New PC: biggest bang for the buck?

[darkownt]

Hey Nocturnal:

I know it sounds ridiculous given how powerful the i7 920 is, BUT if you ever feel the need to upgrade, the best thing you can do next to cheaply speed up your system is cooling.  Go from the mini case to a large performance based case with plently of airflow, maybe even get a 3rd party CPU heatsink/fan and you can reliably overclock your i7 from 2.66 GHz upwards of 3.

I have the same CPU as you running on an Asus Mobo, with a large coolermaster case and ... Thermaltake H/s fan on the CPU, and I am getting 3.5 GHz from it, completely stable.  Even with my vid cards o/c'd (and yes the GTX 260's ARE hot).  Remember, when the Intel guys test the CPUs they give a wide margin so that the CPU is stable at the rated GHz in the worst computer enclosure/case, hottest environment, using the stock H/S and fan...

Congrats on the i7 920, but dont be jealous of ppl who spent $1000 dollars more on a 3.5Ghz chip... you probably have one sitting in your computer right now, and you could unleash it someday for less than $200 ! 

cheers
Colin

[Nocturnal]

Hi Colin,

yes I know the i7 is easily overclockable. I just haven't bothered with it yet My case isn't that small and it has a huge fan on top so I think I should be able to OC it quite nicely even with current coooling. Right now I appreciate my computer being quite fast and still very quiet. And the thought that if I get bored with it a simple overclock is all it'll take to speed it up.

[Yuriy Toropin]

Let me share my expirience in fight for speed of PI.

As a test bench I used StarAlignment process, aligning 24 8 MP shots from QSI583 (default parameters of process were used)

My initial configuration was
Intel Core 2 Duo E8400 (@3GHz) +4Gb DDR2 (@1066MHz) + GeForce 9600GT (1GB) + 1 TB HDD + Asus P5E Deluxe MB, under Win 7 Ultimate 32-bit.
Initially, time required for alignment with mentioned configuration was 176.806 sec

First round of upgrade was relatively simple, I've added another 4 Gb of DDR2 RAM (@1066MHz, total RAM is 8Gb now) and changed to 64bit version of Win 7 Ultimate...
Suprisingly, alignment time went down to 147.109 sec (woo-hoo, -17%!!! or 1.2 times, not bad at all!!!)
Well, this is simply effect of going 64bit (64bit PI was used) having enough RAM 'cause 32bit Win can't address all that 8Gb been limited to something below 4GB naturally by address space available in 32bit version.

Second round of upgrade was in replacement of Duo to Quad, to Q9650 (@ 3GHz), the fastest Socket 775 Intel processor available on the market now.
Alignment time went down to 93.854 sec (Ok, I realize, that couple of last digits are almost meaningless , I've performed 5 runs and present average in the post for all sets ).
This is another -36% or 1.57 times improvement from previous configuration and almost -47% or 1.88 times improvement from my initial PI performance.

So - go 64bit, you will cut ~17% of waiting time, doubling processor cores will give you another 36% benefit. With this I almost decreased my "waiting" time by a factor of 2.

For sure, it can depend on process used, size of image, etc etc.

The next big boost will be in multiprocessor system or - in well-done CUDA support, NVidia Fermi is just in month or so to wait... Juan?

[vicent_peris]

doubling processor cores will give you another 36% benefit.

This is a quite small performance improvement... surely SA is not the right process to analyze paralelization performance. Some other processes will cut down the waiting time to 1/2 using double core number.

I think we could define a standard test bench for PI. We can analize some critical aspects under various machines if we define good standard proceedings and upload use standard images.

What do you think?


Vicent.

[Yuriy Toropin]

Vicent, this is a great idea!

Some script that will be a good representation of "natural processing workflow", starting from alignment to sigma integration, deconvolution, ACDNR and, probably, some other heavy computation things will define good "synthetic" test for hardware performance.

Standard set of data to process should be also provided.

Unfortunately, I'm not the "script guy", but at least I can supply data for "standard performance testing set".

Meantime, I will continue my "home" testing with any new release of PI or hardware change I will make
I'm looking for SSD as the next possible upgrade option 'cause right now storage access speed is my weakest point...

[vicent_peris]

Hi,

tomorrow I will prepare a standard dataset consisting in 13, two megapixel greyscale images, and will do a standard processing. Juan will make a script from my processing icons. Then we will release the script and the dataset as a standard benchmark. The 1.0 version of the benchmark will simply measure the "performance index" of your system (AKA "PI" ). I think this can be a significative number that will tell you how will be your "PixInsight experience". :-P The higher, the better. The unity reference will be Juan's Linux workstation.


Vicent.

[Niall Saunders]

Quite interesting Vincent - I will be interested to see how my PC fares on your standard test 'right now', when my system has been running for several weeks, with a whole load of other processes also running, compared with a clean re-boot, running PI only.

I think that the 'fully-loaded' resuslts will be of greatest interest, because that is how my machine 'normally works'.

It will also be interesting to see how the last of the Core-Quad processors fare against the newer i7 series - especially if the OS is, more or less 'the same', and the 'typical' memory quota is also 'the same' (e.g. two systems both running Vista64 Home, with similarly clock-speed processors, Core-Quad vs. i7, 8Mb RAM vs 6Mb RAM - which would be similar, $for$ )

At least it lets a user determine if a given 'upgrade' can be justified, in terms of benefit versus investment.

Cheers,

[georg.viehoever]

Hi,

... The 1.0 version of the benchmark will simply measure the "performance index" of your system (AKA "PI" ). I think this can be a significative number that will tell you how will be your "PixInsight experience". :-P The higher, the better. The unity reference will be Juan's Linux workstation.

any news on the PI benchmark? I would very much like to see a measure of how fast a specific hardware can be (with all the problems that benchmarks tend to have...)

Georg

[vicent_peris]

Yes, news:

I have designed a processing chain with 20+ icons. The data set is 6x10 min R, 4x15 min V and 3x20 min B of M74 with CAHA's 1.23m scope. You will download the data set (about 60 MB) and the script Juan is writing with these processes. Processing starts with StarAlignment and ends by resizing the image for the web.  


Vicent.

[mmirot]

That will be real cool. 

Max

[georg.viehoever]

Hi,
I can't wait to give it a try... 8)
Georg

[bitli]

Hi,
I plan to buy a new machine, and I hesitate between a Core i5 750 (8GB, upgradeable to 16 in the future) and a Core I7 930 (12 GB, upgradeable to 24 in the future), on Windows 7 64 bits (for professional reasons). I do plan to use a very low end graphic card (I do not use 3D neither games), and I do not think that PI will be impacted, right?. I start with a single 7200 HD to stay within budget (more can be added later or an SSD added in the future).

Do some of you have some feedback on whether the I7 based system is worth the additional cost (about 2 license of PI :-).
--
bitli

[georg.viehoever]

Bitli,

I would go for the i7 930 because it has a 3 memory channels (compared to 2 with the lower end i7 and i5). Image processing tends to be a memory intensive tasks (both in terms of memory bandwidth and memory size), so 3 channels are often better than 2. Having said that: i5 systems are usually cheaper for comparable performance for the usual (not so memory intensive) benchmarks.

Save the money for an additional HD or SSD until you need it/have the money. They are easily added to an existing system - changing the CPU on the other hand can be a lot of headache. And they get cheaper every day...

Georg

[bitli]

Seems reasonable, thanks. Now I have to find a 'decent' screen,,,
I will run the PI benchmark when I have both the benchmark and the machine and keep you informed.
--
bitli

[Nocturnal]

I would think a single process icon (distributed as a PSM) would serve as starting point for a benchmark. Why would it need to be scripted? It already shows the time it takes to execute in the process window:

ProcessContainer: Processing view: LDN1622_4
Crop
Crop: -2, -2, -2, -2: 100%
HistogramTransformation
Processing channel #0: Histogram transformation: 100%
Processing channel #1: Histogram transformation: 100%
Processing channel #2: Histogram transformation: 100%
0.373 s

It would be nice if each individual step was timed so you can compare where the time spent but that's not really necessary.