The T6 Viz Challenge: Final Boss - T2sday Update 3175

Hey everybody, it's T Tuesday. I hope you had a good month. Uh, guess what happened to me on Sunday? Focus is terrible. Is it Is it This might be working. Looks a little better. [ __ ] Hey. Sunday evening. Uh, heading for the, uh, T Tuesday update. It's a start. It's been a long ways. All right. Now we're going to try again with uh Drake. Whoa. Yeah, that's more like it. Okay, cool. Step by step. So, that's all brand new. We got tons to talk about. I have no idea how long this is going to take, so I'm just going to go into it. If you're just joining us, and we have a bunch of new subscribers, welcome. Thank you. uh uh and I think a lot of it is because there's building interest out in the nerd world in the tens torrent architecture chips machine systems and so forth. Uh uh and so the what I've been doing for the last year or so uh um almost a year since last summer is trying to uh build pretty much from bare metal pretty lowlevel uh start uh to build an implementation of the Mughal feast machine stocastic cellular automa to run on the black hole tends to quiet box which is sitting outside uh uh blasting off its heat and with the progress that we've got to report today. We're going to have a live demo in just a few minutes. Now, if all goes well, uh I think I can probably start to say, you know, probably we shall see. All right. Uh uh so these were the goals that I was supposed to get to. I was really hoping to have uh all green, but I do not. uh the cleaning up the code cuz the code's been a terrible mess. And in fact, the code is in much huger mess right now cuz I was fighting bugs, nasty, nasty bugs that I had no freaking clue uh for an awful long time until the very last minute. But yes, I got the new thing that I talked about last time, the t the tag the task tag table. I got enough experience with it that I could start to figure it out and so forth. Real host viz running. That's the main event. Pop yaks. Yeah, it was nice. It was nice. So, here's what I'm going to try to talk to you today. I have no idea how much time it's going to take. I'm going to try to keep it to 20. We'll see how it goes. Uh uh. All right. So, the 10's torrent architecture is a spatial grid of these little tensics tiles. Each thing that has a T at the beginning, T0, T39, that represents a Tensix tile, which has a significant amount of uh hardware. In each Tensix tile, there are five hardware threads. Uh there are 140 Tensix tiles per black hole card, at least if you have the old T black hole cards like I do. The new ones uh all have just 120. And then the box that I've got has four of these whole things in one box. And each of those tiles is connected by two knocks, two network on chips. NOC knock zero routes uh west and south and knock one routes east and north. So you can pick which direction you want to go depending on your destination. So months ago I made the decision that I wanted to cuz the the tensics tiles, you know, the for what I'm using them for once again, you know, we'll see the um I wanted to to batch up batch them up together. So I came up with this regional hub concept where I was going to take a 3x3 grid of Tensix tiles and have them work together to all so eight of them would do little cellular automa event transitions and one of them would be the hub that they would all send things to. And so here is a case where this northwest uh tile has sent back an event window which is going to this little memory up here that remembers them all. And then the T6 viz challenge is how do we get them off of the roof of this building and send them someplace that we can actually see them. Now uh there turned out to be a slight problem with this architecture. Uh uh so what about say this northeast tile here, right? Well, so the NOX zero goes west and south. Well, we don't want to go west. Uh I'm sorry, it goes east and south. Uh, and we don't want to go east because we want to go west and we don't want to go north. So for the northeast tile, any way you slice it, whether you go on knock zero or knock one, you're going to be routing all the way around the chip in order to get to the thing that's right next to you. Uh, so same thing for the southwest tile. you it's just, you know, you can do it, but you have to deliver the packet and then turn it around and switch it on to the other knock. So, it costs a lot. Uh um so what am I going to do about this? Well, uh this is what I did about it. [laughter] I went from a 3x3 regional hub to a 2x two regional hub. So, the problematic ones, the northwest, the northeast, and the southwest, they don't come up. So to get from here to any of these three event window processors and back is all on knock zero or on knock one depending on which way we're going. Now here's a little puzzle for the folks in the in the chat room right now. Uh uh you know how could we get the 3x3 regions back again without having this problem of having to send packets all the way around the world. So this is how it looks in the uh in the declaration file. It's a 2x two cell now uh that is filled with EWP that's event window processors and then at 1 one which is you know 0 0 up here and one one here uh is a hub image and off we go. So that's the architecture that we're working with. The T6 viz challenge is how do we get the events once they've happened they're only stored in these regional hubs all over the place. There's no screen on that thing. So, yes, the problem is we have to take the data off of the 106 tile, off the ones that are being the hub, and get it to the host somewhere which has an internet [clears throat] connection. Uh uh the T6 tiles have no screens. Unlike our beloved T2 tiles, the purpose, the name of this channel, the T2 tile project, was we were building these. Each one of these has a screen on it. So once it decided what things were going to happen, it could display it right there and we took pictures of it to figure out what the entire grid was doing. So [snorts] a little more about the architecture. Each black hole chip, the quiet box got four of them, uh, uh, has a PCIe tile. There's two of them, PCIe0 and PCI1. On the ones I've got, the only one that's used is PCI0. As far as I can tell, no black hole chip uses both of them. It just depends on I don't know what. So every bite that wants to go from anywhere on this entire array of 106 tiles has to go through PCIe0. So for example, here is a little 2x two hub uh uh that we just like we saw and in the lower right is the hub part of it. The other three are the event window processors and it has to route straight up to get to the PCIe. So, it's pretty easy as far as that guy is concerned. But it's really like this. There's this incredible hodgepodge of all of these 2x two regions. Each one of them has a hub that has to send data a significant pile of data to the host so that it can be rendered into an image and sent out to wherever as well as being saved. So the plan is, and I'm not going to go through this, you know, make a 2D spatial list of the atoms that's changed. So we don't have to send the whole uh grid, the whole regional hub grid. We're just going to send the atoms that have changed. We're going to compress them. Uh dedicating an entire one of these little hearts, of which you have five on each tile. We're dedicating one on the hub just to compressing the atom changes down. The compression is not great, but it's about 50%. And at this point, I'll take it. And once they've been compressed, send them through the knot along with everything else that's going there uh to the host. And the host then has a software thread uh dedicating to dedicated to decompressing the changes it's getting from each hub. And it then figures out well given that it's an atom of drag given it's an atom of Facebook atom of fork bomb FB um figure out what color it should be and draw it directly on a array uh that represents an image and then just ship that image regularly regardless of how many changes manage to come in. There's no big synchronization stuff here at all. So let's uh what button do I want? I've completely forgotten. Huh? Right. We want uh that this one. Okay. So, this is the windows that I've got that are connected to the black hole. The what I want to do first is just uh Whoops. Um All right. There we go. I'm going to touch the built not new file. Uh uh which is the trigger that I'm now using. uh because the to build the system takes quite a while uh um and then I have to come and start the thing and then for it to set up it's it's setting up right now but it takes a minute and something uh uh in order to get all the code downloaded and get everything all running. So I used to have to uh do the build and that would take a couple minutes so I would go away and then I would have to come back and start it and that would take a couple of minutes and I go away and come back and it was terrible. So I made it so that when the build se succeeds successfully it it touches it makes a little file that this other script is looking for and whenever the script sees that file it starts up the program. So, I touched it here uh uh to get it going. And now we can go over to the uh Okay, if all goes well, in maybe 20 seconds, something like that, the video stream will pick up. At the moment, it's just routing within the local network here, but you know that there's it's it's internet. It's TCP. Uh um the the live stream will pick up and we'll see what we can see. Uh it works. Uh there it is. Uh uh okay. So that is a fork bomb. Uh so what this current demo is doing is for each of the little hubs, each of the regions uh in the thing in the in the first uh in the upper left in the northwest because that's what corresponds to black hole card number zero. Uh let's make it a little bigger. Uh okay. Uh um and we we can zoom in on it. All right. So these things that you see here that are look like rectangles, those are fork bombs. The maximum fork bomb that spread as fast as they can. The ones that still look like uh blobs, circles, those are the minimal fork bombs that just pick one random neighbor and try to change it to a fork bomb. So they spread much more slowly, but just as relentlessly. And you may be able to see it's a little bit difficult. I try to get the pointer in, but we'll go back to it. uh a couple of things that are just looking like a little sprinkles like we saw in the opening uh little clip. Uh those is Dreg and Rez. One of the earliest earliest uh elements that I ever explored when developing the uh mobile fee machine. And now um uh Whoops. Yow. Uh okay. And so now what has happened is now the second put me back in the uh second black hole card which over here has started initialized. The third black hole card is initializing and finally uh they're going to start picking up over here. And the only reason it's taking that long is because this initialization sequence took a big chunk of time to do black hole card zero and then it went on to black hole card one and two and three. And so these delays in things appearing are just corresponding to the delays that are happening there. Uh uh but you know we can we can zoom in. Oops. Only if we There we go. Okay. So just to pick a little bit in the center. Uh um can I actually do this? I guess I cannot. Okay. No. Um the uh there's one popping. So, we've got the drag, which is the brown background with the yellow. There's that's all that or the greenish stuff. And all those little green sprinkles, those are actually resource atoms. Rez the drag. They're very rare and they're uh they're red. So, you might or may not might or may not be able to see them. It's not clear. Let's go back to the uh full view. Okay. So, at this point, uh, pretty much all of the tiles, and this shows you all the tiles have been seated with something, either FB1 for the slow growth, FB4 for the super fast growth, or drag and res, and they're all just crumping along. I've had these things running now for many hours, and it it's all working. So again, is this what the cellular automa is supposed to look like in the end? No. Uh what's going on here is that each of those little regions, those 2x two groups of tensics tiles, uh uh cannot uh talk to any of their other neighbors. They have events within it, but the interhub communication is still missing. So that's where we are now. Okay. Well, so display isolated hubs. This was a long time coming. I thought we were going to see this demo three months ago. We're seeing it today. Uh better late than never. Okay. So, what's next? I mean, this has been a 100 days or so uh trying to do it. So, what's coming up next is InterHub Communications, and that has a whole rat's nest of possibilities. There's at least two major parts to it. Number one is having the little uh regions on the same black hole card on the same black hole chip be able to share with each other. So something that was an atom that was born on one hub might be able to migrate to another hub and have a exciting life there or whatever. Uh um but then there's the next level which is uh a thing on the entire card from something in the northwest quadrant that's all on black hole card one and it might want to migrate to the east and end up on black hole card one or it might want to migrate to the southwest and end up on black hole card three and that's going to have to take an entirely different uh communications path because the knock can't do that. we have to use these uh these crazy 800 gigabit per second cables that are connecting the stuff all the back that all has to be designed also just going forward with this I've got to gain some experience with all this infrastructure cuz you know there's a lot of new stuff and that was one of the reasons why it took so long is that I really built out a lot of stuff without slowing down and getting some experience with each piece and shining it up a little bit and so forth cuz once you do that. Once you actually go through it and and use it for a while, it it starts to make more sense in my brain. It's it's very strange phenomenon. Even though I wrote the code originally, when I was originally writing it, I wasn't exactly sure what I wanted. But once it starts to work, then in my mind, I say, "Oh, this does this X does Y." And I don't think of it in terms of the code. I think in terms of the tasks that I think it accomplishes. And that in turn changes my view of the code. when I look at it, I say, "Why did I do that? I don't need to do that at the time. I didn't know I didn't need to do that." And so on. And the next big goal that I really, really want, AER subh. I want to explain what that is cuz it's well, for one thing, it's kind of a science of this or it's a piece of the science anyway. AER stands for average event rate. And normally when you have a cellular automa, typically they're all synchronous. They all go kachchunk chunk. So you can talk about the frequency of updates and it's just like one number. This thing is updating at 30 frames per second or whatever it is and that's great but that doesn't apply to the mule feast machine cuz each individual event that uh happens in one of those event processors is asynchronous. It it doesn't happen at the same time as all the other ones. There's all this coordination that goes on and the things get resolved. So we can't just say 30 frames per second. What we actually want to say is how many events does each site in the cellular automa each cell of the cellular automa get on average over the course of a second and that is the average event rate. It's the number of events per second per site on average averaged over an indefinitely large machine. The machine getting bigger and bigger and bigger and bigger. Now in this case the the T2 tiles that we were working with they were indefinitely scalable. So you could actually get an AER indefinitely scalable number out of them. And what we got was very very low like 8 millair something like that like a thousand seconds for uh the sites to all get one event something like that. Here we're talking about AER subh because we're limiting it to the s the size of our grid is HD 1920 x 1080 at least it is in the moment at the moment. If I need to scale it down I will. All right. So that's it. And going forward uh uh so for September 1st that's not very far away. I've try to get myself to clean up uh the code now that at least the communication to the hub is running. Have some kind of inner hub spikes running. I've got little pieces of that, but I have to make a decision. And I would like to uh come up with at least a path towards uh benchmarking AER HD, get the average event rate over an HD sized cellular automa. Maybe expose the live stream to the net. I don't know. Uh maybe we'll, you know, just have a secret on the T2 tile Discord saying, you know, go to this Earl in the middle of the night and have a lot of fun. So that is it. Uh I went a little bit long, but not too badly. Uh uh so I hope this made some sense to you and it did not. It helped convince you that you know uh uh there is a story here. uh um and you know the path to best effort is part of the bigger story uh um that why I'm doing all this is to try to come up with uh technology that fits in with living systems uh better than the technology the digital deterministic computing technology that we have now and you know the AI that's been developed in the last few years uh is all built on top of this deterministic digital hardware and the quiet box is the same thing. Even though the technology, the deep learning neural networks don't really need that and in fact they kind of cook up different answers even if you try to hold things as constant as possible. So this whole project, the T2 tile project is let's demonstrate a scalable kind of computation that would be different and would not take as much energy and would not be focused on maximum gigahertz and pumping tons of gigabytes of data around at incredibly high rates, which is what AI is all about right now. And so instead say, uh uh we can go slow because everybody's [clears throat] going at once. And as long as we can have little reflex loops that get from, you know, seeing something to doing something quickly when necessary, we can then take big long detours and think about the beauty of life and so on and so forth as well. and the T2 tile project and MFMX, the project to get the MFM architecture running on the black hole quiet box from Tensor torrent is a part of that bigger picture. So, I hope you'd be interested in coming along for the ride, you know, once a month. It doesn't hurt too much. In any event, thank you for being here. I hope to see you next time.