Nvidia Tesla K80 24GB GDDR5 CUDA Cores Graphic Cards

Before I get into the review itself, note that this card is intended for servers and datacenters, and requires an extensive cooling system. In particular, the K80 is passively cooled, and thus, doesn’t have any fans built into it.

However, it IS possible to use it in a normal desktop, though such usage is not supported by Nvidia. The motherboard and CPU I used are the Asus X99 Pro and the i7 5820k. You must enable the "Above 4G Decoding" option in the Boot menu in BIOS for the K80 to be recognized. Cooling was tricky; I 3d printed a duct with ABS+ from a 240CFM fan which is admittedly quite noisy, but the K80 is stable at 67C at full load. If you’re going to go for a custom solution like ours with a 3d printer, make sure you’re not using PLA as the glass transition temperature is 60-65C.

Now for the actual review…
We use our K80 for simulations and for deep learning.
With deep learning, you’re probably better off with 2 (or maybe even 4) Titan Xs as a single one of those has nearly as much single precision floating point performance as the K80. However, for use cases which require double precision, the K80 blows the Titan X out of the water. We’ve been very happy here with the K80, but if you’re going to use a K80, you’re going to be better off buying it from a vendor who has already built it into a system.

Most of you think it’s a joke but you don’t capture its value in computing in Artificial Intelligence, especially Convolutional Neural Network. I participated in one Kaggle competition where you have to design an algorithm to classify ocean creatures based on their images. It was hard but tractable via Neural Networks, with a moderately-sized NN I trained it to 80% accuracy in 10 days with a GTX970 (4GB and ~1600 Cuda cores). Now with this beast you can train much larger NNs in just a few days!! GPUs are excellent for massive parallel programming as each Cuda core specializes in simple calculations, which can be used to high efficiency for things like matrix manipulation.

For all the people joking about this card and its price – get a life, read more, learn stuff that exists beyond your universe. There is a product there is a demand. Most successful image processing companies would own this card. Do you think movies like Avatar are made from your desktop GPU GTX980??? Google’s latest deep learning results and Google’s self-driven cars are all based on training all sorts of Neural Networks on such monstrous GPUs, because a NN with 100 million parameters cannot reside on a card with merely 6GB memory!

Will update my review once I got it. (Now in negotiation with a China’s hospital to get sponsored for this card, and help them with diabetic retinopathy based on image recognition through patients’ retina scan.)

If this card was made in 1999, it would have been faster then the fastest super computer (SGI Origin 2000) in the world. An that thing cost 40 million dollars at the time and took up a large room with extensive electrical costs. If you put four of these Tesla together with the latest CPU, it would have held the world record until 2002. But then again that old supercomputer (ASCI White) beast wouldn’t have fit in a full tower and the ASCI White cost close to 110 million dollars, weighed 106 tons, and consumed 6 Megawatts of power for running and cooling. This thing is a bargain in comparison.