Hardware upgrade to cut down reconstruction time

February 20, 2017 01:55

Hi,

We are looking for a hardware upgrade to cut down computation time to about half compare to our
current one and have some questions.
Here is our current situation.

# of Images: 300 ea
Duration 2 ~ 2.5 hours (Mostly consumed at High Detail Reconstruction)
CPU: Intel® Core™ i7-6800K Processor
*https://ark.intel.com/products/94189/Intel-Core-i7-6800K-Processor-15M-Cache-up-to-3_60-GHz
GPU: Two Geforce GTX 1070 connected through SLI
*https://www.nvidia.com/en-us/geforce/products/10series/geforce-gtx-1070/
RAM: 32GB

1] # of CPU cores vs. CPU clock speed? (Which one is more effective?)
2] GPU or CPU upgrade? (Which one is more effective?)
3] Recommended high-performance spec settings to cut down high detail reconstruction time to 0.5~1 hour(Considering full system upgrade)
*Maybe you can just give out your most powerful setting
4] Expected computation time savings with upgraded system

Best Regards,
Justin

Comments

4 comments

chris

February 20, 2017 02:45
are you running out of ram at the moment?

running everything off ssds?
0

Comment actions Permalink
Justin Jang

February 20, 2017 06:28
Hi,

Yes, We are running everything on SSD.
No, we never had trouble with RAM ..
0

Comment actions Permalink
chris

February 20, 2017 07:12
its probably also worth figuring out what takes longer gpu or cpu.

I'm also not sure how well rc handles multiple numa nodes. but think it suffers a bit so your probably best to stay single cpu rather than dual xeon. esp if you don't need the extra ram.

adding anymore video's cards will only add a little bit of extra speed.

for cpu's you could go a 6900k or 6950k. (im running at 6900k @ 4ghz).

or you could go for a single xeon like a e5 2687w v4. just keep in mind keep below 16 cores unless you want to disable hyper threading.

probably by far the cheapest thing would be to get a few machines, at similar spec to what you have.

I think going for 1 very powerful pc, makes sense if you need lots of ram, or are getting cli version.
0

Comment actions Permalink
Castlenock

February 20, 2017 18:56
Hey a couple of observations which are based partly from experience, partly from research, and a good bit of guessing so don't take this as gospel. Also, this is a bit eclectic, major headache and running out the door...

- I think you'll be hard pressed to get 1/2 the speed of your current set up: could be wrong, please someone chime in if I am. CR has suggested that node based clusters are being actively worked on, so I'm getting ready for that.
- I upgraded to 64GB from 32GB of RAM with a similar set-up and was surprised to see that RC happily used it - don't know how well that equated to performance, but it wasn't eating up more than 55% when I use using 32GB either so I was surprised it used it.
- Wishgranter just posted a good article from Tweaktown (there is another on Tom's Hardware I think) about building some really good multi-CPU systems on the cheap given a quirky dip in the CPU market. Can't find it at the moment but chase that down...
- From everything I have seen, I think the best bang for the buck right now is a 6900 overclock it a bit if you want, with 64GB+ of ram, 2 GPU cards, some PCIe-like (nvme) memory speeds, and some crazy good coolers if you want to overclock for 10-20% gains (good luck on that though). You'll get faster speeds with more CPUs if you start to switch over to Xeons, but I don't think by much at all to be honest. Read puget system's multi-CPU/GPU analysis on **************** totally different software, but it's efficiencies for multi-GPU and CPUs are crazy good, RC may be the same, but the multi cpu/gpu return ratio seems to putter out pretty quickly, so going for 32 CPUs isn't necessarily the best decision if price is a major factor.
0

Comment actions Permalink

Please sign in to leave a comment.

Comments

Didn't find what you were looking for?