Graphics Card or Tesla selection of hardware for computations

I am very much new to CUDA and also to GPUs, parallel processing.
My work is to do Monte Carlo simulations PARALLELLY for computational biology applications. Not involving much graphics applications other than plotting variables in runtime.

I have gone through the topics in this forum and not able to decide the hardware I have to purchase.
I have two choices in mind.

  1. Intel Xeon quadro (2.3 - 3.0 GHz) + nvidia quadro FX570 (256MB) graphics card + Tesla C1060 + (4-8 GB) RAM

  2. same as above, but, excluding Tesla C1060, 256MB-graphics card and instead, including a high end graphics card FX 4600 (768 MB).

My questions are, which is a better choice. Suggest me anything better also.
which is effective to do parallel computing using CUDA, graphics card or Tesla C1060?

Otherwise is it possible to use the power of graphics card, Tesla simultaneously for my computations.

Help me getting started. Thanks.

If you just want to do computations, then Tesla is probably the best. If you want to visualize what you’re simulating while it’s happening, then a graphics card may be a better option.

But, is it still possible to do computations with Graphics card also ? Thanks


By the way, Quadro FX4600 has less memory and just under half the processing power of Tesla C1060.


Thanks a lot. I would prefer Tesla, in that case.

I believe the Tesla C1060 is comparable with the Quadro FX 5600, not the 4600.

You might also want to consider a C1070, I don’t know of its Quadro equivalent, the gaming analogue is the GTX 280 (which has many advantages over C1060).

Basically, if you want to render something based on your computations, you might want to use a Quadro or a GeForce. Teslas don’t have a display-out, and you would need to transfer the data from one card to the other (the primary one in your setup). If it’s just some 2D analysis data, then that’ll work just fine without imposing bottlenecks. If you want to display full-fletched visualisations (eg volume rendering) in real time, then this comment is relevant.

There is no C1070.

[edit: deleted because my statement was simply wrong]

mmm. Forget all I’ve said. The C1060 is what I thought was the G8x-based C-Tesla, and it is GT200-based. What was the name of the G8x-based Tesla again? C870. Sorry, I confused myself.

So again:
C870 should be compared to Quadro 5600. C1060 compared to Quadro 4600 is apples with pears as Paulius pointed out. C1060 compares to GTX 280. I don’t know the Quadro analogue that is GT200-based.

In my personal option because of the pricing Tesla is only the better option if you require
a) the extra memory or
b) are unable to power the normal graphics card.

but which are the advantages of a Tesla over a Geforce?

Because A c1070 cost as much as 4 gtx260, and i seriously doubt it can go faster…

To my knowledge, Teslas differ from GeForces in:

  1. How much memory they have - Teslas can have more onboard memory but in the case of C1060 this memory had to be made considerably slower than GTX280’s. C870’s memory also has been made slightly slower than 8800GTX’s.
  2. How stable they are - Teslas get tested more thoroughly to assure there can be no random bad bit errors etc. Not that there are many in off-the-shelf GFs but with Teslas your hardware should be more dependable.
  3. Teslas might get dedicated tech support (I think I’ve heard about that somewhere but can’t really find any info on this now)
  4. Teslas consume a little bit less power (10-15W)
  5. Teslas don’t have monitor connectors
  6. Teslas come in rack units, handy if you want to build a farm of those

As far as I know, Tesla cards sport the same exact chip as the corresponding video card but different memory (more but generally slower). They are not inherently faster or better performing than off-the-shelf gaming video cards. In fact, in terms of pure performance, GTX280 may score better than Tesla C1060 since it has about 30% better device to device bandwidth. However, this remains a video game graphics card and some companies may prefer something more “serious”, dependable, tested etc. even if it costs a few times more.

Do NOT get a quadro. If you need the RAM (and only if you need the ram), get a Tesla. If you don’t, buy a GeForce.

Of course, Big_Mac put it better than I. Especially when he said “However, this remains a video game graphics card and some companies may prefer something more “serious”, dependable, tested etc. even if it costs a few times more.”

Which translates to: Companies buy Teslas (and many kinds of other over-priced garbage, computer-related or not) because they’re fn stupid. (And because they spend more time sitting lazily in the conference room while they’re lied to by sales guys than expend the effort to do actual research.) You have no idea how much of the overal economy’s inflation and unrealized GDP growth is owed to such practices.

I hope to be buying a S1070 in the near future, not because I am stupid, but because it fits in our rack. If you can point me to some way to put a bunch of GTX280s in a rack, I would be very happy. I think I can manage most of my problems with 1 Gb of memory, so the extra bandwidth would be more than welcome.

If you would know how much trouble I have to push CUDA because of form factor issues…

the quadros for the gt200 series are the 5800 which is currently only sold inside the Quadro Plex 2200 and is parallel to the gtx280 and the new quadro for graphics the quadro cx which is parallel to the gtx260. You can always use the gtx280 which will be cheaper and give you the best clocks (almost) and then render with the opengl interops which is really fast compared to moving data back to the cpu and again to the gpu for rendering. ;)