WBPP is the most sensitive to local circumstances. The number and size of the files going into the script have a major effect on all steps in the process. Individual settings for processing steps, selection of which processing steps in the first place, and any manual interference all contribute to significant differences in total speed. And PC configurations can have different effects, depending on the factors mentioned before.
The conclusion here is clear, and that is that whatever system you’re using, the current version of PixInsight is a night and day difference compared to earlier (v 1.9.3 and before) versions. Was WBPP in the past the process that you let run overnight, in its current form it will often be done within an hour. And half an hour or less is well within reach for many. Due to the speed limitations in the past, some users reached out for different tools to do the pre-processing faster, to only come back to PixInsight for the processing of the master stacks. That need is now completely gone.
The WBPP results of my old and trusty Mac Pro from 2019 running PixInsight 1.9.3. In just under 2 hours all 240 images were fully pre-processed and stacked per colour channel. This is the kind of performance that has been very common to me for most of my time in astrophotography.
The same dataset now processed on the new M5 Ultra running PixInsight 1.9.6. With just over 8 minutes to complete, that is 13x improved performance.
Blink
There are few tools in PixInsight that can get users worked up like Blink can. And while the tool could do with some improvements, I have forced myself to learn and use most of the functions in the tool. And overall it works remarkably well in my workflow. The only drawback is that loading a full night of images, which could be well over a hundred, is just uncomfortably long for a quick review of the data. Luckily also here version 1.9.6 has shown some very welcome speed improvements. Overall it is about 3-4x faster than before. In general the Mac Pro is a bit slower than the M2 Max, the latter of which can load the full set of 240 full frame images in less than half a minute. The M5 Ultra offers some reasonable speed improvements here and is about 30% faster than the M2 Max. It loads the full set in only 19 seconds.
I have been using external review tools as well, most notably the software FITS blaster. And while its faster and provides more detailed quality information, the main reason I’m using it is the fast extraction of fits header data that give me quantitative information on weather conditions during the night, focusing performance, etc. But for just reviewing images, Blink is still my preferred choice, not least because it is just within the same platform. And with load times in the order of seconds, Blink has become an even more attractive tool.
Processes
Other than known time-consuming processes such as WBPP, most users will spend most of their time in PixInsight running many different individual processes on individual files. So to assess improvements in this area I selected six processes that most people will use on a regular basis, and that have shown in the past that they can take some time to complete. These processes include ImageSolver, SpectroPhotometricColorCalibration (SPCC), GraXpert, BlurXTerminator, NoiseXTerminator and StarXTerminator.
Also here we see some dramatic improvements of version 1.9.6 over version 1.9.3, irrespective on the machine on which it is run.
ImageSolver is known to be completely re-built and the results show as much. In version 1.9.6 it is 4-6x faster compared to version 1.9.3. And the 4s it takes to run the process on the M2 Max is absolutely fabulous. One claim that is not tested here is the ruggedness of the process. In earlier iterations ImageSolver could sometimes just not find a solution, no matter how hard you tried changing parameters. This is hard to test, and I did not make attempts to do so, but so far in the newer versions I have not had ImageSolver fail me once. It is probably hard to beat such a fast performance, but the M5 Ultra was able to find an astrometric solution in only 2s, which is 2x faster than the M2 Max.
SPCC showed significant speed improvements as well, about 2-3x faster. And also here, the M5 Ultra was cutting the time in half compared to the M2 Max, completing the whole process in 4s.
GraXpert is a bit of a different process, as it relies on an external application. Opening that application can sometimes take more time than the actual gradient correction calculation. For that reason I often use an image container and run GraXpert on all images in that container. The first image takes a bit of time, the second and later images are done much faster. As this is an external process, there is little to no difference between PixInsight versions. Any differences are reflecting the differences in computing speeds between the three machines. The M2 Max is about 3x faster than the Mac Pro and the M5 Ultra is 2x faster than the M2 Max. On the M5 Ultra, the process took 4s to complete.
The RC-Astro ‘terminators’ have become core tools in many astrophotographers tool-kits. They are machine-learning based tools and therefore benefit from computers that have graphics cards or dedicated neural engines that are optimised to work with machine-learning tools. The PixInsight versions have little effect here as most of the heavy lifting is done by the tools themselves, which are developed independently from PixInsight. Their development is mostly in improving the models and my experience is that with newer and bigger models these tools work slower rather than faster with newer iterations. Although we may have seen an exception to this for StarXTerminator.
The Mac Pro is not optimised at all for such work and with recent iterations of more complex models, processing time was well over a minute. The M2 Max with integrated GPU and neural engine is a lot faster and typically gets the job done in about 20s, a 3x improvement. The M5 Ultra with its 64 core GPU and 96GB of available unified memory is designed from the ground up for AI and AI-related tasks. It will come as no surprise that the RC-Astro tools work super-fast on this machine, and finish in 6-9s, another 3x improvement compared to the M2 Max.
One note to make is that StarXterminator has seen some speed improvements that are not related to the computer but to the tool itself. In previous iterations the tool was typically slower or as fast as the other tools. The current version is consistently the fastest of the three. It is not clear to me if this is an improvement in the tool itself, or that it has to do with the way it interacts with the latest version of PixInsight.
All these numbers may be dazzling a bit, and one could argue that a few seconds here and there may not mean much for real-world use. But it is the real world use where the upgrades in PixInsight and the faster computers matter most. If a tool takes 2s instead of 5s, but its used a lot, all these seconds add up. And the much snappier feel of all these tools means less waiting and more focus on the processing. Previews work faster, faster processes entice to try out different settings, provide faster feedback and lead to more experimenting overall. The result is higher quality images and a more joyful experience in the process.
Conclusion
Over the last half year, PixInsight has seen tremendous evolution. Across the board the speed has been improved, new processes have been added, existing processes have been enhanced, and for Mac users there is now native support for Apple Silicon. And when each iteration shows only some improvements, it is difficult to see the overall performance leap. This test was focused on speed improvements. Many astrophotographers spend more time behind the computer processing images than they spend in the field taking them. So speed improvements help us all. And on Macs, PixInsight has become roughly 2-3x faster than it was, regardless of the machine it is running on.
Macs have also become more powerful in recent years, with efficient ARM-based processors, unified memory, neural engines and lots of processing cores in both CPU and GPU. For years the fastest machines to run PixInsight were Linux-based computers. But that has changed. The fastest PC to run PixInsight is now a Mac with M5 Ultra processor. Looking at the published benchmark data, most current Mac models are as fast as or even faster than the Linux machines. And they pretty much all outperform their Windows-based counterparts.
From a personal perspective, with the M5 Ultra and the latest PixInsight version, I now have a processing system that is roughly 10x faster than the system I was using less than half a year ago. That is a massive improvement and I look forward processing many images with it. And you will see the results on this website as well as on my Astrobin channel.