I'm going to be doing some 3D modelling in Blender and I'll be needing a new computer for it, my laptop just can't cut it, I also want to do some pretty heavy number crunching so I'm going to need as much RAM and processing power as possible.
What's recommended for a budget of $3000 AUS?
I don't mind paying more if the extra is worth it.
For my CPU the AMD FX-8350 seems a pretty obvious choice, now I'm looking for a motherboard with dogballs of memory to make the most of it, and some good ports, a cluster has to start somewhere.
Regarding GPUs, seems frivolous to spend a lot on one, then again it's exactly that thought that came back to bite me in the ass, the whole reason I'm having to build a new computer.
Regarding GPUs, seems frivolous to spend a lot on one, then again it's exactly that thought that came back to bite me in the ass, the whole reason I'm having to build a new computer.
The immediate future in computing IS in GPU's. If you're doing Blender work you definitely want a beefy GPU, or two. I'd consider a new Mac Pro which is a custom made machine for the work you want to do, if you can wait a few months. Just make sure Blender supports OpenCL (it supports CUDA but not clear about OpenCL).
You will need: a huge motherboard or two, big sheets of sturdy plastic, the biggest PSU that you can find, some fine metal grating, and a small ceiling fan.
Step 1) Clean your room, specifically the floor. You will need as much space as you can find.
Step 2) Lay a big sheet of anti-static plastic wrap on the floor--cover as much as you can--and hold the wrap down with some heavy things.
Step 3) Cut a huge sheet of plastic and lay the motherboards down in the middle.
Step 4) Lay the PSU down beside the motherboards (five centimeters away at least) and connect it to them.
Step 5) Drop in as much computing power--GPUs, CPUs, RAM--as will fit in each slot.
Step 6) Nail plastic side walls down onto the plastic baseplate, leaving enough space for the fan to spin.
Step 7) Attach the fan to the metal grating as if the grating were a roof and wire it to the PSU or wall outlet.
Step 8) Nail the 'roof' down onto the side walls (or use a hinge).
The result will be a huge, loud, expensive, hideous contraption, but one that is easy to build and expand and will never break down, overheat, or underperform. Your electric bill will explode, though; I recommend installing some solar panels on your balcony walls to power your newborn monstrosity. Also recommended:
--Putting soundproofing foam around the inner walls.
--Cutting slats in the baseplate and putting the whole thing in your A/C duct to improve cooling and better hide it.
--Failing the above idea, putting a nice case around it or a paintjob on it.
Looks like they turned off ATI support due to some instability issues, but it can be turned on. This thread shows some info and the merits of GPU rendering.
Wait to see what the new Mac Pro costs as that is tailor made for what you want. Otherwise I'd get a chimped out CPU (nothing fancy i7) and the Titan, or probably 690 or 680 would be better.
Duxwing, you are awesome
My first thought upon reading the parts list.
Architect, the Blender stuff is only short term, basically I need to 3D print some robot parts and I need some more artistic freedom than TurboCad gives me (specifically in 2D because my laptop sucks balls, heats 'em up too) also I can download prefabbed models which is a huge boon for someone "artistically challenged" like myself. After that I'm not going to be using this as a gaming rig or to render anything, it'll be the number cruncher for an artificial neural net methodology I've been working towards implementing since, forever. It isn't all that complex really, just addition, subtraction and division, but how well it's going to work will depend upon how much memory and processing power I can get my hands on, I'm not even sure there's a theoretical upper limit to what I could make use of.
There is however a very real limit to my budget, especially since y'know I'm building a goddamn robot, although I am building my own servos now which is saving me thousands of dollars, and having the bot's frame and casing printed as one is going to save me hundreds more.
It's going to empty my bank account but I might just actually do this.
Or I will empty my account and spend another year saving up -.-
Architect, the Blender stuff is only short term, basically I need to 3D print some robot parts and I need some more artistic freedom than TurboCad gives me (specifically in 2D because my laptop sucks balls, heats 'em up too) also I can download prefabbed models which is a huge boon for someone "artistically challenged" like myself. After that I'm not going to be using this as a gaming rig or to render anything, it'll be the number cruncher for an artificial neural net methodology I've been working towards implementing since, forever. It isn't all that complex really, just addition, subtraction and division, but how well it's going to work will depend upon how much memory and processing power I can get my hands on, I'm not even sure there's a theoretical upper limit to what I could make use of.
There is however a very real limit to my budget, especially since y'know I'm building a goddamn robot, although I am building my own servos now which is saving me thousands of dollars, and having the bot's frame and casing printed as one is going to save me hundreds more.
It's going to empty my bank account but I might just actually do this.
Or I will empty my account and spend another year saving up -.-
The beastie can be expanded into as much space as you have available and built to arbitrary specifications: just add more motherboards, PSUs, and ceiling fans. In fact, you can build the entire thing in cells of one fully-loaded motherboard, one ceiling fan, and one PSU. Each cell would have an incoming and outgoing: air duct connected to a central system of ductwork capable of running without any cells in it (use an empty cell case to keep the air flowing if one is removed) positive and negative electrical terminal for powering the cells with a massive parallel circuit instead of a tangle of cables, and data in and data out for processing and routing by a manager cell. With a little bit of cleverness, you could thereby build not only laterally, but vertically, stacking cell atop cell for massive power.
Of course, if you can dedicate an entire room, then ditch the cases and fans, put your cells on bookshelves, and use your central A/C to cool everything. Upgrade your air filter to keep dust out and cut huge slats in the backs of the bookshelves (use the leftover material to put little walls on the other side so that the cells stay in) to improve cooling.
-Duxwing
PS Holy cow, now I understand why you use comma splices. Ne visions are so clear, but they just keep coming!
To clarify the AI's program only needs to cycle once or twice a second, the whole AI is an analogue/digital hybrid, the digitally simulated neural net does recognition & reasoning while the analogue handles things like balance, grip strength, power management, etc.
Ideally the whole thing would be analogue, like we are, but if you get into analogue robotics you'll see that, well, as the number of nodes increase the number of potential connections between them increases exponentially, and when there's hundreds if not thousands of nodes, well manufacturing an astronomical number of relays is an astronomical challenge.
On the other side of things simulating an analogue process in digital involves a huge loss of fidelity, I have no perfect solution to that, one idea is specific modules for specific senses, but then there's that weird separation affect, imagine seeing things without actually knowing what they look like. Or maybe going the abstraction route, lets say you can only see black and white but you can still tell what colour something is by how quickly it polarises back and forth, I suppose to an AI with vision like this the idea of colours being wavelengths of light would seem the most obvious thing in the world.
I've been thinking about this stuff obsessively since high school.
The rabbit hole goes deep
The digital processing power is all off board, the robot is really just a puppet for the AI, it could switch between several, perhaps even control several at once if I set it up so the AI can reprogram the microcontrollers and delegate procedures to them, for example if the analogue side of things is handling the walking motion and balance then a microcontroller needs only redirect it from one waypoint to the next, while the AI itself supervises.
You will need: a huge motherboard or two, big sheets of sturdy plastic, the biggest PSU that you can find, some fine metal grating, and a small ceiling fan.
Step 1) Clean your room, specifically the floor. You will need as much space as you can find.
Step 2) Lay a big sheet of anti-static plastic wrap on the floor--cover as much as you can--and hold the wrap down with some heavy things.
Step 3) Cut a huge sheet of plastic and lay the motherboards down in the middle.
Step 4) Lay the PSU down beside the motherboards (five centimeters away at least) and connect it to them.
Step 5) Drop in as much computing power--GPUs, CPUs, RAM--as will fit in each slot.
Step 6) Nail plastic side walls down onto the plastic baseplate, leaving enough space for the fan to spin.
Step 7) Attach the fan to the metal grating as if the grating were a roof and wire it to the PSU or wall outlet.
Step 8) Nail the 'roof' down onto the side walls (or use a hinge).
The result will be a huge, loud, expensive, hideous contraption, but one that is easy to build and expand and will never break down, overheat, or underperform. Your electric bill will explode, though; I recommend installing some solar panels on your balcony walls to power your newborn monstrosity. Also recommended:
--Putting soundproofing foam around the inner walls.
--Cutting slats in the baseplate and putting the whole thing in your A/C duct to improve cooling and better hide it.
--Failing the above idea, putting a nice case around it or a paintjob on it.
The beastie can be expanded into as much space as you have available and built to arbitrary specifications: just add more motherboards, PSUs, and ceiling fans. In fact, you can build the entire thing in cells of one fully-loaded motherboard, one ceiling fan, and one PSU. Each cell would have an incoming and outgoing: air duct connected to a central system of ductwork capable of running without any cells in it (use an empty cell case to keep the air flowing if one is removed) positive and negative electrical terminal for powering the cells with a massive parallel circuit instead of a tangle of cables, and data in and data out for processing and routing by a manager cell. With a little bit of cleverness, you could thereby build not only laterally, but vertically, stacking cell atop cell for massive power.
Of course, if you can dedicate an entire room, then ditch the cases and fans, put your cells on bookshelves, and use your central A/C to cool everything. Upgrade your air filter to keep dust out and cut huge slats in the backs of the bookshelves (use the leftover material to put little walls on the other side so that the cells stay in) to improve cooling.
To clarify the AI's program only needs to cycle once or twice a second, the whole AI is an analogue/digital hybrid, the digitally simulated neural net does recognition & reasoning while the analogue handles things like balance, grip strength, power management, etc.
Ideally the whole thing would be analogue, like we are, but if you get into analogue robotics you'll see that, well, as the number of nodes increase the number of potential connections between them increases exponentially, and when there's hundreds if not thousands of nodes, well manufacturing an astronomical number of relays is an astronomical challenge.
On the other side of things simulating an analogue process in digital involves a huge loss of fidelity, I have no perfect solution to that, one idea is specific modules for specific senses, but then there's that weird separation affect, imagine seeing things without actually knowing what they look like.
You're encountering a problem because the five senses themselves are each composed of three parts--sensors, processing, and integration--a fact that means that you don't need five nodes to model the senses: you need fifteen. Moreover, you can organize these nodes such that you don't need 15! connections.
Or maybe going the abstraction route, lets say you can only see black and white but you can still tell what colour something is by how quickly it polarises back and forth, I suppose to an AI with vision like this the idea of colours being wavelengths of light would seem the most obvious thing in the world.
Inward Path
The node that handles visual sensation simply transforms the state of each sensor on the robot's cameras into an array of hex codes (e.g., red is 036) that is then sent to a second node that 'looks' at the picture by checking it against stored 'known' patterns and looks for new ones by checking it against the X most recent number of pictures. The third node selects the Y most important patterns, transforms them into something comprehensible to the 'decision engine' of the robot, and then sends them there.
The system architecture (I now understand why Archie loves his job!) would be as follows:
Four connections. 3! = 6, so we've come in under budget by a great amount.
Outward Path
The decision engine can also decide to look for something, so it can tell the translation node to change its priorities, the sensors to scan, or the sensor node to change its arraying system.
3! = 6 > 3, so we've come in under budget by half!
The same logic applies to all other concerned senses for a grand total of 5(4+3) or 35 connections--a mote of dust compared to 15! (1.3076744e+12--a number so huge that even Google Calculator overflowed).
I've been thinking about this stuff obsessively since high school.
The rabbit hole goes deep
I hope that I've helped! A little Ti goes a long way.
The digital processing power is all off board, the robot is really just a puppet for the AI, it could switch between several, perhaps even control several at once if I set it up so the AI can reprogram the microcontrollers and delegate procedures to them, for example if the analogue side of things is handling the walking motion and balance then a microcontroller needs only redirect it from one waypoint to the next, while the AI itself supervises.
+3
You should definitely be going with something like this.
EDIT: Specifically, you should be going with intel in my opinion. Also, if you don't need a really powerful GPU, opt for a cheaper graphics card (although, as is already clear, don't go cheap), and also perhaps a better processor.
Overclocking is of course a MUST, even if it is only for modest (25%) gains, just make sure you get a non-stock cooling solution (50-100$).
More threads (as far as multi threading goes) aren't that important, because you can make up for them by overclocking more. In general, i7 processors are higher spec with a bit more raw power (I'm not sure if this is related to die size, I.e. transistors), and also have more cache. What you take away is that, according to benchmarks, similarly clocked i5s underperform similarly clocked i7s, but not by far.
More threads (as far as multi threading goes) aren't that important, because you can make up for them by overclocking more. In general, i7 processors are higher spec with a bit more raw power (I'm not sure if this is related to die size, I.e. transistors), and also have more cache. What you take away is that, according to benchmarks, similarly clocked i5s underperform similarly clocked i7s, but not by far.
The difference in price between i7 and i5 in the same tier is minimal ($100 on Newegg between the 3770k and 3570k). I'd say even a modest performance benefit is worth it, especially if most of the budget is reserved for graphics solutions. Gotta avoid those bottlenecks!
This site uses cookies to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies. We have no personalisation nor analytics --- especially no Google.
