Forum

Author Topic: Is using a multi cpu workstation worth it ?  (Read 3609 times)

cyrilp

  • Jr. Member
  • **
  • Posts: 86
    • View Profile
Is using a multi cpu workstation worth it ?
« on: December 03, 2020, 07:55:46 PM »
Hi,

i currently have a 9900k / 3080gtx / 128go ram computer.

without budget limitation what would be the best upgrade ?

would a workstation with multi cpu or a Threadripper PRO be a lot faster ?

(using windows 10 pro)

cyrilp

  • Jr. Member
  • **
  • Posts: 86
    • View Profile
Re: Is using a multi cpu workstation worth it ?
« Reply #1 on: December 06, 2020, 03:57:50 PM »
Hi, nobody to help me on this ?

Eric Baird

  • Newbie
  • *
  • Posts: 25
  • Quadro K5000 + Quadro K620 / Xeon Hex / 64M RAM
    • View Profile
    • The Brighton Toy and Model Index
Re: Is using a multi cpu workstation worth it ?
« Reply #2 on: December 07, 2020, 02:03:35 AM »
Hi cyrilp! I guess it depends on where your current processing bottlenecks are, and that's something that other people aren't likely to know, because it depends on the sorts of projects you're doing, your personal workflow, and what your priorities are for settings.

Some people will spend a lot of time revisiting and redoing a stage that is mostly GPU-intensive ... others may spend a lot of time tweaking a stage that's CPU-intensive. I guess it depends on which sort of calculation stage you find most annoying!  :)

andyroo

  • Sr. Member
  • ****
  • Posts: 446
    • View Profile
Re: Is using a multi cpu workstation worth it ?
« Reply #3 on: December 07, 2020, 10:11:06 AM »
Hi cyrilp,

Eric Baird is exactly right. I've been building workstations for PhotoScan/Metashape for several years, and I often refer to the Puget System benchmarks by William George to look at how different CPUs and GPUs perform - he covers most of the options. I recently posted some info in this thread that might be useful, but just to emphasize what Eric said - it really depends on your workflow and how you use Metashape. Some steps are CPU intensive and some are GPU-intensive. Some benefit from network processing (having multiple CPU/GPU nodes process chunks in parallel), and some need to run on a single machine no matter how many you have.

I have found that in general once you get up to ~6-8 cores, faster is better than more cores, but again, it depends on what steps. I am a fan of keeping my core count high while trying not to drop too much in CPU frequency, but some of that is because I have other pipelines that benefit from the high core count. William repeatedly says in the various benchmarks he has run that the core count benefit drops off after 6-8 cores too. So if you do a multiple-CPU setup keep that in mind and pick CPUs with a high speed clock. I was building multi-CPU machines earlier, but I've shifted to faster machines, while trying to maximize core count (currently our fastest single machine is a Threadripper 3960x with 2 RTX 2080 Super GPUs), and I'm moving into networking those machines now.

Personally I think that multiple networked fast machines with a few GPUs each are probably the best investment, and depending on your workflow, how much of it uses CPU-only steps, and how big you want to go, you might even want to have an overclocked Ryzen or two with no GPU or a lowly RTX 2070 Super, since you can fairly quickly eat up the price of a license with threadrippers and RTX 3080s. You say "without budget limit what would be the best upgrade?" ... well... a fast network of a lot of fast computers with GPUs and PCIe-connected NVMe storage I suppose...

Back here in budget-limited land, I'm frequently trying to balance performance gain with price increase - and the Puget Systems benchmarks are really good at helping me zero in on the sweet spot for price with GPUs and CPUs. For the work we do (Align/Dense Cloud/DEM/ortho) - the best performing machine is a Threadripper 3960x with 2 GPUs and 256GB RAM, but I'm about to throw in a third GPU.

For your machine, depending on the mobo and your workload, I'd look at adding another GPU before I do anything else if you're GPU-limited, otherwise, move your GPU over to an AMD 5000 series and put in some 3600MHz or faster RAM, or better yet, keep that one and build/buy another with a second Metashape license and network them!

Andy