Fully functional Bipedal social Robot. / General Chatting / Forums

Forums

4hv.org :: Forums :: General Chatting

« Previous topic | Next topic »

Fully functional Bipedal social Robot.

first previous 2 3 4 5 next

Move Thread

LAN_403

Conundrum

Mon Oct 16 2017, 04:07AM

Registered Member #96 Joined: Thu Feb 09 2006, 05:37PM
Location: CI, Earth
Posts: 4059

More RAM is better.
I considered buying a job lot of GT 710 cards (low end 2GB) then overclocking them with my nice sparkling new 25F16 programmer purchased by SWMBO for a different purpose

(cough hacking digital picture frames for spook season /cough)
These typically use a fairly standard layout and Nvidia even provide an unlocked FW on some high end cards with the advice to ensure external cooling is provided and if needed recycling fans from older cards.
Possibly an alternative is initially set up the AI core on a huge system to work on the BCM chipset the RPi uses to reduce overheads.
Then use 16GB UHS-1 cards to preload learning data on the fly having multiple cores running in parallel synchronized using the system clocks and linked at full HDMI speeds using that interface.
This is feasible due to the way HDMI works, as long as sync is valid you have literally 6 data lines at >300 MHz to play with.

I also tried to get the BCM datasheet but the problem is that pesky NDA.
The graphics "core" has some encoding so purchasing an MPEG4 key might be needed then feed that back into the system to unlock advanced features.
Incidentally discovered that some uSD cards can be overclocked by reading back the serial on the back then (possibly) sending a command back over the bus with this attached to get into service menu.
Its handy to access advanced features so that wear leveling can be gauged as have a failing 256GB here.

Chris Cristini

Mon Oct 16 2017, 11:26AM

Registered Member #1749 Joined: Fri Oct 10 2008, 02:04AM
Location: Claremont New Hampshire
Posts: 497

Oh so you mean build an external supercomputer and preload to the robot that's genius :) so less internal mess I love it. This also means more power due to no size limitations. How ever there could still be a small cluster on board for robot vision and DL and communication back and fourth. And 6 data lines at 300Mhz is quite fast compared. Thank you for your wonderful input.

Another thing is I need to start making YouTube videos where I use my voice and show my face that seems to get the views and YouTube and patreon seem to work hand and hand.
So I made this abomination with TTS lol

I also need to get better at grammar and being social in general. Here is CNC Micros Patreon so you all can get a laugh.

And FPGA?

And some good reading.

Now with the research I have done today I have learned it is not necessary to make a cluster with high end cards so I am thinking of two 1080ti and some good Intel hardware. The price hurts.

Conundrum

Tue Oct 17 2017, 05:49AM

Registered Member #96 Joined: Thu Feb 09 2006, 05:37PM
Location: CI, Earth
Posts: 4059

Yeah, the folks at work think I am crazy heh heh..

To be fair this approach was suggested by Turing himself, back in the day.
IIRC "logarithm tables" used to be calculated by roomfuls of human computers by hand using slide rules and math.
He actually suggested a variant of this method to break codes, where you started with the Short Weather Cipher as a crib then stored the complex variables elsewhere and just did the needed calculations, once the wheel combination had been computed it was fed back into the system and used to obtain plaintext.
(google "solving of Enigma")

I looked into getting a bulk discount from pimoroni and they seemed very interested.
Also Arcadia over here have lots of low end GT710 cards which can be modified quite simply.

Chris Cristini

Wed Oct 18 2017, 01:22AM

Registered Member #1749 Joined: Fri Oct 10 2008, 02:04AM
Location: Claremont New Hampshire
Posts: 497

Well I am going to stop disappointing my self with costs yikes. I will see if I can get the eyes done and move forward on the hardware I think I am going to use a clear resin With the half sphere silicone molds on the way I received some of these WS2812

Led Rings I am going to add and I still need to order the EYE cameras to finish up then its back to figuring out arm servos I may use cheap screw gun Motors/Gearbox and POTs for now I have these PWM motor drivers

already. I just started a Kick starter campaign for this awaiting approval.

Conundrum

Wed Oct 18 2017, 03:34AM

Registered Member #96 Joined: Thu Feb 09 2006, 05:37PM
Location: CI, Earth
Posts: 4059

Have used WS2812s before, don't forget the input resistor!!
Re. 1080Ti, a lot of "broken" cards can be trivially repaired using simple reflow of affected memory chips.
I think when the miners use them often they run hotter than normal so even marginal parts fail under these conditions.

Chris Cristini

Wed Oct 18 2017, 08:01PM

Registered Member #1749 Joined: Fri Oct 10 2008, 02:04AM
Location: Claremont New Hampshire
Posts: 497

Conundrum wrote ...

Have used WS2812s before, don't forget the input resistor!!
Re. 1080Ti, a lot of "broken" cards can be trivially repaired using simple reflow of affected memory chips.
I think when the miners use them often they run hotter than normal so even marginal parts fail under these conditions.

Thank you for the tip on the LEDs. It turns out its not a Kickstarter ready project? OK they say its not complete enough.
What do you think of a pair of GTX 690s

I think it would be an alright choice If I want multiple GPUs at a low cost while having DDR5 and a ton of cuda cores another benefit is 4 GPUs per cheap X16 SLI Motherboard and its not like I am doing human genome research.

I am still very new to research regarding DL so try not to laugh to hard lol

I wonder?

I like your idea of used miner equipment I also wish riser cards they use could be used for DL I seen kits that allow 6 Video cards per MB but that gives up lanes and would slow stuff to a RPI level and waist tons of energy I hope to achieve my goals for less than 2KW I figure decent OPEN CV recognition can be achieved with a PI3 and voice recognition on small embedded systems this may be over kill? but for $1000 or so can make a pretty bad a$$ computer with older hardware.

hen918

Wed Oct 18 2017, 09:08PM

Registered Member #11591 Joined: Wed Mar 20 2013, 08:20PM
Location: UK
Posts: 556

For parallel processing tasks you don't need a very high bandwidth from PC to GPU. A single lane would be enough.

Chris Cristini

Thu Oct 19 2017, 05:21PM

Registered Member #1749 Joined: Fri Oct 10 2008, 02:04AM
Location: Claremont New Hampshire
Posts: 497

hen918 wrote ...

For parallel processing tasks you don't need a very high bandwidth from PC to GPU. A single lane would be enough.

Ah yes I am understanding more and more and reading this

will help make a decision before spending the money.

Conundrum

Fri Oct 20 2017, 05:58AM

Registered Member #96 Joined: Thu Feb 09 2006, 05:37PM
Location: CI, Earth
Posts: 4059

Measure twice, cut once and all that.

My Litz wire showed up, have you considered using tinsel wire for some of the low power interconnects?
I'm pretty sure thats what they used for some of the industrial robots where the wrist is fully mobile.

Chris Cristini

Fri Oct 20 2017, 12:01PM

Registered Member #1749 Joined: Fri Oct 10 2008, 02:04AM
Location: Claremont New Hampshire
Posts: 497

Conundrum wrote ...

Measure twice, cut once and all that.

My Litz wire showed up, have you considered using tinsel wire for some of the low power interconnects?
I'm pretty sure thats what they used for some of the industrial robots where the wrist is fully mobile.

Tinsel wire? That's a good idea. I have worked with tensel wire at Whelen engineering hooking up a spray booth arm it was weird cable because it had some weight to it but flexed more like a rope. I could source tons of it from phone cable as twisted pair that could be good if I actually decided to go with small switches in the fingers to simulate touch?

first previous 2 3 4 5 next

Moderator(s): Chris Russell, Noelle, Alex, Tesladownunder, Dave Marshall, Dave Billington, Bjørn, Steve Conner, Wolfram, Kizmo, Mads Barnkob