Understanding and comparing PixInsight Benchmarks - Forums

Topic starter

Dec 22, 2023

I've been doing some research on building a PC and have discovered the PI Benchmarks. I had been looking at the Ryzen 9 7950X3D. But I immediately noticed that the cheaper Ryzen 9 7950X consistently scored higher across the board. So, I'm curious what factors might lead to those surprising results. Would I be wasting my money going for the 7950X3D or is something else at play?

I also was surprised that the amount of RAM seemed to make little to no difference. Most of the machines were 64Gb. The 128Gb machines ranked middle of the pack. In fact, there was a 32Gb machine in the top 5. I was looking at going with 192Gb of RAM. And my understanding is that the amount of RAM makes a huge difference. So why are the 128Gb machines underperforming compared to the 64Gb and 32Gb machines? It again makes me wonder if I'd be wasting money on that much more RAM. There's obviously something I'm missing.

Well Written Engaging

Andrew Ford

Dec 22, 2023

From reviews I have seen on the X3D feature, it is basically just additional cache on-chip which can assist "small" CPU workloads which require frequent cache transfers quickly - this benefits (some) games which rapidly load/unload relatively small bits of memory but is not ideal for large dataset workloads like PixInsight.
For software like PixInsight, where you may be processing 100's of GBs of data between CPU, RAM, and SSD (e.g. when stacking), the X3D stuff doesn't really assist much. I've also seen reviewers mention the algorithm that controls the X3D cache sometimes does wonky things especially in benchmarks, sometimes choosing the wrong (slower) cache and actually hurting results.
The things that benefit PixInsight the most: A very fast (and large capacity) SSD, as many CPU cores as practical (hence why the 5900X, 5950X, and 7950X benchmark so well), and a reasonable amount of high speed RAM (DDR4 or DDR5, minimum of 32-64GB). For processes such as RC Astro's plugins (BlurXterminator, StarXterminator etc) and starnet 2 which benefit from CUDA acceleration, get a decent GPU like an RTX 3080, 3090, 4080, 4090 etc.

I would honestly stick with the 7950X or even save some money and get the 5950X.

For what it's worth, I run a 2TB 980 Pro M.2 SSD, 5900X with 64GB of Corsair DDR4 3600MHz RAM, and a RTX 3090 with the CUDA acceleration stuff set up to improve CUDA accelerated processes in PixInsight. This is already "not the fastest configuration" with the new 7950X and 6000MHz DDR5 RAM now on the market, however, I do not feel like I need anything faster at the moment. This system is plenty fast enough for me.

Some of the benchmarks you are seeing in the PixInsight benchmarks that use large amounts of RAM are configuring it as a RAM disk - I personally do not see the need to do this as M.2 storage has become so fast now and is cheaper than RAM per GB.

Also, look into adding multiple temp file SWAP directories - I added 8x duplicate temp SWAP directories and improved my benchmark scores significantly.

Helpful Engaging

Bob Rucker

Dec 22, 2023

From what I understand about the 7950X3D, the added L3 vcache primarily eliminates CPU bottlenecks associated with gaming allowing higher frame rates. Unless you plan on hard core gaming, I doubt PI would experience any performance lift since the core and thread count is identical.

I’ve played around with the benchmarks while building my new imaging rig. CPU and memory is important but similar rigs can have large benchmark variations due to swap file optimizations. In addition, HDD speed is a major performance factor. I upgraded to gen 4 high speed (7,000MB/s) NVME drives which provided a huge performance boost.

I take the PI benchmarks with a grain of salt. Certain upgrades like memory seem to have diminishing returns. 64MB has been adequate for my needs but I’m sure things might be different with a full frame camera vs. my 4/3rd sensor. Beyond that, high core CPU’s and high speed drives provide the best real world processing speed. If you use Russell Crowmans tools, an Nvidia GPU with lots of CUDA cores will speed up his tools dramatically. It’s kind of a pain to configure but well worth the effort.

Helpful

Otto Vehviläinen

Dec 22, 2023

I have a 5950x with 128GB RAM and a 3070 GPU and I run M.2 disks as well. I’m not sure what the practical limits are but at least the amount of memory isn’t that. The GPU operations are of course heavy and it feels like adding a faster GPU would improve things, but going from 20s to 10s per operation would have quite high price tag. I’ve seen my memory usage never peak above 64GB, so I’d look more at the speed and latency of the memory than the amount beyond that. Stacking and preprocessing is both CPU and I/O intensive so fast disk and a lot of cores rules once you have maybe more than 32GB of memory.

Helpful

Konrad Krebs

Dec 22, 2023 · Edited: Dec 22, 2023

I run a 5950x with 128 GB of ram with PCIe4 SSD. I upgraded from 64GB because it was constantly filled to the maximum. I think it depends in your files and amount of them. I have a DSLR with very large files and if you have a huge amount of them the bigger ram will help.

Edit: And what I recognize right at the moment on another machine, when you deal with DSLR data and want e.g. to Blink, your ram will be full quiet fast and then it will work with SSD space instead.

Concise

SemiPro

IOTD Staff

Dec 22, 2023

https://pixinsight.com/forum/index.php?threads/experimental-tensorflow-gpu-acceleration-repository.22325/

If you have a Nvidia GPU that is relatively new, you may want to check this out. It massively speeds up processes like StarNet, NoiseX or BlurX.

Well Written

Jaymz Bondurant

Topic starter

Dec 22, 2023

I appreciate the responses. Very insightful. I plan on “downgrading” to the 7950X.

The only thing I’m iffy on is the GPU. I’ve seen others using old GPUs that aren’t even made anymore who say the XTerminator programs went from 6 minutes down to 30 seconds. So I’d planned on going with the cheap GTX 1650. But I’ll admit I know next to nothing about PCs. So if anyone wants to play Devil’s Advocate, I’d be willing to hear your arguments.

Franklin Marek

Dec 23, 2023 · Edited: Dec 23, 2023

I had a GTX 1660 super and found the acceleration of the RC tools over just the cpu was great for normal sized files. I then got a RTX 3070 on the used market and it was like flipping switch. I can use it to accelerate in decent time very large images like mosaics. Just my two cents but you really dont want to build a great PC with top tier cpu and ram then throw a GPU in that someone would probably not even use for a display output on a server.

Jaymz Bondurant

Topic starter

Dec 24, 2023 · Edited: Dec 24, 2023

I have another question I thought you guys might be able to answer rather than me starting a new topic. I've compared the benchmarks for all the available components of this build and have come to another crossroad. Now that I know that two components with similar specs can perform very differently, I'm not sure how to weigh my options. What's more important; the amount of RAM or the efficiency of the RAM? On one hand, I've found the 2x32Gb of RAM with the highest benchmarks. On the other hand, there's a 2x48 with slower speeds and lower benchmarks for right around the same price. The 2x48Gb was my original choice before discovering benchmarks. So is it better to sacrifice some RAM for better efficiency or sacrifice a little speed to get more RAM?

Well Written Respectful Engaging

Ruediger

Dec 24, 2023 · Edited: Dec 24, 2023

The PI benchmark is not very memory intensive. So no challenge to RAM amount, but to RAM speed.Systems with more than two modules (typically 128GB with 4 modules) run on lower RAM speed and latency. You lose easily 10% or more in memory transfer and this penalty you will notice/measure everywhere.

You really should measure, whether you need more than 96 GB RAM (biggest amount possible with 2 DDR 5 modules). To buy 128 GB because you think you might need it, may lead quite to opposite effect.
I have a fullframe ASI, and never needed more than 64GB.

CS
Rüdiger

PS. Looking at the task manager is not sufficient. A smart operating system always tries to make use of ALL memory. E.g. for disk caching. So to determine the actual need, is a bit more sophisticated. A good guess is also having a look on the amount of images you have in a typical stack and whether you work with drizzling. Based on that you can calculate an educated guess.

Helpful Insightful Respectful Engaging

Jaymz Bondurant

Topic starter

Dec 25, 2023

Ruediger:
To buy 128 GB because you think you might need it, may lead quite to opposite effect.
I have a fullframe ASI, and never needed more than 64GB.

That’s definitely something to look into. I actually planned on going with no less than 128Gb. Even with the 48Gb, I planned on going with 4x48 for the full 192Gb. I know very little about PCs and am just now learning about this stuff. Everyone says about RAM, “the more, the better”. That’s why I was leaning toward the 48Gb-5600 over the 32Gb-6000 originally. I’m using the ASI2600MC right now but plan on upgrading to the ASI6200MM eventually. I’m also using an 812mm focal length. So I’ll never be undersampled enough to drizzle. I’ve got a while before I actually build a PC. I wanna make sure I do plenty of research before I spend big (especially if it’s not needed).

Monty Chandler

Dec 25, 2023

I run the 7950X with 64gb of ddr5 memory. I looked into getting 128gb, but the speed of ddr5 memory vs ddr4 led me to think that 64gb would be enough and it has proven to be true. There's also a timing issue with ddr5 memory if you're using 4 cards vs 2, so I went with two 32gb ddr5 cards. cheers

Helpful Concise

Jaymz Bondurant

Topic starter

Dec 25, 2023

Monty Chandler:
I run the 7950X with 64gb of ddr5 memory.

That’s exactly what I’m looking at at this point. Both of my options were DDR5. But one was a 2x32-6000 and the other was a 2x48-5600. My biggest issue is not knowing exactly how these options affect PI. Sure, you could tell me how long a stack takes on your PC. But there are way too many variables that make comparing them meaningless. I’m thinking that maybe, as I get closer to an actual build, I’ll see if I can supply a small data set to a few people and ask them to stack it using identical settings and then compare those results with my little i5 laptop. My biggest fear is spending a large chunk of change just to shave a measly 10% off my stacking time.