GPU+CUDA cards on Fedora Core 8 Which cards work?

Can anybody please recommend graphic cards that are sure to work
as a CUDA GPU on Linux Fedora Core 8 64-bit?

Low to mid-range price would be the best, not top of the line.

My system has a Tyan S2915 (Thunder n6650W) (Rev 1.0) motherboard,
and dual AMD Opteron 2214 (dual core) processors.
However, the onboard GeForce 6200 LE video controller is not CUDA-enabled,
according to the NVIDIA web site.
I have vacant PCI Express slots, not sure if x8 or x16.

Thank you.
Gus

Any CUDA card will work in your system - or at least I’d be badly surprised if it didn’t - there’s a couple recent card recommendation threads around. Check these out, see if they have the info you need.

Yeah, just make sure you have the right power supply connectors (at least one six-pin PCIe connector for 8800GT-level cards, 2 6-pin PCIe for GTX 260, 6-pin + 8-pin for GTX 280).

Hello kristleifur

Thank you for our answer.

I am just trying to avoid to get “badly surprised” as you said. :)

Recently a simple Gigabit NICs would not install or be

recoginzed by several versions of its (official) Linux driver on a machine here,

and I had to return it.

So, better get any available compatibility info before buying.

I also saw this posting reporting kernel panic with an NVIDIA-based common card,

BFG GeForce 6200 OC, on Fedora 8, although it is a pre-CUDA card generation:

http://forums.nvidia.com/index.php?showtopic=64301&hl=

I couldn’t find the threads that you mentioned yet, recommending the cards that work,

but I’ll keep trying. Are they on this forum?

Thank you again.

Gus

Hello tmurray (and all who answered)

Thank you for the extra info about the power connectors.

I will go open the box and check

.

I saw you mentioned high end cards.

Do the NVIDIA 8000 card series require power connectors too?

By the way, are there any adapters from, say, standard molex connectors, SATA, or other,

to PCI 6-pin or 8-pin?

Well, maybe it is a different voltage.

I am just getting to PCI express devices now, so I am not familiar …

My goal is to use CUDA for programming, number crunching.

The card is not gaming or computer graphics.

How much card memory would be OK to do programming?

Many thanks,

Gus

I started with Cuda 1.1 and a 9600GT video card, running Fedora Core 8 64-bit. I haven’t had the time to upgrade that computer to Cuda 2.0 beta 2, but wouldn’t expect any problems. Likewise, any G8x, G9x or GTX2x0 card should work, provided you have the correct power connectors.

Thank you Rick Poore and all who helped me so far.

Sorry for taking this long to answer.

I was busy with other projects.

I have additional questions about:

  1. Hardware requirements, and

  2. Memory.

Your comments, suggestions, and clarifications are greatly appreciated!

  1. Hardware/Connector requirements

I checked inside the computer box.

There are vacant PCI express x8 and x16 slots.

Two are side by side, so that a wide card like these can fit.

The power supply is 600W.

There are two idle PCI express 6-pin connectors from the power supply,

and several idle Molex connectors.

There is no 8-pin PCI express connector, though.

Please correct me if I am wrong.

Given the hardware I have, I would be able to buy a GeForce card

up to GeForce 9800 GTX, which uses two PCI-E 6-pin connectors.

(Some models I saw come with Molex-to-PCI-E-6-pin adapters.

Do they work?

I don’t have to use them anyway)

However, I would need another power supply to support card

models GeForce 9800 GX2, GeForce GTX 260, or GeForce GTX 280,

because they require also a PCI-E 8-pin connector which I don’t have in my power supply, right?

I haven’t seeen any Molex-to-PCI-E-8-pin adapters on the Internet either.

Do they exist?

Do they work?

I wonder also if a 600W power supply would be powerful enough to suport the three

upper range card models I just mentioned.

What is your experience or recommendation?

  1. Memory

All GeForce 8800 GT, 8800 GTX, and the 9800 GTX models that I saw come

with 512MB of memory.

I know a GPU is a different processing beast than a CPU.

I presume in a CUDA program the GPU is only actively working and using memory while a

“kernel” (device C function) is loaded on it and at work.

Hence, the memory requirements to program effectively with a GPU may be quite

different (hopefully smaller) than for a CPU,

where data segments can be large and stick around for longer periods.

However, compared to current standards of memory available to a regular CPU,

which nowdays can be typically anywhere between 2GB and 128GB,

the 512MB GPU memory sounds a bit small to me.

Questions:

Is a 512MB good enough for CUDA programming?

Any experiences, bad or good, recommendations, suggestions about card memory?

OK, OK, I know some will tell me that “Big is better” … :) …

… but I can’t afford cards on the $700-$3000 dollar range.

Many thanks to all.

Gus Correa

Either buy a new power supply with 6- and 8-pin PCIe power connectors, or add one of the dedicated VGA power supplies that fit in a 5.25" drive bay. These provide just +12V for graphics cards, providing two 6-pin and two 8-pin PCIe power connectors:

Thermaltake W0157RU 450W $70

ePower EP-350CD 350W $90

FSP Group BoosterX 3 400W $90 (only has four 6-pin connectors)

512 MB is enough to start learning. Whether it’s enough for your application depends on your application. What are you trying to do?

There are some cards with more than 512 MB that work with 6-pin PCIe power connectors:

9600GT 64 cores, 1 GB, $180

8800GT 112 cores, 1 GB, $210

GTX260 192 cores, 896 MB, $300

Thank you Rick Poore for the answer rich in information!

I apprecciate it.

I didn’t even know there were graphic-card-dedicated power supplies that

fit in the 5.25 inch bays!

Somehow since yesterday I can’t reach the Newegg site,

to check the 1GB cards that you mentioned .

Thank you for the tips anyway!

I will probably buy a 512MB card, 8800GT or 9800 GTX,

This should be OK for learning the CUDA API, as you said.

Hopefully the memory and number of processors will be good enough

to do some meaningful projects, to accelerate some Matlab calculations,

to parallelize a few programs, etc.

I hope the experiment can tell me how much effort is required to use the

GPU for scientific applications, and if the results are worth it.

I saw some impressive speedup results in the Nvidia site and other places.

However, I confess I found the CUDA API a bit complex and dominated by low-level

thread and memory control.

The more standard shared and distributed memory parallelization tools for CPUs,

such as OpenMP and MPI, hide these details from the programmer,

and look more friendly to me.

But my bias may be only because I am used to those tools already.

Trying the GPU and CUDA will tell.

Anyway, the point of buying a GPU card is more to learn the technology

and check the potential of the GPU/CUDA on our applications in Earth Science.

It is not to develop production code.

Thank you again.

Gus Correa

External Media PS - Sorry, yet another question that I just came across with.

How about PCI-E 1.1 versus PCI-E 2.0?

On the NVidia site the GeForce 8800GT and 9800 GTX cards appear as PCI-E 2.0.

The motherboard I have (Tyan Thunder n6650W S2915) is PCI-E.

Unclear which release, but most likely it is PCI-E 1.1, not 2.0, as the system is already one+ year old.

I found this posting in Tom’s Hardware which says that video cards built for PCI-E 2.0

are compatible with PCI-E 1.1, although they will use only the lower bandwidth of PCI-E 1.1.

Is this correct?

Any experiences with NVidia GeForce PCI-E 2.0 cards on motherboards with earlier PCI-E bus?

Thank you.

Gus Correa

It works with no problems. My 9600GT is a PCIe x16 v2.0 card, it’s plugged into a GigaByte GA-P35-DS3L motherboard which only has a PCIe x16 v1.1 slot. It means the card’s memory bandwidth for CPU ↔ GPU transfers is limited to the 4 GB/s bandwidth of v1.1 versus 8 GB/s for v2.0.

Helllo Rick Poore and list

Sorry for the late reply.

Yes, you are absolutely right.

I bought a GeForce 9800 GTX PCIe x16 v2.0 card (512MB),

installed in the Tyan Thunder n6650W S2915 motherboard, wich is PCI v1.1,

and the card works fine.

Later I found on the PCIe site a FAQ that confirms what you wrote,

that the only restriction in this case is that the PCIe bandwidth is going to be v1.1, not v2.0.

But I can live with this, I don’t intend to change the computer or the motherboard.

Also, the card required two 6-pin PCIe connectors, which the power supply had.

The power supply is 600W, which is above the minimum 500W suggested by the card manufacturer.

This particular card came with two “molex-to-PCIe-6-pin” adapters.

Not sure if this type of adapter is available for the 8-pin connectors to the higher end

NVidia cards.

Somebody (perhaps NViidia) could post a checklist of minimum hardware requirements for the

various NVidia cards. This would be very helpful.

People coulld then buy a card that would safely work with the computer motherboard and power supply they have.

Regards,

Gus Correa