Are AMD processors better than the Intel ones for PixInsight?

Hi Jeff,

Many thanks, but is not the processor more important than the GPU? And also more the RAM than GPU?
I know about this GPU acceleration and I post a doubt in another thread ( https://www.astrobin.com/forum/c/astrophotography/equipment/gpu-cuda-yes-but-memory-matters/ ) but I think that GPU acceleration with NVIDIA CUDA is not working in all the processes, just in some of them like StarNet, or StarXterminator but for example I don't think so in WBPP (maybe I'm wrong).

Clear Skies!

Daniel.

You can empirically compare the Pixinsight speed of different processors and other hardware combinations Here:

https://pixinsight.com/benchmark/

Kevin Morefield:
You can empirically compare the Pixinsight speed of different processors and other hardware combinations Here:

https://pixinsight.com/benchmark/

Thank you, Kevin. I'll ckeck this link ;)

Die Launische Diva:

Daniel Arenas:
Hi Jeff,

Many thanks, but is not the processor more important than the GPU? And also more the RAM than GPU?
I know about this GPU acceleration and I post a doubt in another thread ( https://www.astrobin.com/forum/c/astrophotography/equipment/gpu-cuda-yes-but-memory-matters/ ) but I think that GPU acceleration with NVIDIA CUDA is not working in all the processes, just in some of them like StarNet, or StarXterminator but for example I don't think so in WBPP (maybe I'm wrong).

Clear Skies!

Daniel.

PixInsight does not use the GPU for the heavy processing yet. Only a limited number of third-party plugins can use the GPU. These plugins are trendy nowadays (I know...) but are actually they are not a required at all for producing a final image. Unless you are doing anything else with your computer which requires a decent GPU, your only concern as a PI user is to have many cores, lots of memory and fast storage. Regarding price per core (for meeting the first requirement), an AMD processor is hard to beat. My AMD-based Beelink "jewelry box" costs ~600€, scores ~18200 at the PI benchmark and suits all of my needs. I can see no reason spending 5x-10x more to be at the top of the PI benchmark.

All right, Die Launische Diva,

I think I understand you, AMD has more "normal cores" and Intel make differences between performance cores (p-cores) and efficient cores (e-cores) so not all the cores in intel are going to be used at the same time in a heavy process (if I'm not wrong).

Clear Skies!

I built my PC in 2018 and runs off an old Intel i5 6 core chip. The GPU is also not the greatest but was more middle of the road back then. Pixinsight runs just fine for me with my outdated equipment. I did add on Cuda for gpu processes like RC-Astro plug ins etc and I recently upgraded my RAM as this was the cheapest upgrade I could do outside of building a new rig.

My old CPU is the limiting factor (or bottle neck) here. Running PI is not a bother for me. I know which scripts/processes take the most time (subframe selector and wbpp) so when I run those I'll read or do a few chores around the house then return. I am seeing some speed boost in wbpp compared to my old RAM sticks but even with 32gb its my CPU that would need to be changed in order to see any real difference.

Benny Colyn:

• A CPU with good multi-threaded performance helps for most PI processes. You can somewhat compare their relative performance here: https://www.cpubenchmark.net/singleCompare.php but as always take benchmarks with a grain of salt. I have found however that the relative numbers for multi-threaded performance largely hold for PixInsight. With that in mind, shop around and compare prices - see how much you are willing to pay for the extra performance.
• PixInsight likes writing temporary data to disk, even with ample RAM available. Therefore a speedy SSD helps a lot (particularly for WBPP).
• A GPU is only used if:
• You use one of the Tensorflow-based processes (only RC-Astro's plugins and Starnet)
• You installed the CUDA-enabled tensorflow.dll + CUDA libraries
• You have an NVidia GPU

Given that a GPU does nothing for the other processes, the price of GPUs nowadays, and the fact that CUDA support on Windows has already been removed from the latest Tensorflow versions (only Linux going forward - see https://www.tensorflow.org/install/lang_c) meaning this hack will probably stop working when RC-Astro or Starnet starts to require Tensorflow 2.11 or higher - I would not invest in a GPU unless you want/need one for other purposes (gaming?) or you run PixInsight on Linux.

Did not know that about Cuda support on Windows, bunmer. But I agree with all the above regarding CPUs. I installed a new SSD drive on my PC which I use only astrophotography processing in PI which had helped some before boosting my RAM.

I built a PC specifically for PixInsight processing two years ago.  I’m using the AMD 5950x CPU, and an NVidia 3060 Ti GPU.  I have two NVMe drives, one for the system, one for PI data processing and 64GB RAM.  

If I were building a machine today, I’d use the 7950x CPU, NVidia 4070 Ti GPU, Gen 4 NVMe system, dual Gen 4 NVMe (RAID 0) data, and 128GB RAM.

I tend to process large numbers of images, running WBPP on 500+ subs is common for me, and I’ve done several projects with more than 1000 subs so the additional cores, faster storage, and more RAM are definitely helpful in reducing overall time to process.

Something I would recommend against is using a laptop, because they are always inferior to desktops when it comes to cooling.  On paper the specs of a laptop may look similar to a desktop, but their inability to cool as effectively means you never see that actual performance, or if you do it’s only briefly until the system has to step down speed due to excessive heat.  PI, when running WBPP on a large number of subs, will max out your CPU for hours.  You need a desktop with AIO Liquid Cooling to manage that heat to keep the CPU operating at maximum speed.  No laptop can do that, and will step down the CPU speed to a point that it can keep cool.

Die Launische Diva:

Daniel Arenas:
All right, Die Launische Diva,

I think I understand you, AMD has more "normal cores" and Intel make differences between performance cores (p-cores) and efficient cores (e-cores) so not all the cores in intel are going to be used at the same time in a heavy process (if I'm not wrong).

Clear Skies!

I believe this is incorrect. At the time PI developers were writing the System Requirements, AMD had the advantage of offering more cores per buck. This may still be true nowadays, but since this is all about your money and how to spend it, you have to do your own research. Kevin's advice on having a look at the PI benchmark webpage is great.

My advice is to be conservative when spending money on hardware since very soon, even the high-end hardware will be obsolete and will need replacement. My personal experience from fellows working in academia, (and others), is than most of the time their expensive hardware is underused or not necessary at all. I see no reason to invest on a cutting-edge computer, unless it is the tool for your work and puts food on your table.

I was talking about the differences on how intel and AMD uses its cores and wondering if the Intel strategy is more or less interesting for PixInsight.

I totally agree with you about the obsolescence. Maybe using last technologies in mother boards, RAM and processors can give you some more years but computes evolve a lot in short periods of time son maybe in 5 years the requirements for AI or something else makes the actual configs not good for that time.

Jason Coon:
Something I would recommend against is using a laptop, because they are always inferior to desktops when it comes to cooling.  On paper the specs of a laptop may look similar to a desktop, but their inability to cool as effectively means you never see that actual performance, or if you do it’s only briefly until the system has to step down speed due to excessive heat.  PI, when running WBPP on a large number of subs, will max out your CPU for hours.  You need a desktop with AIO Liquid Cooling to manage that heat to keep the CPU operating at maximum speed.  No laptop can do that, and will step down the CPU speed to a point that it can keep cool.

Totally agree, and that's one of the reasons I want to buy a desktop computer.

You should also remember that usually you run the WBPP once at night, or day, and you don't need to look at it. But the post processing often requires revoking and repeating the same actions, so the proper GPU may help. This is in my case.

My 2 cents worth. I have been using an AMD Ryzen 7300X based system since ~2019 when that processor first came out. It has worked quite well for me and still gets the job done very well for PI processing. That being said, I feel like a similar Intel CPU would have been pretty equivalent. 

I am assuming you are talking about a desktop system (TL;DR above - maybe you mentioned that already…)

Personally, I think if you are comparing like priced systems with CPUs of similar costs, you will not notice big differences between CPUs. As mentioned in posts above, multi core processors are well utilized in PixInsight, especially in the preprocessing phases when the application is crunching through lots of files.  The news about tensorflow removing GPU support in the future is unfortunate, and does probably mean that whatever GPU is included in the system you pick is adequate, although I would still prefer at least a separate card implementation rather than something that is incorporated into the motherboard, for upgrade purposes down the line if nothing else. After those things the disk drives are the bottleneck, and so having an NVMe internal drive to process your data on will make a huge difference over other approaches. These are getting cheap, and a 1TB one is more than adequate form your 2nd drive to be used for data processing. I would get a 2TB one for the main boot drive. Lastly, you will want some longer term storage/backup for your data after you have finished processing. External drives of various versions work well for that.

ML

How many cores you have really is the answer.  Even my Biggest WBPP stacks have never exceeded 2 hours, and that was when I unecessarily ran LN.   IMX571 data

i9-12900k, 32GB ram, RTX 3090, all SSDs, but pix runs on my slower SATA drives.   The new Ryzens take the cake on stacking times, but Intel still hangs.    I just ran a full frame 2 panel mosaic, StarX and BlurX all were around 1 to 1.5 mins, longest process was LHE and it took roughly 2 mins.   My Embryo stack was something like 500 frames and it took somewhere around an hour.