Scientific Computing: 980x v. Xeon W3680

[alphaten]

I'm going to be running a lot of finite element analysis for PhD work the next couple of years, and I am about to choose my computer. Unfortunately, the computer program I will be primarily running is 32-bit and single threaded (right now); however, and I can (and will) be running multiple instances of the program.

I have a $4k budget for hardware and have been looking at the 980x and Xeon W3680 (both 3.33 GHz). From Dell, with 12GB of RAM, I'm looking at around $3.7k for the Xeon system (w/ ECC memory) and $2.7k for a 980x system. Any thoughts as to what the smart move might be?

I've also considered dual Xeon processors, but it looks like they might be out of my price range unless I go with the individually slower E5507s (2.26GHz).

 

[degibson]

Smart move depends on where the $4k comes from. If its from an equipment grant, get the best stuff you can that comes with a warranty. If it's personal funds, buy the 980x -- depreciation is a pain.

If you can run many instances, look into Condor -- see if your university runs it somewhere on a communal pool.

Info: Condor is a way to share spare CPU cycles opportunistically on unused machines. Google it for more information.

 

[EliteRetard]

Looks like Intel will be releasing the 990x soon...thatd be a free speed bump if you can wait (probably 3.46GHz)

 

[aigomorla]

Xeon system from Dell for a mission critical machine.

And get ECC.

 

[jacktesterson]

The Xeon

Consider a self build?

 

[Sp12]

The Xeon

Consider a self build?

I say Xeon, but definitely not self-built.

 

[schenley101]

I say Xeon, but definitely not self-built.

second that.

 

[alphaten]

Thanks! I'll go with the Xeon system.

Self built? If this were a personal computer, I'd build it myself (not Xeon though). Since I'm spending other people's money (taxpayers), I feel like I should buy it through typical channels.

 

[Idontcare]

Thanks! I'll go with the Xeon system.

Self built? If this were a personal computer, I'd build it myself (not Xeon though). Since I'm spending other people's money (taxpayers), I feel like I should buy it through typical channels.

It's really up to your confidence level in knowing for sure that you are capable of servicing a DIY machine versus having a pre-built.

When I was in grad school I secured a $25k grant to build a Beowulf cluster. My choices were to go with a 12-node pre-built turnkey cluster or a 24-node DIY cluster.

I was comfortable in my expertise at handling the build as well as managing the serviceability of the computer afterwards so I went the DIY route. It worked great. And having 2x the computing resources meant I finished my doctorate in 5yrs instead of 6yrs.

But on hearing of my success in the chemistry dept there was a guy in the physics dept who wanted to duplicate it so he met with me to go over my build plans. After about 5 minutes it became painfully clear to both of us that DIY was simply above this guy's capabilities...so it was a no-brainer for him, he went with the pre-built that had the service contract.

He got half the hardware for his money (and his projects took all the longer to finish) but in his case going DIY really wasn't a viable option.

I wouldn't let the whole "taxpayer money" etc mental stigma narrow your choices, but I would definitely argue to let your assessment and comfort level of your DIY capabilities to narrow your choices.

 

[RaistlinZ]

Shoot, for something like that I'd be tempted to go with the pre-built with full warranty. If the system craps out... new, free system for you.

 

[PsiStar]

I run electromagnetic software (multi-threaded/64 bit) that takes GB's of RAM & can run for hours/days. Over the years I have had a variety of machines & memory. I see no advantage to using ECC and/or dual Xeon if your software is single threaded/32 bit. For this kind of work ECC has never been an advantage or even noticeable ... just more $$.

Well, the added cores would be nice, but being organized enough to run upwards of 6 (or 12 or 24) instances at one time suggests a pretty incredible self control and organization. Maybe you are?

Because your software is single threaded, get the processor with the highest clock speed available. With my software, the mesher and solver are multi-threaded & I also run multiple instances ... works well as long as there is enough memory. But, there are pre-/post- processing steps that are single threaded and can be a significant delay so I am always looking for that highest clocked CPU myself.

Last, check out Dell's refurbed machines ... on Dell's web site search "refurb". I think they still keep the same warranty as new boxes, but usually you can get a little more speed/RAM/something for less $$.

 

[cbn]

It's really up to your confidence level in knowing for sure that you are capable of servicing a DIY machine versus having a pre-built.

When I was in grad school I secured a $25k grant to build a Beowulf cluster. My choices were to go with a 12-node pre-built turnkey cluster or a 24-node DIY cluster.

I was comfortable in my expertise at handling the build as well as managing the serviceability of the computer afterwards so I went the DIY route. It worked great. And having 2x the computing resources meant I finished my doctorate in 5yrs instead of 6yrs.

But on hearing of my success in the chemistry dept there was a guy in the physics dept who wanted to duplicate it so he met with me to go over my build plans. After about 5 minutes it became painfully clear to both of us that DIY was simply above this guy's capabilities...so it was a no-brainer for him, he went with the pre-built that had the service contract.

He got half the hardware for his money (and his projects took all the longer to finish) but in his case going DIY really wasn't a viable option.

I wouldn't let the whole "taxpayer money" etc mental stigma narrow your choices, but I would definitely argue to let your assessment and comfort level of your DIY capabilities to narrow your choices.

Great information.