Navigating the trade-offs of using server-grade Xeons for PixInsight and heavy stacking?

I had a similar problem when using GESD rejection and switched to Winsorized Sigma Clipping when stacking a lot of frames. I couldn't see a quality difference and it cut the stacking time from 16+ hours to under 2 hours.

It’d be helpful to know how many subframes and what size they are to know your use case.

I’ve run 1000+ subframe integrations with an IMX571 sensor and it usually takes less than an hour on my Mac Studio, so if things are taking your computer 16 hours, I’d expect you’re processing 10,000+ subs at a time.

Some three years ago I used a dual Xeon workstation (HP Z8 G4, 2 x Xeon Gold with 18 cores each, 72 threads total, 192 GB RAM, 2 × 512 GB NVME SSDs, Windows 10/11 for workstations).

PI then could not properly handle a dual CPU setup (two sockets) with 72 threads as the performance with hyperthreading on (72 threads) was lower than with hyperthreading off (36 threads). Two years later I built a new PC for PI, using an Intel 14900K CPU and 64 GB of RAM. This new PC is about 2x faster in all PI chores than the old workstation (and at a fraction of its original cost).

So unless you have an older Xeon workstation hanging around it is not worth the hassle in my opinion.

I did the same some time ago. I did a mosaic with 190 panels, but I used Astropixelprocessor instead. I bought used server-parts and built a machine with 768 GB RAM and two CPUs with 28 logical processors/54 threads over all.

In it‘s final stage this machine took 36 hours to finish 190 panels. But it worked.

I still have that machine, selling the RAM nowerdays would more than pay back all investments. 🤔

Hey Jack,

Second hand server rig is definitely the way to go for a cost effective astro processing machine.

I’ll only warn about AVX2 support, depending on how old the CPU is it might not have AVX2 support, and if it doesn’t then you can only install up to Pix 1.8.9.

I do not know how big are your stacks, but a general rule of thumb is the more RAM/cores the better.

I do not know what you currently have, but 14 might be similar to a 8 core CPU @5Ghz (tho this is not linear but gives an idea), so you might want to increase core count if you can.

As you noted server cores are slower than desktop cores, so some things (single threaded ones) will be slower but the most demanding ones will be faster as you compensate speed with more cores.

Last advantage of servers is ECC RAM, which will guarantee stability.

To give perspective my desktop has 32 cores (AMD threadripper) and 256GB RAM and I do max them out with “as little” as integrating 600 full frame images.

My two servers have 24 cores and 512GB RAM each. I tend to run automated pipelines and bespoke software I develop on them.

So long story short, get as much RAM/cores as you can and be careful that the CPU has AVX2 support.

Also I noticed, if you find the sweet spot (like won’t be true for DDR5), second hand market is not so much affected by current RAM shortage and high prices.

Cheers and CS,

Carl

Jack Hicks · Jun 29, 2026, 11:42 AM

Hey everyone,

I’m finally reaching a breaking point with my current image processing workflow. I’ve been using a fairly mid-range consumer laptop for the last two years, but as my integrations are getting longer and my camera resolution has increased, the "wait time" has become unbearable. I recently spent nearly sixteen hours just trying to run a drizzle integration on a large stack of Subframes, only to have the system hang right at the end. It’s heart-wrenching to wait that long and have nothing to show for it.

I’m looking into building a dedicated, budget-friendly workstation using some refurbished enterprise gear. I’ve come across a deal for an older Xeon 14 Core / 2.3GHz-9.6GT-QPI processor. On paper, having 28 threads sounds like a dream for multi-threaded applications like PixInsight or even just massive batch conversions in Siril. However, I’ve always been a bit wary of the lower clock speeds on these server-grade chips compared to the high-boost consumer CPUs we usually see.

One specific point I’m trying to verify before I pull the trigger is the impact of that 9.6GT/s QPI (QuickPath Interconnect) speed on our specific type of class="ng-star-inserted" style="box-sizing: border-box; font-size: 14px; line-height: 21px; font-family: Inter, sans-serif; font-optical-sizing: auto; font-weight: 400; margin-bottom: 18px; color: rgb(43, 45, 49); font-style: normal; font-variant-ligatures: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; white-space: normal; background-color: rgb(255, 255, 255); text-decoration-thickness: initial; text-decoration-style: initial; text-decoration-color: initial;">On a personal note, I’ve always enjoyed the "tinkering" side of this hobby almost as much as the imaging itself. I remember my first successful shot of the Andromeda Galaxy—it was grainy, poorly tracked, and processed on a machine that sounded like a jet engine taking off. There was something satisfying about squeezing every bit of performance out of "inadequate" hardware. But now that I'm trying to do more serious science-based imaging and complex mosaics, I really need stability over sentimentality.

I’m curious to know if anyone else here has gone the "used server" route for their processing rig. Does the sheer core density of a 14-core Xeon make up for the lower frequency when it comes to the heavy math involved in local normalization or star mask generation?

Do you think the architectural stability of these older workstation platforms is still a viable "bang-for-your-buck" option in 2024, or has the IPC (Instructions Per Clock) of newer, fewer-core chips simply left these older Xeons in the stardust?

If the 16-hour-long stacking is long, then what about 38 hours?

Big loud and not sure if you will be much faster. I am suing a 14900k workstation with 128 ram.

My biggest data set was 900 subs that took a good 8 hours. all my everyday data sets of 50-200 subs finish under 2 for sure. learnt from it that LONG projects do call for the longest sub I can use without burning stars.

Palmito · Jun 29, 2026 at 01:30 PM

Hey Jack,

Second hand server rig is definitely the way to go for a cost effective astro processing machine.

I’ll only warn about AVX2 support, depending on how old the CPU is it might not have AVX2 support, and if it doesn’t then you can only install up to Pix 1.8.9.

I do not know how big are your stacks, but a general rule of thumb is the more RAM/cores the better.

I do not know what you currently have, but 14 might be similar to a 8 core CPU @5Ghz (tho this is not linear but gives an idea), so you might want to increase core count if you can.

As you noted server cores are slower than desktop cores, so some things (single threaded ones) will be slower but the most demanding ones will be faster as you compensate speed with more cores.

Last advantage of servers is ECC RAM, which will guarantee stability.

To give perspective my desktop has 32 cores (AMD threadripper) and 256GB RAM and I do max them out with “as little” as integrating 600 full frame images.

My two servers have 24 cores and 512GB RAM each. I tend to run automated pipelines and bespoke software I develop on them.

So long story short, get as much RAM/cores as you can and be careful that the CPU has AVX2 support.

Also I noticed, if you find the sweet spot (like won’t be true for DDR5), second hand market is not so much affected by current RAM shortage and high prices.

Cheers and CS,

Carl

My goal wasn't to build a fast computer, but rather one capable of handling this task with 190 panels in the first place. While looking for motherboards with maximum RAM support, I ended up in the server section on eBay. The CPUs cost €50 each. The 24 x 32 GB Registered ECC DDR4 modules cost significantly more, but at the time, a single stick was €40—whereas today, they are almost unaffordable.
But you're certainly right that the latest versions of PI no longer run on my server, whereas AstroPixelProcessor still does.

Steve Walkenshaw · Jun 29, 2026, 02:57 PM

$2700 for a server processor that is older makes no sense. Last year I built a brand new processing computer using an AMD Ryzen 9 9950x CPU. I added 128 GB of RAM (when you could get it at a decent price) and some quick SSD drives . The whole setup cost around $5000 and blows the doors off my previous computer. This year I purchased a 62MP Mono camera and while it puts a strain on it, I’m very happy.

It’s not only the CPU, but also the memory and drive setup as well. Servers are like Semi’s, they are not made for speed, they are made for serving out large amounts of data. PixInsight to me seems to like the sports car computers more, especially in terms of swap file speed and working drive speed.

Good luck with that, but I would put my $$$$ on faster drives and memory.

A good place to start is the PI benchmark results.

9950x here too, same findings, put the 9100Pro on the PCIe5 lane. 44000+ on PixInsight benchmark. If you are an Apple guy, the M3 Ultra my be the beast I’d buy.

Ok, I am going to dive into this with my own equipment as a baseline. So ymmv. I am just trying to wrap my head around how you are taking 16 hours to do a drizzle integration. Are you processing massive datasets (like 1000+ subs)? Because I just can’t comprehend this based on my use case unless you are.

Some caveats. I am not a pc / pixinsight expert. I can only use my equipment as a baseline. I did notice a significant decrease in processing time when I switched to Linux from Windows a couple years ago (to Fedora in my case). But even on my old quad core with 64 gb ddr3 ram and windows, a large (at least what I consider large e.g. 100-200 subs) dataset took about an hour or so to drizzle. Please keep in mind that is going strictly by memory.

We currently use a zwo294 in bin2 for our imaging. I recently bought a bottom of the barrel $300 Thinkpad (8 cores, 4gb RAM) to take with me to NM for our maintenance trips. Our data set for M63 was about 37 hours over 360 subs HaRGB. My laptop would WBPP the dataset in a little under 3 hours (2hrs 43mins if memory serves). These were with settings that placed quality over all else and a 2x drizzle.

I compared that to my desktop when I got home. My desktop is a 32 core intel with 128gb DDR4 RAM. It processes the same dataset in 23 minutes. So the extra cores and RAM is definitely a game changer in that regard. If this scales linearly (and I don’t know if it does), I would be able to process approximately 15,100 subs in 16 hours.

So are you just processing insanely large datasets? If not it seems like there is something else going on rather than just a pure hardware issue. At least to me, but as I stated, I am no expert. Hopefully, this is somewhat useful. If not, sorry for wasting your time.

Hi Brian,

could you tell me, if your M4 Macbook Pro has 16 or 24GB ?
CS

Jürgen