4 ms·

I want something like this for deep learning at home (and gaming on my free time)

by jorgemf 10y ago

Dylan16807 10y ago
So go buy a GTX 1070?
- jorgemf 10y ago
  only the 1070 is more expensive that the console probably, and it only has 8GB of RAM. The console would have a price around $500-$700 and it will have 12GB of RAM + CPU + HHD. Only a decent GPU with 12GB will cost you more. And that is my point
  - Dylan16807 10y ago
    A 1070 is $340, while this console is probably $400 or $500. And while it technically has 12GB of ram, a big chunk of that is reserved by the OS. In practical terms you would only be able to GPU on 8GB of it. That's not even considering how narrow the niche is that absolutely needs more than 8GB but fits under 12GB.
    - jorgemf 10y ago
      Well, it was more a dream thing. Obviously I would want it to run Linux (which takes very few RAM) and use Nvidia instead of AMD. And probably the hardware will be outdated when they release it to the market
      - Dylan16807 10y ago
        And I'm saying that your dream is already reality, just in a slightly different form.
ansible 10y ago
What's stopping you from setting up a PC to do this now? GPUs with 4GB of RAM like the Nvidia GTX 1050Ti aren't expensive. I just built a few Kaby Lake workstations with NVMe SSDs and those GPUs, and it is a slick little platform.
- sremani 10y ago
  One of the best ways to put life back on to long unused desktops lying around is 1. 1050Ti ($150) 2. Ubuntu 16.04 LTS I did this to a 6 year old Dell XPS 8300, and now that thing rocks! (I do a little bit of Tensorflow playing around on it).
  - ansible 10y ago
    Yup. General comments to the folks reading this from home: For machine learning, you'll want as much RAM on the GPU as possible. I can't recommend a 2GB GPU card. 4GB cards can be less than $140, and then there is a jump in price to the 8GB cards (6GB cards aren't cheaper enough to consider, in my opinion). Check that your power supply has the extra PCIe 8-pin power connector, though some of the GTX 10xx series cards don't require any additional power. 8GB system RAM minimum is also what I'd recommend. Some NN applications can benefit significantly from a faster CPU, even though it is mostly running on the GPU. For example, for training a Caffe model, an Ivy Bridge i7 was vastly superior to a Kaby Lake i3, using the same GPU card.
- jorgemf 10y ago
  Well a video console would be around $500. You cannot have a computer with 12GB DDR5 in the GPU and a lot of cores for $500.
bsamuels 10y ago
unfortunately, you're not going to get very much machine learning mileage out of an AMD GPU like Scorpio is using
- jorgemf 10y ago
  Yes, I know. I was just dreaming about a complete machine for deep learning for $500-$700.
roel_v 10y ago
I have a workstation roughly like that (ddr 4, 16 gb ram, gtx 1070), for around $2k, and some of that is because I went overboard on the mobi. I have a spare cpu socket on the mobo with which for $300 I can upgrade to 16 cpu cores, too. It's well within reach for a regular workstation build.
- jorgemf 10y ago
  No you don't. Your GPU only has 8GB of DDR5, and only the GPU is more expensive than the complete video console with 12GB of DDR5 + CPU + HHD. And that is my point, the hardware for the price.
  - roel_v 10y ago
    Oh it's 12 GB on the GPU, I misread, I thought it was regular RAM - which confused me because of the DDR5. Either way, sure, 2k is more than the console will cost. But you said you 'wanted it', not 'wanted it at console price'. My point was - it's not particularly high powered (compared to other things that can still be build at home), and well within reach of most people here.
dsacco 10y ago
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.
- kayoone 10y ago
  i'd like to hear more on that $10k PC :)
  - dsacco 10y ago
    I don't have a ton of time at the moment, but here's the cliff's notes: * 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.
    - artmageddon 10y ago
      > For the rest I do financial modeling (which has a fair overlap with the cryptanalysis from a signal processing perspective anyway) What kind of modeling are you doing with a beast of a setup like that?
      - dsacco 10y ago
        The approach has a few tiers: 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.

I want something like this for deep learning at home (and gaming on my free time)

by jorgemf 10y ago

I want something like this for deep learning at home (and gaming on my free time)

Dylan16807 10y ago

So go buy a GTX 1070?

jorgemf 10y ago
only the 1070 is more expensive that the console probably, and it only has 8GB of RAM. The console would have a price around $500-$700 and it will have 12GB of RAM + CPU + HHD. Only a decent GPU with 12GB will cost you more. And that is my point
- Dylan16807 10y ago
  A 1070 is $340, while this console is probably $400 or $500. And while it technically has 12GB of ram, a big chunk of that is reserved by the OS. In practical terms you would only be able to GPU on 8GB of it. That's not even considering how narrow the niche is that absolutely needs more than 8GB but fits under 12GB.
  - jorgemf 10y ago
    Well, it was more a dream thing. Obviously I would want it to run Linux (which takes very few RAM) and use Nvidia instead of AMD. And probably the hardware will be outdated when they release it to the market
    - Dylan16807 10y ago
      And I'm saying that your dream is already reality, just in a slightly different form.

ansible 10y ago

What's stopping you from setting up a PC to do this now? GPUs with 4GB of RAM like the Nvidia GTX 1050Ti aren't expensive. I just built a few Kaby Lake workstations with NVMe SSDs and those GPUs, and it is a slick little platform.

sremani 10y ago
One of the best ways to put life back on to long unused desktops lying around is 1. 1050Ti ($150) 2. Ubuntu 16.04 LTS I did this to a 6 year old Dell XPS 8300, and now that thing rocks! (I do a little bit of Tensorflow playing around on it).
- ansible 10y ago
  Yup. General comments to the folks reading this from home: For machine learning, you'll want as much RAM on the GPU as possible. I can't recommend a 2GB GPU card. 4GB cards can be less than $140, and then there is a jump in price to the 8GB cards (6GB cards aren't cheaper enough to consider, in my opinion). Check that your power supply has the extra PCIe 8-pin power connector, though some of the GTX 10xx series cards don't require any additional power. 8GB system RAM minimum is also what I'd recommend. Some NN applications can benefit significantly from a faster CPU, even though it is mostly running on the GPU. For example, for training a Caffe model, an Ivy Bridge i7 was vastly superior to a Kaby Lake i3, using the same GPU card.
jorgemf 10y ago
Well a video console would be around $500. You cannot have a computer with 12GB DDR5 in the GPU and a lot of cores for $500.

bsamuels 10y ago

unfortunately, you're not going to get very much machine learning mileage out of an AMD GPU like Scorpio is using

jorgemf 10y ago
Yes, I know. I was just dreaming about a complete machine for deep learning for $500-$700.

roel_v 10y ago

I have a workstation roughly like that (ddr 4, 16 gb ram, gtx 1070), for around $2k, and some of that is because I went overboard on the mobi. I have a spare cpu socket on the mobo with which for $300 I can upgrade to 16 cpu cores, too. It's well within reach for a regular workstation build.

jorgemf 10y ago
No you don't. Your GPU only has 8GB of DDR5, and only the GPU is more expensive than the complete video console with 12GB of DDR5 + CPU + HHD. And that is my point, the hardware for the price.
- roel_v 10y ago
  Oh it's 12 GB on the GPU, I misread, I thought it was regular RAM - which confused me because of the DDR5. Either way, sure, 2k is more than the console will cost. But you said you 'wanted it', not 'wanted it at console price'. My point was - it's not particularly high powered (compared to other things that can still be build at home), and well within reach of most people here.

dsacco 10y ago

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.

kayoone 10y ago
i'd like to hear more on that $10k PC :)
- dsacco 10y ago
  I don't have a ton of time at the moment, but here's the cliff's notes: * 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.
  - artmageddon 10y ago
    > For the rest I do financial modeling (which has a fair overlap with the cryptanalysis from a signal processing perspective anyway) What kind of modeling are you doing with a beast of a setup like that?
    - dsacco 10y ago
      The approach has a few tiers: 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.