[SOLVED] Making use of a discrete GPU card (Tesla K20)

[jhsachs]

I've got a Tesla K20 GPU card that I want to try out. I've found very little documentation. For example, the K20's user manual contains nothing but specifications -- no instructions. It seems to be written for system integrators who don't need to be told how to use the card because they just naturally know.

Does anyone know a source of practical information on the use of GPU cards in general, or the K20 in particular? Following is a list of questions I have... I'll appreciate answers but what I really want is a good source for answers. I'll probably have a new set of questions soon after I answer these.

Is the K20 compatible with my GeForce GTX 680? (The 680 and the K20 were both introduced in 2012, so the answer should be "of course it is," but the general lack of information makes me skeptical about everything.)

How do I connect the two cards... or don't I? I've seen flat statements that the K20 doesn't support SLI, but if it doesn't, why does it have SLI connectors?

If I need an SLI bridge, what type? The 680 and K20 both have three fingers on the top of the board. I've seen bridges that connect the contacts on the front and rear fingers, and bridges that connect the front or rear fingers, but not both. (And some bridges have both configurations for different cards!) i've never seen a bridge that connects the middle fingers, making their existence a mystery.

All of the rigid bridges I've seen have connectors that are spaced so they can't be aligned with both cards at once. (That may be confusing, so I'll state it a second way: if the bridge's end connectors are attached to the card in the first PCIe slot, the fingers on the card in the second slot are squarely between the bridge's second and third connectors.) I can connect two cards with a ribbon bridge, but again, the weirdness makes me suspicious: Why would NVIDIA make rigid bridges that can't be used with a standard motherboard or case? I'm missing something here, but what?

And what about software? The K20 specs say it supports DirectX and OpenGL. If I run an application that uses one of those APIs, will it automagically make use of the K20 once the K20's driver is installed? Or must I recite some spells to make it work, or must I use special applications that are written to take advantage of a discrete GPU card?

 

[kanewolf]

Tesla cards were designed to be used in servers. They don't have any video output capabilities. At least there are some K20s that are actively (fan) cooled. Many Tesla cards are passively cooled and can't be used in desktop cases.
You can't SLI. You could use it for rendering. It is very limited in usefulness to a desktop PC.

 

[jhsachs]

I may have worded my question carelessly. When I said "I want to try out" a K20, I meant that I want to find out how to use it. In other words, I want to learn something.

If I must set up a server to do that, maybe I'll learn to set up a server. Solving a problem that has to be solved to solve another problem can be very interesting and productive.

You've answered one question I had, about the applicability of SLI, but several others remain. And my primary question, where to find a good source for information, hasn't been addressed. I hope someone can remedy that.

You raised a new question too: about cooling. The K20 does not have a fan. I assumed that meant it doesn't need one. You told me such cards have to run in a server case. Can you tell me why, and what a server case does? I always thought that a standalone server case was simply a case that has room for a lot of disks and other gadgets and allows easy access for maintenance. This is the first I've heard that it can actually do something a consumer case cannot do.

At the moment I'm operating my test system with no case at all. I'm planning to build a new one in a bench case, whose top will remain off. From what you've said so far, I can't tell whether that means the cooling problem you mentioned won't affect me, or that I won't be able to solve it!

 

[kanewolf]

I may have worded my question carelessly. When I said "I want to try out" a K20, I meant that I want to find out how to use it. In other words, I want to learn something.

If I must set up a server to do that, maybe I'll learn to set up a server. Solving a problem that has to be solved to solve another problem can be very interesting and productive.

You've answered one question I had, about the applicability of SLI, but several others remain. And my primary question, where to find a good source for information, hasn't been addressed. I hope someone can remedy that.

You raised a new question too: about cooling. The K20 does not have a fan. I assumed that meant it doesn't need one. You told me such cards have to run in a server case. Can you tell me why, and what a server case does? I always thought that a standalone server case was simply a case that has room for a lot of disks and other gadgets and allows easy access for maintenance. This is the first I've heard that it can actually do something a consumer case cannot do.

At the moment I'm operating my test system with no case at all. I'm planning to build a new one in a bench case, whose top will remain off. From what you've said so far, I can't tell whether that means the cooling problem you mentioned won't affect me, or that I won't be able to solve it!

If the K20 you have does not have a fan, then it was designed for specific 2U server cases that have high airflow. When we purchased some at work, we had to use Dell R720 servers because they were some of the first that were certified for the K20s. You need a case with a high positive pressure to force air through the K20 and out through the exhaust vents in the end.

Tesla K series GPUs are useful in computation (CUDA or openCL). To utilize it you either need to start learning how to program in CUDA or use software that uses CUDA/openCL.

 

[jhsachs]

Now the open front end on this type of card makes sense. That must be where the air flows in.

It seems to me that I should be able to confirm that the hardware and software are working by loading the card too lightly and briefly to heat the GPU to a dangerous level. I assume I can monitor the GPU temperature, just as I can for a graphics card, to know when I'm pushing it into dangerous territory. I've seen the heat sinks on these things; the GPU shouldn't heat up so fast that I don't have time to ease off.

Beyond that, it shouldn't be hard to make an end-to-end air cooler from a couple of little fans and a scrap of aluminum sheet. The shroud has threaded holes on the end where I can attach one. Whether it would have enough capacity is another question, but an answerable one.