Ai Dreams Forum
Software & Hardware => General Hardware Talk => Topic started by: LOCKSUIT on March 09, 2018, 12:49:08 pm
-
Keeping this simple:
About 1.8 years ago I built my PC. It's a large tower. My laptop in the kitchen is less powerful but still stark and I could fit 10 of them in my tower easily winning in power yet same size.
I know you can construct a supercomputer server stack or do whatever korrelan did (?) and this is very powerful yet as small as it can be. But. I don't know how. Right now I have a Large Tower - it is stronger than an avergae pc, but in the future I would like something say 5 times more powerful (keep in mind new tech) yet not 5 times as big as my Large Tower. In other words how do you go athletic powerhorse yet keep the size of my Large Tower? Tell me da way of the terminator secrets. Is it stacking 3 xeon severs on your desk???
-
From my experience of running machine learning stuff is gpus are the cheapest way to give yourself an obscene amount of cores, but they have limiting factors - theres not much memory and reads and writes to and from it are slower. And sli only gives more cores, doesnt double up the memory.
And if your doing something really seriously, (but u have to understand the HARDWARE videos, not just the software) actually printing out your own custom brain for it isnt too much to ask for. if you have a micron density accuracy on a 3d printer, over a cm^3 you can get ~100 gigs of ram. doesnt that sound smart - even if its harder? the printer to finish in a week the robobrain has to plot 100k "voxels" in a second. its a mega highway in cm^3 - takes AGES to plot it. most ppl with 3d printers probably would say it was too hard - but im not that type of guy.
-
Build your best tower and then tie them connect it with Gigabit Ethernet to last years build:
https://en.wikipedia.org/wiki/Gigabit_Ethernet
Or you could cluster a bunch of PC together. to make your own super computer:
https://en.wikipedia.org/wiki/Computer_cluster
The super computer of today are cluster computers of highly trick out pc with highly optimized software. And also use PCIe bus to connect to
other pc in the cluster.
-
If I'm constructing a NLP AGI, and am going to go deep, what should I better have? And is CPU better for LSTM?
-
GPU for the heavy lifting. Cpu for the light work and multitasking with other mother boards.
-
But NLP AGI is sequentially demanding no?
And will a LSTM run faster on GPU or CPU?
-
For serious applications we never lay hands on the hardware ourselves. We lease server systems from places like Softlayer with dozens of CPUs, hundreds of gigabytes of RAM, terabytes of SSD, and petabytes of disk space. For some tasks we need the whole thing, but mostly we split each one up into dozens of different virtual machines with a few cores dedicated to each one and then allocate them to specific functions. You'd be amazed what you can get for a few thousand dollars a month.
Nowadays the trend is towards just going to the cloud. You can set up instances on Amazon Web Services or Google Compute Engine that start out free but accumulate cost as you use more resources. Note that you can use them for free for quite a while if you are very careful, but it is easy to start running up a bill so be very cautious if you try this. This is the way to go if you want unlimited power.
I just upgraded my personal server with another 4 terabytes of SSD and while I was shopping I had a look around at options for upgrading the rest. I was delighted to find that for five thousand dollars I could put together a new machine with 16 cores and 128GB of RAM. Maybe next month. My research is all in the symbolic domain so I don't have a lot of use for GPUs and I didn't price those this time round.
-
CPU starts up and runs main operating system like Windows, Linux, or Apple.
Mains CPU start and may stop other CPU and GPU programs. CPU is the master organizer of other CPU and GPU programs.
When CPU does not have the muscle it uses GPU power. ALL deep LSTMs, Deep NN, RNN, and GANs will need GPU power.
And also game with good graphics, Video editing software, and CGI software.
-
Oh ya the cloud. Just use Google's supercomputer. Only possible issue is algorithm running on the net haha...
What do you think about that?
-
Locksuit, as far as Google is concerned, your porn profile (#) is more valuable than any ideas that you might have. As long as you can maintain the illusion that you're a raving nutter, nobody will suspect you of having anything worth stealing, so you're safe for now. If you're serious, you should be more worried about your creation escaping into the wild and taking over the world than anyone knowing what you're doing.
(#) Ex Machina: who doesn't love that movie :D
-
https://cloud.google.com/security/
https://cloud.google.com/security/whitepaper
https://cloud.google.com/security/security-design/
https://cloud.google.com/security/encryption-in-transit/application-layer-transport-security/
https://support.google.com/cloud/answer/6262505?rd=1
-
What a beautiful site.
I will definitely bookmark and note this site down as top priority.
-
What a beautiful site. I will definitely bookmark and note this site down as top priority.
That's great, I'm glad you think it will be useful. Just be very careful not to run up any bills that you can't afford to pay. I'm pretty sure you'd get lots of warning emails if you were starting to get out of your depth, but when you're dealing with hardware that powerful it can get out of control scary fast.
-
Don't worry, I have so much time to think and focus. It shouldn't be a problem for me.
-
As long as you can maintain the illusion that you're a raving nutter, nobody will suspect you of having anything worth stealing,
haha made my day.
-
Don't forget that I'm only 22.6. I can't know that much!
Surely any Google Brain computer scientist with big paychecks knows way more than I can learn!
-
I think guys like us is all their is, big paychecks is just for show. the less money you have the more you DIY, who gains more?
-
We the Power Rangers do.
-
ay locksuit - im developing a mental picture of you - have u got black hair?
-
Thou shalt not depict nor idol any symbols of thy father.
-
AGI need a body. Raspberry Pi 3 to do physical interface with the world in real time. Raspberry for touch , for motor control, for encoder reading,
for temperature reading, and for smell. All raspberry are connected to small laptop for sight and all audio. During sleep cycles AGi is trained
on a deep NN on recorded video data when AGI is was awake.
-
Did yous know that vision is symbolic?
The only thing that has meaning is the body.
Because the body interacts with physics - by rewards.
That's why we love to relax and eat popcorn and save lives.
-
Apparently Google Complex can't even build straight buildings:
Evidence:
https://en.wikipedia.org/wiki/Google#/media/File:Googleplex-Patio-Aug-2014.JPG
And all they do is have fun look:
https://www.youtube.com/watch?v=8sOtjBDPQdU
Just joking. Google & AI is everything today.
-
Did yous know that vision is symbolic?
The only thing that has meaning is the body.
Because the body interacts with physics - by rewards.
That's why we love to relax and eat popcorn and save lives.
all ai is symbolic, the phrases that appear in a neural network could be considered symbols each.
-
The AI field might start calling us the Power Rangers.
I always liked how when the evil tech-creatures are defeated, they grow larger.
Here is some powerful moments of up-sizes and laser beam/sphere blasts.
3.33
5:53
https://www.youtube.com/watch?v=3xF1zubfpds
5:18
7:21
https://www.youtube.com/watch?v=cwKC8ZOJ7wc
36:01
https://www.youtube.com/watch?v=9My93C_6ysg
-
Also 15:16 that is like how fast the programmable-bio-nanobot-wireless-fog-brain-creatures will move. Self healing too.
https://www.youtube.com/watch?v=9My93C_6ysg
-
Wow Google Cloud Compute it's only 1.46USD for 1 NVIDIA® Tesla® P100 for an hour. 5.84USD for 4 for an hour.
-
your better off with a geforce, i think the tesla cards arent even faster, they are just a ripoff.
-
Do I have to be in ex. Iowa to use Google Compute wtf guys help...I tried looking it up and can't find much answer...
but it's a 16,000 dollar card bro....
Besides...I can use multiple processors from their cloud instead of having to maintain some monster computer.
I can (if I'm correct) use ex. 32 of their super-expensive GPUs/CPUs.
They also keep there stuff up to date meaning I can come back after some usage and be on a better processor! Instead of losing ALL that money!
In the pc I built I put in a 1060 6GB Nividea GeForce.....and my next computer (depending on my estimate of what my algorithm would require) I just may put in a better one and might be able to just use my home brew pc for my NLP algorithm.
Other than that, there's not much point I think in investing money into a Beowulf monster that may break down and get outdated. I'll wield a home pc, and use Google's supercomputer superstorage supermachinelearningdatabase facility.
Can't you see the future? Humans weild a small PC. And when they need power they tap into the big Google facility CLOUD it makes so much sense.
-
Lock,
Since you don't Program and don't use any formulas or algorithms for any purpose, what would be the other purpose for you wanting to build a "Power-house" computer? Deep Neural Networks rely on lots of number crunching and heavy math. Just trying to figure your logic here.
Surfing the web and visiting YouTube or other such sites doesn't require a powerful computer. I can do that with my Tablet.
Just saying...why throw a goodly sum of money just to have the latest, greatest thing in computing only to have it become outdated by the time you Unbox it at home?
-
Haha, nope, I don't code, or do math. (I'll code one day when a super-easy-visual-PL is released). For now I use my buddy to do the code/math. He also knows physics and AI quite well. I can concentrate more on the mechanism.
Art...that's exactly what I said...I'll own a normal-ish pc and use Google's Cloud to do anything big. - I never have to maintain/upgrade to the latest....I'll simply use Google's many expensive new updated processors/storage/etc. ~ I'm so excited over Google Cloud.
-
A cluster is only useful for certain specific types of AI modelling… not for the novice or faint hearted lol. You have to be running huge simulations before the network overhead/ lag times become irrelevant.
I agree with Art, just keep a decent tower/ desktop machine with a fast internet connection. When AGI tech becomes generally available it will require a huge amount of resources, more than the average user can supply… just have a decent terminal ready… that includes a decent microphone, HD web cam etc.
:)
-
Gee I don't have a microphone or cam on my tower...Next build !
:)
-
I'm curious about your buddy. Assuming this isn't just another imaginary friend, it would be great if they could sign on here and talk to us directly. We could all exchange insights faster that way.
-
Did you see my simulated spiderTable learn to crawl? Or my simulated super-realistic baby randomly jiggle? He coded those by my instructions.
However I halted further work on the baby because I concluded by strong evidences that a body isn't needed, against my previous proposition that a body was definitely needed.
-
Yes I remember those demonstrations but I don't ever recall you mentioning an associate before. Maybe you did and I missed it, I'm not always paying attention.
-
Last I recall, Lock was looking to spend some coins to hire a programmer. If you indeed have a "friend" that can do it, so much the better!