I built a ~$10k PC for deep learning (among other types of computation). Feel free to email me if you have any questions about choosing components or building it.
I wouldn't expend that much money right now. I think we are just starting to see the evolution in the hardware for deep learning and that in the next 3-4 years we are going to see a really big jump in terms of performance and performance per watio. So I would expend less money and update the system more frequently. I would want to know what do you think after taking the decision of expending so much money in a system like that. Just see how in a year nvidia released 1080, Tital P and 1080Ti. I consider this the first generation designed with DeepLearning in mind.
* i7-6900K CPU
* 4x GTX 1080 GPUs
* X99-E WS/3.1 Motherboard
* 128GB DDR4 2400Mhz RAM (8x 16GB sticks)
* 14 TB between SSD and HDD (Plex, datasets and applications on Ubuntu 14)
* 14x 120mm fans run at lowest RPM
* Custom liquid cooling loop with dual pumps
* 4x Radiators (1x 120mm, 1x 240mm, 2x 480mm)
It's not just deep learning, a lot of what I like to do with this is research or projects parallelizing things like cryptanalysis. For the rest I do financial modeling (which has a fair overlap with the cryptanalysis from a signal processing perspective anyway). I used to work primarily in application security, but have been slowly moving into cryptography and data analysis; the latter mostly as a supplement for cryptography as a research interest.
What kind of modeling are you doing with a beast of a setup like that?
1. Use one stock that I believe I have unique and actionable data on as the input barometer. Take it as a premise that this stock is currently mispriced by the market, and I have an idea of what it should be. We'll call this the primary stock.
2. Find all stocks that have a consistent positive or negative correlation with the input stock, (do they share volatility?), at different significance levels. (This is not particularly hard, it's just getting accurate data with real granularity that is the challenge.) These are the corollary stocks. If there is a calendar event upcoming, I restrict the backtest to the values of other stocks on similar days.
3. For each correlated stock found, calculate an options model with a new terminal price distribution (as opposed to what is currently priced by market consensus).
4. Simulate different options trading strategies and their outcomes across different projected prices for each stock and present the ones with the highest probability of success versus return. At a minimum, the primary and corollary underlyings should have a large enough misprice that trading them can be profitable across the bid/ask spread. So for example, a $1 increase or decrease will not generally be meaningful enough to make the trade, especially if there is a lot of volatility priced in (such as around earnings releases).
You'll notice that there is nothing truly sophisticated about any of this. It's just a bunch of computation used to extrapolate the best leverage available using derivatives. The core thesis involves having material data that the rest of the market has not yet priced in, and I don't often have that (or rather, enough of a directional conclusion based on it). When I do, I like to do this because 1) I don't trust most profitability calculators which assume constant volatility and 2) it maximizes potential results by not just looking at the primary stock I know about.
tl;dr: Assuming one stock is mispriced, do a bunch of derivative re-pricing and choose an optimal one.