Does adding cpu cores really make much difference if not rendering with them?

[Paul King]

I have a PC with an old 4 core 8 thread i7 4790K CPU, 32 GB RAM, M.2 drive SSD, and GTX980Ti GPUs.

I have been considering an upgrade to something that can handle more CPU threads, in the hope that ArchiCAD23/24 will become more responsive on complex hi poly count , multiple hotlinks etc projects.

I am a freelancer and generally have to work with projects as I find them - model structure and efficiency are often out of my hands, and I tend to have to follow whatever 'lowest common denominator' office standards are already in play when I am brought in.

I am also at the mercy of whatever imported structural and services IFC models are produced when attempting to coordinate (generally no fancy BIMcloud services on tap - just old fashioned manual IFC file exchanges via DropBox). Switching one or two of these on I find can bring any ArchiCAD session to its knees, particularly if the engineer responsible has not been very good with separating things into useful layers, meaning every bolt has to be shown when beams are switched on, and when everything appears on a single story etc.

Obviously am also impacted by whatever remote Bimserver and VPN folder structures setup I need to dial in to but appreciate more CPU cores won't help with this one.

Sometimes I need to have more than one such project open at the same time. As well as maybe a session or two for hotlinked modules being edited, the normal Email, PDF viewer and browser windows open (generally not too demanding in themselves)

Other than model efficiency, the only variable I do have much control over is my own hardware.

I don't use CPU rendering at all - all GPU based.

All things being equal, am I likely to notice much day to day benefit while wrangling large cumbersome models by switching to something with 10-12 CPU cores? Jumping between views, loading 3D geometry into GPU etc. Would more than 12 make a lot of difference? (where do diminishing returns kick in)

Any benefit at all in moving to a PCIe 4 bus, with all that data needing to be shunted between harddrive and GPU?

Any tips most welcome.

[Lingwisyer]

You could log your current computers CPU and RAM utilisation over a couple of days to see where your primary bottlenecks are. This can be done through Windows Performance Monitor.



Ling.

[Paul King]

Lingwisyer wrote:
You could log your current computers CPU and RAM utilisation over a couple of days to see where your primary bottlenecks are. This can be done through Windows Performance Monitor.

Hi, thanks for that.

Yes, I have noticed that RAM is not ever an issue in my case, but CPU utilization bumps into 100% when generating 3D views or large sections or elevations, but I also see something like 100 - 200 threads are allocated to each ArchiCAD session open

What I am not able to identify is how many of those threads are dormant at any given time vs how many are responsible for pushing overall CPU utilization to 100% during view generation - it could just be one or two threads being allocated all the work for all I can tell, suggesting additional cores may not help.

[Erwin Edel]

The examples of poor drafting seem to be mostly impacted by GPU, also in 2D plan view (redraw time).

CPU might come into play with mutliple tabs and having to regenerate.

I once had to deal with a 3rd party project where all the 2D details were on one giant worksheet and contained many, many, MANY exploded DWG files with loads of double lines etc. These were then placed as individual drawings (resized) on layout. However updating one of these layouts would mean updating the entire worksheet and that took forever.

You could hotlink the IFC export from a PLN where you fix most of the poor drafting/layers and keep importing to that seperate PLN to keep your master file clean.

Sadly, I think going through the junk of the 3rd party and cleaning it up is sometimes the best time spent v.s. time lost on updates. In the above example I purged the entire worksheet of all the double elements and split the worksheet up in to two or three seperate ones to keep the amount of 2D elements per view low enough for speedy redraw/regeneration.

I do notice Xeon processors deal better with Cinerender than i7 processors, but I've not noticed a huge difference on 2D work. The 3D window navigation seems entirely tied to the GPU.

Given the by now old hardware of my HP workstation and it still handling ArchiCAD projects well enough, I think it's more to do with the poor skills than lack of power on your setup (which seems more than adequate even for large projects).

[Paul King]

Thanks Erwin
An excuse to buy an RTX3080 maybe? (to future proof myself for ArchiCAD 35)

[Lingwisyer]

If you can get past all of the scalpers sniping the cards...

[Paul King]

Lingwisyer wrote:
If you can get past all of the scalpers sniping the cards...

Yeah. And even the 'honest' local retailers here in New Zealand are throwing a few hundred dollars on top of the recommended retail price I see.

[Erwin Edel]

I'm not 100% sure about this (as in no real testing / proof), but my old Quadro K2000 seems better equiped at 2D redraw than the newer 'gaming' mobile cards in my 2013 iMAC at home and the macbook pro machines we also have at the office. In 3D window it seems reversed.

I do like the robustness and quality of the Quadro hardware and the drivers are a breeze to update, when I feel like updating.

[Paul King]

Erwin wrote:
I'm not 100% sure about this (as in no real testing / proof), but my old Quadro K2000 seems better equiped at 2D redraw than the newer 'gaming' mobile cards in my 2013 iMAC at home and the macbook pro machines we also have at the office. In 3D window it seems reversed.

I do like the robustness and quality of the Quadro hardware and the drivers are a breeze to update, when I feel like updating.

From what I have experienced (again no real empirical testing to quote) if money is no object, the very best 2D performance probably does come from the best Quadro, but where there is a limited budget, the best performance seems to come from the best consumer Gforce card that same budget will buy - simply because with any any given Quadro you could buy with a given budget, the same budget will buy a Gforce card that is a generation or two newer and faster, and with more memory. And of course driver development is more active for consumer cards (new driver not always for the better with either Quadro or Gforce, but at least with Gforce, there are usually more recent roll back options)

Certainly I recall the odd issue with Quadro cards when using them under AC17 and 18 that went away after dropping in a newer GForce card as a replacement. I think most of the time you are generally fine with any recent card now, but very large projects may test this, especially when using trace reference a lot in addition to complex hotlinked IFC's etc. Although I notice CPU usage and disk writes both spiking when panning and zooming around and editing in demanding views - not sure if this represents a GPU bottleneck or not. The disk writing might be somewhat explained by data safety set to 'save every step'

[Nader Belal]

@Paul King

The i7 4790k is getting a little bit outdated in general, and for the model size and complexity that you imply is starting to show its age. But again, may be there is something that is within your model that is causing this lag.

So I would advise:
1. Try to open your model on another PC with a higher core count such as AMD Ryzen 7 1700 and up.

2. Make a copy of your project, and start deleting things, such as layout, models, etc.

That way you would have a more precise idea on what is causing such lag.

PS:

• I didn't know if you're working on MAC or PC, but anyway, check the OS updates.
• If you want to buy new hardware, just wait 2-3 weeks, AMD will release the new CPU Ryzen and Radeon GPU, and the competition is getting hot, and that my push the prices down

[Paul King]

Thanks for the suggestions

Unfortunately I don't have access to a high CPU core machine to test with ArchiCAD - this is why I was hoping someone on here might have tried it.

Sadly I often don't have a lot of say over what is included or excluded from models I am assigned to work on as a freelancer (I am usually bought in when the model and documentation is already quite far along, but the practice recognizes they need help), so I can't always rely on good model optimisation techniques or structure to solve my problems.

Consistently it is editing and refreshing while trace reference is on (it needs to be on a lot, because a lot of what I end up doing is cleaning up discrepancies from one story to another) , or with sections and elevations in large complex models with lots of SEO and/or IFC models referenced in etc. or editing in 3D where elements I need to snap to are very large, or there needs to be a lot of potential snap candidates visible at same time etc that I notice lagging.

Am on PC, with always up to date OS

Yes, I am intrigued about the new AMD hardware coming up, and given shortages of Nvidia rx3080 GPU anyway, it may not be crazy to wait longer.

[Erwin Edel]

With trace and reference it always speeds things up if you trace from the Project Map rather than the View Map. Even if all the view settings are the same with your active view, I think. Certainly so if they are different.

I think SEO does benefit from raw CPU power, not sure if it is optimised for multiple cores. You can maybe see that with Task Manager in your current setup.

Sloppy SEO usage can really bring things to a crawl. Like selecting all the walls and using a roof as an operator, regardless if they walls need to be trimmed by that roof or not.

[Paul King]

Erwin wrote:
With trace and reference it always speeds things up if you trace from the Project Map rather than the View Map. Even if all the view settings are the same with your active view, I think. Certainly so if they are different.

I think SEO does benefit from raw CPU power, not sure if it is optimised for multiple cores. You can maybe see that with Task Manager in your current setup.

Sloppy SEO usage can really bring things to a crawl. Like selecting all the walls and using a roof as an operator, regardless if they walls need to be trimmed by that roof or not.

Thanks - some great tips there! Esp. using project map rather than view map for tracing - will give it a try.
Have not been able to discern the extent to which core share the load for any given task unfortunately.

You also just suggested a very useful wish for Graphisoft - filter seo selections for actual contact either before or after the operation (eg clean them up in the background), discarding non connected objects from recorded operator and target selection sets.

[Barry Kelly]

Paul,
Have you read this in the HelpCenter?


https://helpcenter.graphisoft.com/knowledgebase/25850/


My take on it is more cores are only good if you have sufficient RAM for each of those cores.
And there is a diminishing return when increasing the number of cores, but it seems like it is difficult to quantify it, otherwise I am sure they would have.


Barry.

[Paul King]

Thanks Barry - no I had not seen that article - very helpful.

Sound like the takeaway is yes ArchiCAD will use more than 4 cores if it can get them, even outside of cpu rendering, but for anything but rendering the marginal gains per core diminish a lot with every added core, such that moving from 4 to 8 is not going to have the impact of moving from 2 to 4.

Yes I see the article also seems to bear out what you are saying about having enough RAM per core.

Sounds to me like an upgrade from my current 4 core to an 8 core would be beneficial, as much as anything due to the increased single core performance more modern cpus have, and because 8 is the current default number of cores in a new workstation cpu anyway, but that any more cores than the current 'sweet spot' unlikely to pay for themselves given the steep price jump thereafter.

On that basis Intel is currently looking a bit better than AMD, based on single core performance of their respective 8 core cpus, at any given price level. May all change in a couple of months.

[Nader Belal]

@Paul King

You can even go with 12 cores with Ryzen ....

Anyway, if you're going to AMD, make sure that you check Ram specs, because the RAM velocity have a greater impact on AMD than on Intel.

[Lingwisyer]

Though if you start looking at memory speeds, you might realise that higher is not always faster due to latency increases...

[Paul King]

OK, I can now answer my own question.

I upgraded from 6 year old 4 cores i7490K to a fresh new box with 19 10900K (10 cores, fastest single core speed in retail market, at least this month), 32 GB DDR4 Ram, ITB NVIe SSD, RTX 3080 GPU etc.

Short answer is, yes there is definitely some improvement, but far less than hoped, much of the time. Certainly things are a bit more snappy in 3D, and when selecting an item from a complex scene etc, but subjectively I would rate the upgrade, between CPU anf GPU, as giving maybe a 30% 'experience'/frustration level improvement overall in my current large teamwork project .
Still quite slow working with trace reference on.

Am still also seeing about as much lag during any teamwork operation, despite the BIMserver just having been moved to a new dedicated machine + upgraded from HDD to SSD drives (it had been doubling as a file server apparently!) - and while far less crashing is now occurring now, speed not actually much changed during any teamwork operations.

I don't think this is a straightforward network issue - every time I click Send/Receive- there is a several minutes delay while a steady 0.1MB/s network activity is maintained before, during and after the send and receive, despite both BIM server PC and my PC being on 200MB/s fiber connection in same city. No surges to speak of in CPU or RAM utilization at my end during this process. Data is being exchanged at a trickle in other words, far from saturating network or system capacity.

Am guessing any bottleneck must be with the BIM server conducting some CPU intensive task at other end, maybe scanning database for differences and implications based on my 20 or 30 changes (which in themselves probably don't entail much data)?

This lag remains frustratingly true even late at night with no other users connected/potentially competing for server resources and bandwidth.

[Erwin Edel]

Only tip I can offer is a bit lame perhaps, but: don't use send and receive too often.

Sometimes re-uploading the teamwork file speeds things up and gets rid of the 'bloat' of it, I think.

It's been a while since I've done a teamwork project.

Thanks for the feedback though. Do you feel like schedules have gotten any faster? These are our biggest bottleneck at the moment and I think most of that is done on a single core (if I look at CPU activity).

[Paul King]

I could not say what impact the upgrade has had on scheduling - not a feature very extensively used in my part of the world (too much legal liability, depending what you are scheduling).

In theory even single core operations should show some improvement - maybe around 20-25% from i7 4790k to i9 10900K, though any multi core operations should benefit far more. I am not so far noticing many (or any) multi core operations in ArchiCAD, though I know they exist.

I would be happy to benchmark schedule generation time if you wan to send a test file.