[biznatch]

The larger the amount of RAM you want you will have to invest in ECC/Registered RAM (which is not cheap) to avoid problems.

Is this just because with more RAM there's a greater chance of something going wrong, or is there something inherent about having lots of RAM that requires ECC/Registered RAM?

[ymc]

Ymc,

There are several workstation models (from Dell, Lenovo, HP to name a few) that can have > 100 GB RAM (this is a representative link for Dell workstations: http://www.dell.com/us/enterprise/p/precision-desktops).
For example, the T7500 workstation from Dell can have a maximum of 192GB of DDR3 RAM. If you are going to invest in a machine with that much of RAM you should consider getting a proper workstation (or better still a proper server). The larger the amount of RAM you want you will have to invest in ECC/Registered RAM (which is not cheap) to avoid problems.

But this is assuming you put in 16GB sticks in the DIMM slots. These sticks are super expensive and the fastest ones are only DDR3-1066.

On the PC with 8 DIMM slots, you can put eight 4GB DDR3-1600s for only US$200.

[ymc]

Is this just because with more RAM there's a greater chance of something going wrong, or is there something inherent about having lots of RAM that requires ECC/Registered RAM?

It is the former.

But as far as I know, some supercomputers use non-ECC RAM anyway because of ECC's speed penalty and cost concern. They would rather write code to detect memory errors and re-run the subtasks.

[tonybolger]

Is this just because with more RAM there's a greater chance of something going wrong, or is there something inherent about having lots of RAM that requires ECC/Registered RAM?

You need ECC because of the former, and registered RAM because of the latter, since driving all those RAM chips is too "heavy" for a single bus, so you proxy it. The Xeon EX takes it further, and has another memory buffer / proxy chip on the motherboard.

[GenoMax]

There is a reason these machines are called "workstations". They use premium enterprise quality components (you could perhaps build one cheaper with the same components but then you would need to deal with the manufacturers independently for any warranty issues) and are meant to be workhorses for 24/7/365 computing. These also use motherboards that have additional features/chipsets that are not present in consumer level boards.

If you were building a gaming machine then you can afford to have a crash or two now and then. On the other hand, if you are looking to assemble/align a large dataset you would not want that crash to happen when your process is 90% done.

If you are putting this rig together on your own with no institutional support then your budget should be the first and only consideration. Otherwise think about long term use, reliability and stability.

But this is assuming you put in 16GB sticks in the DIMM slots. These sticks are super expensive and the fastest ones are only DDR3-1066.

On the PC with 8 DIMM slots, you can put eight 4GB DDR3-1600s for only US$200.

[lvcosme]

Is anyone using a SSD as swap for de novo assembly? I have an I7 overclocked to 4.2Ghz, 32GB RAM DDR3 1600, and I also have 3 SSD of 120GB. I was thinking to use one of them as swap, and then I would end up with ~150Gb RAM + swap. Other option is to change the motherboard and get one that supports 120Gb RAM, but memory still expensive. The transcriptome of my organism is around 400Mbases and its genome is 1.3Gbases. I have 500 million reads (50bp - HiSeq2000).
Thank you.

[ymc]

SSD swap is no use for memory intensive application given that it is 10x slower than memory

[adaptivegenome]

What do people consider a reasonable price for "cheap" desktop system?

[ymc]

What do people consider a reasonable price for "cheap" desktop system?

US$3000???

For that, you should be able to get
i7 3930K
64GB RAM
GTX 580 3GB
2x 128GB SSD raid0

[adaptivegenome]

US$3000???

For that, you should be able to get
i7 3930K
64GB RAM
GTX 580 3GB
2x 128GB SSD raid0

Ok thanks, I just wanted a reference point in comparing hardware...

[arvid]

Is anyone using a SSD as swap for de novo assembly? I have an I7 overclocked to 4.2Ghz, 32GB RAM DDR3 1600, and I also have 3 SSD of 120GB. I was thinking to use one of them as swap, and then I would end up with ~150Gb RAM + swap. Other option is to change the motherboard and get one that supports 120Gb RAM, but memory still expensive. The transcriptome of my organism is around 400Mbases and its genome is 1.3Gbases. I have 500 million reads (50bp - HiSeq2000).
Thank you.

Why not buy cloud computing time or run it on XSEDE Blacklight or something similar if this is a one-off thing? Or ask someone in your neighbourhood with a beefy server to collaborate on that assembly?
I favor the view of getting the right tools for the right task - if you're in an institution make it clear to the managers that they should spend money on decent computer hardware. It would be a waste of their money to pay a bioinformatician to build and maintain unsuitable hardware and wait for jobs to (hopefully) finish on slow and unreliable machines. Just my 2 cents.

[arvid]

Ok thanks, I just wanted a reference point in comparing hardware...

You wouldn't need to spend that much more money to get a decent server. We just got a 64 core server (4xAMD 6276) with 512 GB ECC DDR3-1600 RAM (32x16 GB) for ~14 000 €. With 256 GB RAM (32x8 GB), it would have been much cheaper (about half the price). We get a good academic discount, however, the list price would be over 20 000 €.
~6 times the money for 8 times more CPU and 8 times more RAM... and higher reliability. No SSDs or GPUs included, however.

[GenoMax]

One would not want to use a regular SSD for swap but there are PCI-e based SSD drives that could potentially be used.

One such example is Intel SSD 910 (there are others but intel is one of the more reliable vendors) . At $2000 for 400GB (MRSP) it is not meant for desktop applications.

This is what we are specifically talking about:

A single PCIe 2.0 lane is good for 500MB/s of data upstream and downstream, for an aggregate of 1GB/s. Build a PCIe 2.0 x16 SSD and you're talking 8GB/s in either direction.

Below are specs from Intel.
---------------------------------------------

New Intel 900 SSD Family Expands SSD Product Line with PCIe Interface SSD for Accelerated Data Center Storage

PCIe-based 400GB and 800GB Intel SSD 910 Series provides extreme performance, endurance and reliability for rigorous data center demands.
Intel SSD 910 Series offers easy-to-install, seamless post-deployment server storage upgrade with no changes to existing server design.
Using Intel High Endurance Technology, Intel SSD 910 offers 10 full drive writes a day for 5 years for 30x endurance.*

[adaptivegenome]

You wouldn't need to spend that much more money to get a decent server. We just got a 64 core server (4xAMD 6276) with 512 GB ECC DDR3-1600 RAM (32x16 GB) for ~14 000 €. With 256 GB RAM (32x8 GB), it would have been much cheaper (about half the price). We get a good academic discount, however, the list price would be over 20 000 €.
~6 times the money for 8 times more CPU and 8 times more RAM... and higher reliability. No SSDs or GPUs included, however.

Arvid, I agree. I actually have several nodes composed of 4x AMD 16-core chips. We stuck gtx480s in the nodes in hopes that we could squeeze some more computation out by offloading things to the GPU. For us the AMDs were great because they are not much slower than the intel chips and FAR cheaper.

I was curious what people thought a good price for a commodity computer was...

[GenoMax]

I was curious what people thought a good price for a commodity computer was...

The answer would be dependent on what one is comfortable with .. in terms of cost (personal/institutional funds) and tolerance for associated risk based on intended application (personal research or wider support for others).