Supercomputing adventures with a MegaST/4

[Badwolf]

Ye flippin' gods! Are you planning to bitcoin mine and jumpstart your car at the same time?

One problem with this is the connector is a big unobtanium thing with metal teeth

My immediate thought was spade connectors, but I can't quite get a handle on the dimensions.

Let's hope your backplane doesn't pull so much power the low temp solder starts to melt.

BW

[dml]

Ye flippin' gods! Are you planning to bitcoin mine and jumpstart your car at the same time?

I probably don't need to worry about dropouts

My immediate thought was spade connectors, but I can't quite get a handle on the dimensions.

You're right of course - it would be better if I can find the correct spades to fit but the ones I had seem to thin.

Let's hope your backplane doesn't pull so much power the low temp solder starts to melt.

It did cross my mind but I checked the conductivity etc. for this one isn't worse than tin-lead so it should be ok with the 3 cables so long as the joints are good. The main risk with it is tearing since it is quite soft and the cables are heavy.

Good enough for now but will try to do something a bit more permanent later.


The gameplan for this thing has been:

- figure it out
- try to make it work with one TRAM
- try to make it work with two or more
- give it enough power for a full card
- fill the card and make sure the whole thing can be addressed and works

next part should be:

- make code run on it
- escalate things in all directions
- render some Babylon 5 on it

[Cyprian]

nice progress.
I wonder what the actual power increase will be with this number of transputers, e.g. in XPtr.

[dml]

nice progress.
I wonder what the actual power increase will be with this number of transputers, e.g. in XPtr.

I am thinking that for expensive scenes (most of the ones I have provided are expensive! its all floating-point in C) the time spent in each rendered tile will dominate over transferring tiles over the links so the gain should be nearly linear with the number of chips. The gain will drop a little with each one added but still should be close to linear.

For quick-rendering things it will probably be much worse as the requests/replies for each tile and the tile pixels will be pass-the-parcel up and down the chain and the first few transputers in the chain will be 'very busy' trying to serve all of that traffic. The links have background DMA engines but they are still only 20mbits/sec and processes block to read/write if the link is busy or nothing is ready yet.

So the simple chain-of-pairs topology I am using here is not great for realtime/low-latency stuff but I expect it is fine for slower work like raytracing.


However by rewiring links on a backplane you can make better networks with less bottlenecking, usually at the expense of writing more complicated code to deal with it. e.g. you could use free links on the first Transputer to fork into 2 or 3 parallel chains instead of one (and have the first T do primarily routing/scheduling/caching). More optimal arrangements include grids/planes, cubes and hypercubes o_O where there is a known maxmimum link distance between any two nodes on any axis - you then need to write software which knows how to route stuff across the optimal free path at any moment.

A downside with Transputer networks - one of the reasons they went out of favour - is the need to write (or at least compile) code which knows what the topology is. While you can link them any way you like, the code being uploaded doesn't organise itself automatically onto free resources - you need to design around the arrangement you have set up and tell the compiler/configurator what it looks like so it makes the appropriate bootloader.

[rubber_jonnie]

Can't wait to see some B5 scenes, I am a fan

10 transputers though, that is just incredible, well done sir.

[Darklord]

Yeah, Babylon 5 rocks - still go back every couple of years and watch
it, end-to-end.

[SteveBagley]

Yeah, Babylon 5 rocks - still go back every couple of years and watch
it, end-to-end.

I found a 'Making of' video on YouTUBE once and you could see a MegaST keyboard in Christophe Franke's studio -- at 15:25… Of course, by half way through season one they'd stopped using Amigas anyway and were running Lightwave on NT, which begs the question was the ST being used on Babylon 5 longer than the Amiga was

Steven

[dml]

... and you could see a MegaST keyboard in Christophe Franke's studio -- at 15:25…

Ha! that was a nice find and good point on the ST probably surviving a bit longer

[Darklord]

I found a 'Making of' video on YouTUBE once and you could see a MegaST keyboard in Christophe Franke's studio -- at 15:25… Of course, by half way through season one they'd stopped using Amigas anyway and were running Lightwave on NT, which begs the question was the ST being used on Babylon 5 longer than the Amiga was

Steven

Nice. Now, go drop that as a conversational "tidbit" at your local Amiga users meetings...

[SteveBagley]

I found a 'Making of' video on YouTUBE once and you could see a MegaST keyboard in Christophe Franke's studio -- at 15:25… Of course, by half way through season one they'd stopped using Amigas anyway and were running Lightwave on NT, which begs the question was the ST being used on Babylon 5 longer than the Amiga was

Steven

Nice. Now, go drop that as a conversational "tidbit" at your local Amiga users meetings...

Hah -- I'd quite like to live… Although we have chatted about doing a Computerphile video on the Toaster (which I suggested we call 'Why the Video Toaster failed (in Europe)'


As far as I can tell, It's all a bit of a urban myth anyway, yes Lightwave was used to render the graphics for B5, but the actual images were written out to be loaded into a professional piece of video equipment (an Abekas disk recorder iirc, I think it was an Abekas model) via an 8mm drive on the Amiga, and preview was done using the Toaster's NTSC framebuffer (judging by that video) so the Amiga's graphics and video hardware isn't actually being used.

It'd be fun to port early Lightwave to the ST though and see how the rendering speed compares on a TT (or 50MHz PAK/3) -- or Doug's transputers to bring it back on topic.

Steven