I like what you are doing. I've been running folding@home for almost 3 months now on 2 computers. Im at about 123,000.000 work units. It was nice to see what titans could achieve. I'd like to see a video if you could compare the GPUs to the CPUs.
Why did you put game cards instead of K80? Or do you make a separate server for K80, and Titans only for graphics on a large monitor? Using gaming graphics cards for important computations is crazy, I've tried. There were a lot of problems and I sold my 4 GTX 1070, then I bought 2 Tesla K80.
The main reason is that each Titan Xp cards is 2 to 3 times faster than a K80 and they have the same 12 GB of RAM per GPU. I have had no problems with computations on these they are just quicker, much quicker.
@@robingrosset6941 You're in luck, I got errors in computational cores and memory failures on the GTX 1070. At the same time, the cards worked perfectly in tests and in games, which is why I was able to sell them in one day and buy a pair of K80s. In addition, one Titan, like yours, is more expensive than two K80s and it does not have much fewer double precision cores and no memory error control. If they work better for you than K80, I congratulate you. But there was already a story when calculations on Titans on one task gave a difference of 20-30% for each pass. On the claims, Nvidia said buy Tesla, important things are not considered on the cards for games.
Saber Toothed I am very interested in the type of calculations, I am doing machine learning models typically deep learning. Would really like to understand what is causing the problem. What kind of calculations are you doing? What language and library are you using.
@@robingrosset6941 Linux Ubuntu 16 Server, CUDA 11, Python, TenzorFloy. In my case, the problems were caused by old video cards bought on the secondary market after mining. They overheated and this led to a violation of the crystal structure of the graphics chip and memory. In games, single core and memory failures are invisible, while in calculations - cuda kernel panic.
Saber Toothed I use the same libraries. I have heard of old mining cards failing more quickly as they have overheated. I have run my cards K80s and Titan Xp for months at 100% utilization. The cooling solutions I found for the K80s keeps them below 70 degrees C. The Titans stay around 84 degrees C on max with their stock coolers but never had a kernel panic. The cards should also thermal throttle, lower power consumption and slow down to avoid damage. NVIDIAs own manufactured cards seem to last really well, the only card that I ever had fail was from another vendor and died after heavy loads. I had a component explode on the board. I think it was a capacitor. There was nothing left of it.
The Tesla K80 was launched quite a few years ago now and waterblocks are hard to find in production still and also not cheap. Here is one which is $140 www.ebay.com/c/24017789794 which is a pretty good price. But compare this to 2 80mm high CFM fans for $40. It’s cheaper and simpler to air cool and it works well. If I needed a silent system I might do water cooled and I have other water cooled stuff, for me it was mostly a price issue.
Do you think there's a way to get a display off of this card? I've been looking for a while, and I happened to stumble on your F@H video, and thought you could test this for content or if you've tried it already. Like take a PCIE display adapter and try to get high frame rates in some games. Or do a passthrough like what LTT did with the "mining 1060" a few years ago
I like what you are doing. I've been running folding@home for almost 3 months now on 2 computers. Im at about 123,000.000 work units. It was nice to see what titans could achieve. I'd like to see a video if you could compare the GPUs to the CPUs.
I will take a look at this and see if turns up some interesting data.
Why did you put game cards instead of K80? Or do you make a separate server for K80, and Titans only for graphics on a large monitor? Using gaming graphics cards for important computations is crazy, I've tried. There were a lot of problems and I sold my 4 GTX 1070, then I bought 2 Tesla K80.
The main reason is that each Titan Xp cards is 2 to 3 times faster than a K80 and they have the same 12 GB of RAM per GPU. I have had no problems with computations on these they are just quicker, much quicker.
@@robingrosset6941 You're in luck, I got errors in computational cores and memory failures on the GTX 1070. At the same time, the cards worked perfectly in tests and in games, which is why I was able to sell them in one day and buy a pair of K80s. In addition, one Titan, like yours, is more expensive than two K80s and it does not have much fewer double precision cores and no memory error control. If they work better for you than K80, I congratulate you. But there was already a story when calculations on Titans on one task gave a difference of 20-30% for each pass. On the claims, Nvidia said buy Tesla, important things are not considered on the cards for games.
Saber Toothed I am very interested in the type of calculations, I am doing machine learning models typically deep learning. Would really like to understand what is causing the problem. What kind of calculations are you doing? What language and library are you using.
@@robingrosset6941
Linux Ubuntu 16 Server, CUDA 11, Python, TenzorFloy. In my case, the problems were caused by old video cards bought on the secondary market after mining. They overheated and this led to a violation of the crystal structure of the graphics chip and memory. In games, single core and memory failures are invisible, while in calculations - cuda kernel panic.
Saber Toothed I use the same libraries. I have heard of old mining cards failing more quickly as they have overheated. I have run my cards K80s and Titan Xp for months at 100% utilization. The cooling solutions I found for the K80s keeps them below 70 degrees C. The Titans stay around 84 degrees C on max with their stock coolers but never had a kernel panic. The cards should also thermal throttle, lower power consumption and slow down to avoid damage. NVIDIAs own manufactured cards seem to last really well, the only card that I ever had fail was from another vendor and died after heavy loads. I had a component explode on the board. I think it was a capacitor. There was nothing left of it.
For the K80, did you ever think to use liquid cooling via waterblock? If so, why did you choose not to do it?
The Tesla K80 was launched quite a few years ago now and waterblocks are hard to find in production still and also not cheap. Here is one which is $140 www.ebay.com/c/24017789794 which is a pretty good price. But compare this to 2 80mm high CFM fans for $40. It’s cheaper and simpler to air cool and it works well. If I needed a silent system I might do water cooled and I have other water cooled stuff, for me it was mostly a price issue.
Do you think there's a way to get a display off of this card? I've been looking for a while, and I happened to stumble on your F@H video, and thought you could test this for content or if you've tried it already. Like take a PCIE display adapter and try to get high frame rates in some games. Or do a passthrough like what LTT did with the "mining 1060" a few years ago
Could you try to run the GPT2-1558M model on this & show us the AI responses, as it wasn't able to run on the Jetson Xavier
Nice!