![]() ![]() I needed to pick it up from a very messy car, but that is for another day’s story.Artificial intelligence (AI) technology, powered by advanced computing power, a large amount of data, and new algorithms, becomes more and more popular. In case you are curious, the monitor shown above costs $35. To me, $450 for a 3.47 TeraFlops machine is a pretty good trade. A lightweight OS with a minimalistic GUI makes things much simpler. I also found this ubuntu linux desktop is a delight to code on even without the GPU firepower. My little HP box took 34 seconds, well ahead of AWS Tesla K80 instances (71s) and mid-level CPUs (289s). This is not a perfect test for a DL workstation, just a sanity test. I ran a simple Keras MNIST benchmark (described in ) on this $450 machine. When you start a Tensorflow session, the tf will automatically identify the GPU and use it automatically. The GPU build of Tensorflow makes GPU computation effectively a transparent layer. I managed to assemble everything and install the Ubuntu, CUDA, cuDNN and all the machine learning packages onto SSD, and some standard benchmark datasets onto the the HDD, all on one seating after dinner, with a lot of help from. I am amazed how little have changed with the desktop PCs. Putting it up is much easier than I thought. So they ended up with a brand new PC with the original video card extracted for sale, with a reasonable price. Some gamers have acquired this HP Pavilion, which was on sale months ago, only for the Radeon RX 580 Card. Then I have figured this HP is also a gamer’s special. ![]() The HP Pavilion Desktop is an open-box, new item. (i5 6400 2.7GHz, 8GB DDR4, 128GB SSD, 1TB HDD, built-in wifi and card reader) HP Pavilion Desktop - 560-p015HVR PC minus Radeon RX 580 Card, $300 I did everything in 10 hours, and came up with this very short part list: So I set a goal: I will just grab the cheapest GTX 10 series card and a tiny minitower machine on local craigslist and see how much that would cost. (It is very similar to what triathletes do with their bikes) They usually sell their old card with a very reasonable price to finance the newer ones. And gamers have a tendency: they tend to renew their video card much faster then others. In contrast, CPUs are not extremely critical (it’s used in preprocessing, basic operations, and visualizations), I can just use a mid-range, pre-assembled general purpose PC so I can save some cash.ĭuring the quest I noticed one trend: the hardware configuration needed for deep learning has a lot of overlaps with hardcore gamers. GPU memory is also critical since it limits the model scale and the mini batch size. Most builds use GTX 1070 or GTX 1080Ti, which are the higher-end models in GTX 10 series. Majority of the workstation builds I go over online use GeForce GTX 10 series GPU, which is the latest (as of July 2017) of nvidia’s consumer GPU line. The central part of a deep learning workstation is the GPU, or more specifically, nvidia GPU. $1200 is not a price point that you can shell out without suffering the buyer’s remorse. These are powerful machines, but I am looking for something at even lower price point. Further search reveals many other PC builds, mostly priced between $1000 to $2000, with a massive power supply and a transparent case featuring many shiny LEDs. ![]() A quick search online gives you a lot of good starting points, like this or this. Then I decide to build a workstation for deep learning. ![]() Or sometimes there a glitch and the AWS machine goes away, your work would just disappear in the digital wasteland. Sometimes the SGD takes one night to diverge, wasting whole night's work. Put in a bad learning rate you will watch the model die SLOWLY. Training neural net on CPU is just like watching a horror movie in 10x slow motion. Also, I am wondering how things have changed since I got my last desktop.Įven thought I have experimented with neural nets for a year or so, I have to confess that I have been training most of my models on CPUs. So why do I want to put up a clumsy, heavy machine now? Because I figured this is the way I probably can get the most computation power for the least upfront cost. I have not assembled any desktop PC for ages. I got my last desktop from work in 2011, when I found no one claims the lonely HP workstation in the corner. This is when you can use a lot of help from a powerful desktop computer sitting on your desk.ĭesktop PC is a bit out of fashion during the last few years, when everything slowly and steadily moves to the cloud. Even through the neural net / deep learning frameworks are much more sophisticated and easy to use than it used to be one year ago, you still need some (read: a lot) tuning toward a usable DL model. As a data scientist, I do a lot of experimentation on models. ![]()
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |