Clone #1: PC for scientific computing

[mornington]

**No0B**

My intention is to build a sequence of rigs, essentially clones of each other to put together in a cluster. I want to have an easy way to learn distributed computing, MPI etc. None of this is directly relevant for this post except that I would like to control costs.

So, here is what I have in mind:

1) CPU: AMD Phenom II X6 1090T Black;
http://amzn.com/B003FVI2KQ
Yep, I am cheaping it on the CPU.
Price: $199.99

2) MoBo: Gigabyte AMD 890GX;
http://amzn.com/B003A764XM
For some reason, I prefer this to the GA-870A which "mrSHEiK124" recommends in his mid-range mobo review.
Price: $134.99

3) Memory: Corsair vengeance;
http://amzn.com/B004CRSM4I
Saw these on LinusTechTips. They are 1.5V, hence good?
Price: $119.99

4) GPU: Nvidia GTS 450 from PNY
http://amzn.com/B0041HN6P4
Basically, have heard a lot about GPU computing and what support Nvidia provides for GPU computing and would like to have a solid mid-range GPU to try out some of that stuff on.
Price: $122.99

5) Cooling: Cooler Master RR-B10-212P-G1 Hyper 212 Plus
http://amzn.com/B002G1YPH0
This seems to be everyone's favourite mid-range cooling solution on these boards.
Price: $28.43

6) PSU: Thermaltake TR2 W0388RU 600W;
http://amzn.com/B002RWJGHM
Seems good enough for my purposes. No idea how good or bad this is.
Price: $49.99

7) VDU: ViewSonic VA2323wm 23-Inch 16:9 1080p;
http://amzn.com/B002R0JJYE
Seems cheap and reliable.
Price: $176.99

8) HDD: Seagate Barracuda 7200 1 TB;
http://amzn.com/B00272NHOK
Price: $59.99

9) Case: Cooler Master Elite 310 ATX
http://amzn.com/B0028ACYMC
This case is destined to remain open most of the time anyway - don't want to spend too much on it.
Price: $36.70

Throwing in an optical drive brings in the total to a nice $950 - which is more or less how much I would like to spend.

Questions:
++++++++

1) Suppose I wanted to stick in more GPUs at a later date, would my mobo allow me to do this (using 3x crossfire/SLI - I have heard of this stuff, but have no idea what any of this means)? Presumably this involves having enough PCIe slots or something like that.

2) Is the RAM good for the motherboard? I heard some stuff about some mobos not being able to use memory beyond 4gb...

3) Is the case big enough to fit everything in comfortably?

ALL suggestions welcome!

Thanks!!

 

[oynaz]

A bit of overkill on the PSU, you might want to find a quality 350-400 watter instead.

 

[mornington]

Thanks oynaz. I was thinking about it, and the Corsair CX430 (430w) costs only like $5 less, so in case I get another GPU, do you think it is a good idea to stick to a bigger PSU?

Also, any comments on the memory and the motherboard will be highly appreciated.

 

[mnewsham]

Unless you are wanting to play with more cores i would get an x4 or wait for bulldozer. RAM is overkill no need to get that ram, get 1333mhz CL9 1.5v RAM should be mayb 70-80 bucks.

 

[mornington]

mnewsham, the six cores is something I agonised about quite a bit. I know not a lot of software explicitly gets written for six cores, but I would like to experiment with my own hyperthreaded/parallel code.
I hadn't been following bulldozer development closely but it seems slated for 2012, so a long wait.
Also, the x4 970 black is only $20 less... Any more comments on this are welcome...

Could you possibly give me the brand/link for an 8gb RAM kit that you would recommend?

 

[mnewsham]

BD is slated for a may/june release at this point. and here is some ram i would get two of these. They are out of stock atm but the ETA is the 28th if you cant wait that long then these would work too.

 

[mornington]

Thanks for the RAM links, the GSkill looks good. I am giving some serious thought to the x4 965 black...

Hopefully, BD is an AM3, so I can just drop it in later.

 

[mnewsham]

Thanks for the RAM links, the GSkill looks good. I am giving some serious thought to the x4 965 black...

Hopefully, BD is an AM3, so I can just drop it in later.

AMD said BD would NOT be am3 compatible so you sadly cant just drop it in.

 

[KGB]

If you want to use these for developing distributed computing apps/etc., hang out in the DC forum here. Those guys do wonders in building low-cost 'herds' of munchers.

 

[mornington]

Cool, thanks, that was going to be my next stop!

 

[fffblackmage]

Thanks oynaz. I was thinking about it, and the Corsair CX430 (430w) costs only like $5 less, so in case I get another GPU, do you think it is a good idea to stick to a bigger PSU?

It's not always about the watts. You need to consider the quality of the unit itself. The crappiest PSUs (and there are a lot of them), don't even output their rated power, and if they do, the voltage regulation and ripple may go out of spec (very bad!).

If you really want to add a second GPU in there, get a quality 600W PSU instead. Remember, you get what you pay for.

 

[mornington]

It's not always about the watts. You need to consider the quality of the unit itself. The crappiest PSUs (and there are a lot of them), don't even output their rated power, and if they do, the voltage regulation and ripple may go out of spec (very bad!).

If you really want to add a second GPU in there, get a quality 600W PSU instead. Remember, you get what you pay for.

Thank you, that is a good point. I can't always tell between brands. Actually, I don't think a second GPU will happen anytime soon (I'd prefer to get the second rig going), so I think I will switch to the Corsair CX430.

 

[jchu14]

*take the following with some grains of salt. I am just a HPC user and do some basic code developing. I don't know too much about the hardware side. The machines just work (usually) when I log on remotely.*

The overall specs looks okay. Though the PSU is not from one of the more trusted brands (Corsair, Seasonic, Antec). The RAM is pricey and overkill. Get cheaper/basic RAM rather than enthusiasts RAM.

If you're just learning/developing, the speed of RAM and CPU won't be that important, and if this computer is doing production runs, then you wouldn't want to be overclocking anyways.

Does that motherboard play nice with Linux? You may want to learn the ropes in a linux environment since they're more popular in the HPC community.

As for your questions
1) i don't believe so, at least you can't in gaming SLI sense unless you have an AM3 board with nVidia chipset. I don't know if you can use it to do computing, but I doubt it. You can go with an intel setup for better SLI support.

2) I haven't heard of any modern boards having issue with more than 4gb of RAM. It will work fine (just about any ddr3 modules from a reputable manufacturer will be fine).

3) Should be.

You can find some lecture notes on basic parallel computing here made available by TACC (Texas Advanced Computing Center).

 

[fffblackmage]

Do you have to buy from Amazon?

If you don't mind buying from newegg:

$60 ($44 after promo core and MIR) - CORSAIR Builder Series CMPSU-600CX
http://www.newegg.com/Product/Produc...82E16817139019

$40 - Antec NEO ECO 520C
http://www.newegg.com/Product/Produc...82E16817371030

Corsair Builder are not top of the line or anything, but at least they're far from being crappy PSUs.
The Antec Neo Eco are based off Seasonic S12II, so they're solid PSUs for the price.

 

[mornington]

*take the following with some grains of salt. I am just a HPC user and do some basic code developing. I don't know too much about the hardware side. The machines just work (usually) when I log on remotely.*

The overall specs looks okay. Though the PSU is not from one of the more trusted brands (Corsair, Seasonic, Antec). The RAM is pricey and overkill. Get cheaper/basic RAM rather than enthusiasts RAM.

If you're just learning/developing, the speed of RAM and CPU won't be that important, and if this computer is doing production runs, then you wouldn't want to be overclocking anyways.

Does that motherboard play nice with Linux? You may want to learn the ropes in a linux environment since they're more popular in the HPC community.

As for your questions
1) i don't believe so, at least you can't in gaming SLI sense unless you have an AM3 board with nVidia chipset. I don't know if you can use it to do computing, but I doubt it. You can go with an intel setup for better SLI support.

2) I haven't heard of any modern boards having issue with more than 4gb of RAM. It will work fine (just about any ddr3 modules from a reputable manufacturer will be fine).

3) Should be.

You can find some lecture notes on basic parallel computing here made available by TACC (Texas Advanced Computing Center).

As you have cottoned on fairly well, I am quite new to this. The point about RAM was made by mnewsham as well, and I have switched that.

I have never developed in a Linux environment, and intend to for this right from the start here. So I anticipate a steep learning curve (but am looking forward to that). However, I have no idea if the MoBo plays nice with Linux. Any ideas?
I will either end up using FermiLab's Scientific Linux, or may chicken out and end up using Ubuntu. In either case, I do not know how they interact with the MoBo. Any comments are welcome.

Re: GPU computing, Matlab under Windows allows using GPUs for computing fairly painlessly. I was under the (possibly naive) impression that it would be possible to do CUDA programming if I had an Nvidia GPU under any OS. Or were you specifically talking about two GPUs in SLI configuration? That I'd have to use them as independent GPUs?

Thanks, this info is great!

 

[mornington]

Do you have to buy from Amazon?

If you don't mind buying from newegg:

$60 ($44 after promo core and MIR) - CORSAIR Builder Series CMPSU-600CX
http://www.newegg.com/Product/Produc...82E16817139019

$40 - Antec NEO ECO 520C
http://www.newegg.com/Product/Produc...82E16817371030

Corsair Builder are not top of the line or anything, but at least they're far from being crappy PSUs.
The Antec Neo Eco are based off Seasonic S12II, so they're solid PSUs for the price.

Antec Neo it is. Thanks!

 

[mfenn]

As you have cottoned on fairly well, I am quite new to this. The point about RAM was made by mnewsham as well, and I have switched that.

I have never developed in a Linux environment, and intend to for this right from the start here. So I anticipate a steep learning curve (but am looking forward to that). However, I have no idea if the MoBo plays nice with Linux. Any ideas?
I will either end up using FermiLab's Scientific Linux, or may chicken out and end up using Ubuntu. In either case, I do not know how they interact with the MoBo. Any comments are welcome.

Re: GPU computing, Matlab under Windows allows using GPUs for computing fairly painlessly. I was under the (possibly naive) impression that it would be possible to do CUDA programming if I had an Nvidia GPU under any OS. Or were you specifically talking about two GPUs in SLI configuration? That I'd have to use them as independent GPUs?

Thanks, this info is great!

<--- HPC Sysadmin at his day job.

I agree with jchu's recommendations. You don't really need fast machines learning/development. Here's a few tips and answers to your questions:

• Get something cheap like an Athlon II X4 640. It has plenty of cores to play with and costs half what the Phenom II X6 does.
• I'd start with a 4GB kit like this G.Skill. If you find yourself wanting to run larger problems, it is painless to add more memory.
• You don't need a fancy mobo. A cheap-n-cheerful 870 board like this ASRock will do fine.
• No point in getting an aftermarket cooler for this machine
• The Neo ECO is a much better PSU
• You probably don't need a 1TB HDD
• I would probably skip the monitor altogether and get comfortable with SSH or VNC because that's what you'd be using in practice.
• I wouldn't worry about multi-GPU right now. You're going to find that just using a single one is hard enough. SLI is strictly for gaming, so for GPU compute you'd be using the cards independently anyway.
• You'll definitely want to use RHEL or derivative (CentOS, Scientific Linux) because that's what 98% of the HPC centers use and it is well supported by ISVs.

 

[mornington]

<--- HPC Sysadmin at his day job.

I agree with jchu's recommendations. You don't really need fast machines learning/development. Here's a few tips and answers to your questions:

• Get something cheap like an Athlon II X4 640. It has plenty of cores to play with and costs half what the Phenom II X6 does.
• I'd start with a 4GB kit like this G.Skill. If you find yourself wanting to run larger problems, it is painless to add more memory.
• You don't need a fancy mobo. A cheap-n-cheerful 870 board like this ASRock will do fine.
• No point in getting an aftermarket cooler for this machine
• The Neo ECO is a much better PSU
• You probably don't need a 1TB HDD
• I would probably skip the monitor altogether and get comfortable with SSH or VNC because that's what you'd be using in practice.
• I wouldn't worry about multi-GPU right now. You're going to find that just using a single one is hard enough. SLI is strictly for gaming, so for GPU compute you'd be using the cards independently anyway.
• You'll definitely want to use RHEL or derivative (CentOS, Scientific Linux) because that's what 98% of the HPC centers use and it is well supported by ISVs.

It is amazing how much I have been able to learn from one day on this board. Thank you mfenn for your superb input. I am swayed, and I will put together the cheaper system you recommend which means that once I have that up and running, I can start work on a compute node right after. Thanks again.

 

[mfenn]

It is amazing how much I have been able to learn from one day on this board. Thank you mfenn for your superb input. I am swayed, and I will put together the cheaper system you recommend which means that once I have that up and running, I can start work on a compute node right after. Thanks again.

No problem, feel free to PM me if you have any software setup questions.

 

[mornington]

A quick question regarding the video card. Since I am going for a considerably less powerful PC, I was wondering if I could switch video cards? For the one listed above, this is the one I am now thinking about:

Evga GeForce GT 430 1 GB DDR3
http://amzn.com/B0046HAW7Y
$59.99

Is this compatible with the cheap-n-cheerful Asrock mobo mfenn recommended above? Since this appears to support CUDA, and I intend to use my setup only for protoyping, I guess this should be fine.

Thanks!

 

[mfenn]

I'd still go with a GTS 450. CUDA isn't that abstract yet, and you want something that at least resembles the high-end Tesla chips

 

[Ruger22C]

"7) VDU: ViewSonic VA2323wm 23-Inch 16:9 1080p;"

Samsung 2333T

 

[Ruger22C]

"4) GPU: Nvidia GTS 450 from PNY
Price: $122.99"

This is just the 8800GT re-labeled, really.
Regardless, for $10-$20 more you can get a GTX 460 768mB

 

[mfenn]

This is just the 8800GT re-labeled, really.

Uhhhh no. A GTS 250 is a G92. A GTS 450 is a GF106.

 

[Ruger22C]

Uhhhh no. A GTS 250 is a G92. A GTS 450 is a GF106.

Sorry, thanks for the correction.

The 460 is still a much better deal.