If you've ever had the opportunity to train a neural network, you've probably realized that it's a major time saver to start with weights from a pre-trained model. Training a model from scratch can take hours, days, or even sometimes weeks depending on the complexity of the model, the amount of training data, and the resources available.
Here at KickView, we already spend a good deal of time and resources creating datasets and optimizing models. So it was an exciting morning when a large box arrived with a brand new DGX Station!
For those of you who have lived under a rock for the last year or have not kept up to date with NVIDIA's news releases, the NVIDIA DGX-1 is currently the top of the line in machine learning hardware, and the DGX Station is its little brother. This machine is a little larger than a standard desktop but will easily fit under a desk. Boasting 4 Tesla V100 GPU's, 64GB of GPU memory, 2,560 Tensor Cores and 20,480 CUDA Cores, this beast can crank through data with ease.
The specs of the machine are impressive, and NVIDIA has done a great job making the machine look powerful and aesthetically pleasing. A push of a button on the back of the gold colored custom machined chassis pops open the side of the machine, and you are greeted with a view that would put a smile on the face of anyone who has has the pleasure of putting a machine together. Four V100 GPU's are the obvious centerpiece, complemented perfectly by a custom GPU liquid cooling loop.
We decided to run a few quick tests to compare the efficency of using the DGX Station to some of our current dev machines which have two Titan Xp GPUs each.
Here are some of the results:
|Data Set||Model||Epochs||DGX Time||Dual Titan Xp Time||Performance Gain|
|Kitti||DetectNet||30||17 Min 54 Sec||107 Min||597.8% Faster|
|POC2012||GoogLeNet||30||9 Min 14 Sec||15 Min 5 Sec||163.36% Faster|
|POC2012||AlexNet||30||4 Min 23 Sec||7 Min 7 Sec||162.36% Faster|
|Cifar10||MiniGoogleNet||70||27 Min 24 Sec||16 Min 32 Sec||60.39% Faster|
As you can see, most of the time and performance gains are quite significant. If you have small images in your dataset, you will not see as much gain in using the DGX, but when processing large images in a large net... that's when the DGX shines. With over 5 TB of raided SSD disk storage and 10G ethernet connections, getting data on and off the machine is fast and easy.
If you'd like to read more about the DGX station here's a link to the specsheet.
Even though this machine costs more than other lower performance options, we believe that the amount of time saved during experimentation, refining and optimizing models makes owning a DGX a wise choice.