tived wrote:

Bo, I know we only talk Pano work here, but the computers will obviously be used for other tasks as well, and though you are keeping your thesis to this topic, I suspect that both Hankkarl and I are in the back of our mind also thinking about the all the other things that this computer is meant for, though I think the pano work is by far the most intensive work, at least for me it is. It does help to have a fast computer when going through 2-3000 files from a wedding shoot :-), when you shave off a few sec here and a few sec there, it adds up when you have that many images to do something with. (no they are not getting all that many images, but there about to be some blinkers) :-)

Yes, there are other things I do, like my day job.  I also do a bit of video work with Premier.  My current machine can transcode a 30 minute DV video to H.264 in 30 minutes (the old machine took over eight hours).  PS is faster.  And I keep many windows open at the same time.

And there are the real time aspects of things.  I don't like to sit around and watch the task bar move.

Bo, the equation differs for us.  I can get Dell systems fairly cheaply.  And US salaries may be different than the salaries in Bulgaria.

Why do I buy Dell instead of home-brew?  Because I don't want to fiddle with incompatible boards, drivers, etc.  For example, I had a cheap firewire board in my old system. Never got it to work.  So I bought a Dell-recommended board, which worked like a champ.

Hi Folks - read your comments with much interest.
I have the opportunity to buy an 8 core MacPro from 2007 - with 2 4-core Xeons of 3GHz each, 10GB RAM, Nvidia GeForce 7300GT - for 1500.-€.
In my eyes the only relevant problem is that it uses fully-buffered RAM which is really expensive.

In my understanding related to APG this machine equipped with four internal big and fast SATA drives definitely i an advantage to my
2004 G5 using 2 IBM PowerProcessors . . . - but of course not as fast as the current version using hyperthreading of which i don´t know
how relevant it is related to APG and APT.

So how important is it to use the very newest processor technology in spite of APG: What counts most?
What´s more important: drive speed? drive size? Processor speed? Hyperthreading? RAM size? RAM bandwith?
RAM clock?

As o learnd so far the current generation of MAcPro is around 15-20% faster than the one i´m talking about.
1) that´s not really impressive and 2) i doubt they run the test using APG . . .

The upcoming generation of Mac Pro will be faster of course - but how much would i profit of the newest
technology related to APG/APT?

best, Klaus

Last edited by klausesser (2010-07-22 11:41:37)

I think i would take that MacPro but since i dont like Mac its nothing for me

Mmmh i thought that the Open-GL and GPU abillities of the Quadro Series would improve the APG Performance a lot, but when "even" a GTX is faster then i should try more in this direction becuase even a mid-range Quadro is more expensive then to top GTX in SLI.

In fact i supposed the SSDs not for the OS-Partition but as a big TEMP Folder

I would be very nice if someone from Autopano / Kolor might help us out here especially with the GPU/CPU thing, because when GPU is just for preview it is not so essential then the actual rendering(just my opinion), so it would be better to annihilate "Waiting-Time" by a big bunch of Brute Force in the CPU.

I think i will go in a few month with the D20, because the SystemBoard allows two CPUs and up to 96GB of Ram whether i will never reach them:lol:

I think we should fall on Kolors nerves to focus on one System eg. GPU/CUDA for everything, so you dont need a CPU which clears your Bank-Account and a Graficadapter that clears your Wifes Bank-Account and some SSDs paid by the savings of your child

No just kidding, but i think it could be realy usefull to let things like detection and rendering will use the GPU as well, but maybe thats already in development and we are not patient enough

I just phoned with the "Tesla" Guy, who talked to nVidias Support, they said that a Tesla Card will not benefit in ANY Way to APG, because TESLAs are just optimized and usefull for CAD, SW, Maya renderings and programs which "want" to be improved by a unique driver for the progam and the Device itself and to do so the program has to have massiv CUDA implementation

At least thats the answer the Support-Guys gave my boss this morning.

Hi Klaus, yes, the macpro sounds like a good deal, though only you can tell if it is a good deal for you :-)

Interesting about not being able to us Tesla, but not unexpected.

I think, regardless of the new engine, good GPU, plenty of ram, and very fast hard drives, will not go astray. Justifying it, is a different matter.

One thing I have gotten from this, is to revise my workflow and utilise my resources better. With my next new machine, I will have three machines in the office, hopefully the new one will be quiet. The devise a way to spread the work across the machines to make room for more hands on jobs on the main computer.

Bo, if we somehow can settle on a Benchmark test, or tests. that would be good. And we would also have to devise one before the new engine comes out as we need point of reference. I am open to suggestions, but I do have one question, how do we make this uniform, so that everyone does the same test with the same restrains.

Suggestions : x number of images, with xy dimensions (eg.2000px by 3000px), if we could create a synthetic one, then it would not be depending on whether or not someone had images with detail or not for the engine to find control points.

Now, I unfortunately know nothing about writing scripts so I am not the most helpful person for that job, only did first year at comp sci (twice)

No need to go overboard with the test, max a couple of hundred images would be max or even one hundred :-) to increase the memory, you can just increase the number of pixels, that should push both memory and cpu.

what would be really good, if, it could be broken down in the various stages, so you could see where you needed to add more fuel. but that is far beyond my own capabilities.

Just a thought

Henrik

Hi Bo,

thats not a bad effort, but if you look careful, the top systems are all "supercomputers :-)" or they are single processors with special rendering software, by the professor Mr Panotool himself :-)
Now add the special sauce to the SC, then they will fly. :-) obviously they cost 4-5 times more then the hightest single processor :-)

I will see if I can get some time to run this over the weekend.

Your idea of the test is sound I think, but I am worried that people are not likely to want to download 100's of megabytes for a test, which is where a synthetic test would be more suited, but again, I can't make that so I better keep my mouth shut, so we can get it done :-)

thanks

Henrik

The main problem with that set/test is the size of a single image! Those are 300+ images, 200KB each How about using the same count, but full size TIFF or PNG files, not some lowres 6MP sources?

Other problems - it's not clear how the project was rendered in Autopano. Was is using Smartblend? I guess not - if it was, the HDD influence would have been much bigger! Also, the project comes with render size set to 98%, why is that?! I render at 100%. Also, RAM has almost no effect in such small pano sets. So why bother at all What's more, we need to test Autopano 2.5, not PTGUI or APP 2.0.x.

Anyway, what I hope for is to get those same people to take a future Autopano-based test, even with APP 2.5 Trial installed, and then compare the results!

Mehlsack, I'd really like some "official" test sets + project files, that come out of Kolor. This whole project will benefit the company as well and will help potential users evaluate their configuration.

Clearly I never made it past Computer Science 101, but when developing software on a commercial level, don't you use set some sort of test to set minimum requirements. Also to see if you have maximized your coding.

This should be of great interest to Kolor and beneficial for both them and us, users, unless there is some information between the competing Pano Software companies that prevents this from being published. Is any at Kolor listening in here???? It would be really good if they didn't favor one Processing family over another (AMD vs INTEL) but kept both equally well supported if possible and relevant.

Ok, in order not to send around a huge set, but somehow, make something that is portable and still accurate. Could have a set of JPEG's at a certain size, (we are not looking of image quality here, so we could run an action in Photoshop or similar that would up-size the images to X-factor bigger, if this would make a difference and perhaps save as TIFF's

Question is, does APP look at byte size or the pixel size of an image (eg, would it matter if it was JPEG or TIFF??).

Also if the set of images is known, and the program settings are set in *.pano file would it not then be possible to be repeatable, like the Giga-pixel Panorama Speed-test seems to be.

We could outline, the settings, that one would have to apply to. such as which blender etc should be used.

anyway, at the end of the day if the set is X megabytes size, then that is just it, and we can just get on with it. Personally I don't have a problem with it, it could even be 1Gb but then it would be great if it was download managed so one could either resume if the link got broken. (I had to download a set of videos recently that I had bought, it was originally 5GB, and I ended up spending over 15Gb of downloading it, because their end did not support download manager)

Ok,so does anyone know how to implement a timer?

do we know anything that can monitor resources and the various platforms?
On XP and perhaps on Vista and 7, there should be administrative tools/performance where you can set the various things to be graphed. but probably not precise enough.

...I think I am rambling here, but I hope you get the idea

Henrik

Last edited by tived (2010-07-23 02:09:11)

Hi everyone,

just read this from Alexander from June this year, regarding version 2.5

Quote:
"For v2.5 series, the answer will change :
- The old multiband is not available anymore. This was a fast but limited multiband algorithm.
You now have access to a full multiband algorithm which have the same demand as smartblend.
- Smartblend engine has been renamed : it's now only a part of the new global rendering engine.
Yes, it will use disk if needed. But if everything fit in memory, you don't need fast drive. It will depends on
the targeted size. My guess is that for gigapixels images, you will first need fast drive because it won't fit in memory."
End Quote

I know Bo, will tell me I am speculating again, and that I have to wait till the release :-) <grin> but if we are making big pano's then we can expect to have to have a faster Hard drive sub-system
thats what I am reading from this, and a bit extra RAM won't hurt either.

Bo, et al.... have any of you given a test more thought? as in what its going to consist of, and how we will confirm the process?

No rest for the wicket here!!! :-)

Henrik

[bo] wrote:

This quote is about a month older than the 2.5a release, so it's actually less relevant than the sticky topic in 2.5 forum. Anyway, I wouldn't disregard any of the system's components - you are absolutely right to think that everything should contribute to the performance. My main concern is *how much* will each component contribute, so I can make an informed decision when I decided on my next PC configurations (to replace older dual core, x86 systems).

As for the test - I've written an email to Alexandre; I guess they're pretty busy now, so I'd rather see 2.5 finalized sooner and *then* worry about testing, instead of the opposite. My current proposal (which, in some time in the future will have its own topic for general discussion) is this:
- handheld set, 15-30 images @ 24mm or similar, parallax issues, moving people or clouds, etc; the general "quick snap" pano
- panohead set, 200mm+, partial pano of 30-50 images, probably planar projection
- panohead set, fisheye, full 360 deg. pano, 6+2 or some other common setup, probably bracketed
- gigapano set, 100 images
- gigapano set, 300 or 500+ images
- potential FUSION set, exploring HDR fusion and/or focus fusion (but we need to know if this is a separate step in the rendering that occurs only on specific settings, or it is always invoked)

All images should be taken with 5Dmk2 or an equivalent (and probably distributed as JPGs). Having in mind that a even a current P&S camera has 12MP+, it's best to test with high-res images. Also, we're talking about Autopano PRO here, after all

The tests should not be synthetic and should depict real-life usage. Right now, the only thing I can think of testing via the Autopano GUI is the render speed. Queue all the projects, wait for the render to end, report system spec and times.

However, knowing that Kolor certainly have command-line driven Autopano engine, it's feasible to think about some custom GUI/cmdline tool that'll be fed with images and report detection time, optimization time and render time. Personally, if I was on the Kolor team, I wouldn't invest much time in this. After all what need does that tool satisfy? A few enthusiast wanting to measure their... machines' performance? So even the rendering test, which will be easy to organize and conduct among users, without any endorsement or participation of Team Kolor, seems quite enough for me.

Hi Bo,

The testing specs sounds good.

Hmm, If you were on the Kolor team you wouldn't invest much time on this, maybe you personally can't see the benefit of this, and that is fine with me. However, I do believe there is something to be gained from this both for Kolor and for us end-users. There are people here ranging from the odd pano shooters, to people who makes a living from it, and in-between there is the people who does similar work, and find that offering pano's to their clients as a value added product, the later could be the two of us, or at least I fit into this category.

Like so many other topics here on this and other forums is based on the content everyone brings on to the table, some is more relevant then other things, its often from the hard work many people put in, mostly off their own back, such as the Merlin/pano head that makes it interesting to come to a certain forum. Its also what makes this forum what it is. This topic here may be considered noise, it may be what people in the future will use as a reference for when they buy or build their next computer. Based on our experiences, maybe its willy wagging to some, to others its a necessity, to have hardware that performs to certain expectations, and will pay the price.

In this thread you, yourself have contributed greatly, you have a different view on things then I do, and that is what makes this an interesting place, that we can learn and contribute, and exchange our different opinions. There are things here that you and others have mentioned that I had not thought off, and hopefully there is something in it here for you.

All the best

Henrik

About benchmarking computer. We care about that. That part of our release check list : do our big panorama database and bench it ( in timing and in quality : meaning, 96% of panoramas were perfect with default setting detection / rendering ).

One note : with 2.5 beta 1, you'll get a better benchmark output in the batch render. Everything needed for bench is now displayed ( detection, optimization, color correction, previewing, rendering ). With the right global settings in the automate ( render checkbox checked ), everything that is detected will directly be rendered too. This will make Autopano perfect tool for computer benchmarking.

I agree with [bo], we need to build a repository of typical stitching jobs for benchmarking. Once that done, with a simple browse folder on that database and the right settings, we could bench really nicely any computer.

It's already available on the 2.5 alpha. When a rendering is finished, click on it and nearby, you can see the resume of timing ( that's in the batch renderer).
What I said is that it has been improved for 2.5 beta.

Hi Bo,

its a guide, for the average user. Editing Pano's I would not fit in the average user category :-) but I could be wrong, or have very different expectations.

I am contemplating, I have placed to order at my friends store.

for
Also I am walking down a path, that I normally wouldn't the over-clocking path and hoping I can get 3.8-4Ghz on air :-)

2x Intel X5650 cpus
EVGA SR-2 w555 mainboard http://www.evga.com/products/moreInfo.a … y&sw=5   :-) 
RAM: MP24GX3M6A1600C9 - DOMINATOR, 24GB (6x4GB), 1600MHz, 9-9-9-24 with Airflow Fan or Kingston HyperX if I can get the 2000Mhz that would be better. (6x4Gb will leave me room for another 24GB later)
2x  SSD C300 Crucial sata-3 ( 1 for boot and one for scratch or both for boot, not sure yet)
4-6 x WD Black Caviar 1TB
GPU; Starting with one, probably a nVidia 460 1GB but I will wait and see till everything comes in
Case: here lies the problem as this mainboard, is 13x15in HTPX size xtra large. Possible Lian Li V2120X when it comes out in October
Some optical drive, not fussed which.
CPU fans Dual Noctua NH-D14s
PSU: Antec 1200w

this would give me 12 core and 24 threads with HT at 2.66Ghz at stock speed or 3.8-4Ghz overclocked (yes I know the pitfalls, been there before in my youth :-) many moons ago)

Henrik