John Carmack on PS3 Video
- Thread starter Ben-Nice
- Start date
patsu said:
The limited amount of local memory could be a problem when you need to know whether an agent sees another agent. This is not a local problem.
The fact that real world works locally tends to produce an illusion that a game world can be simulated locally. But it's not true. For example, when a sniper shoots, you don't simulate the bullet flying through the game world. The bullets are too fast so they are normally regarded as "infinitely fast." The same goes for lights. In real world, I see you because the lights reflect ed from you reach me. But obviously you can't actually do that in a simulation.
So consider this: you need to know, for one agent, which agents s/he can see. Now it's a global problem because you need to know the position of all agents in his/her proximity and the terrans/buildings between him/her and these agents for occlusion tests. These kinds of works can dominate the process of these agents.
This part is realistic ...patsu said:
In this part they are behaving like they are capable of instantaneous communication and cooperative problem solving ... how often would you want to simulate that?
Do people doing musical chairs use a shared algorithmic approach to finding a seat? Do people blocking eachother on the sidewalk use a shared algorithmic approach to finding a path? No, they try, hesitate, retry ... as many times as it takes.
_xxx_ said:As said, maybe it was ripped out of context or whatever, but I surely don't know what his motives are. I said that I would never comment any such stuff
Heh, true.
I was just curious about your opinion. In some ways it's the curse (or blessing depending on how you look at it) of being famous. You can and will cause massive contraversy by your slightest thoughts and actions. In Alan Greenspan's case, it even moved markets. It takes a lot of finese to be able to control that kind of power.Nite_Hawk
Could we please stop telling that Carmack is lazy ? He is still searching and finding new ways to solve problems. The perfect example is the megatexture-thingy. He talked about it back in Trinity days, full surface unique texturing was something he was experimenting with in 1999. He is not the kind of guy doing a patchwork Engine with all the bells and whistles that everybody does, so repeating that he is lazy ad nauseam is unfair.
Carmack doesn't particularly love the Cell ok so what ? Maybe he is more interested in solving problems that fighting the hardware. And Carmack has used other platforms than the PC, one of the more exotic one was the Atari Jaguar with one Motorola 68k and two proprietary RISC CPU that couldn't access main memory and forced you to segment your code and data in 4KB or 8KB pieces, that reminds me of something
pcchen said:
It is a local problem and is quite realistically approximated as a local problem, as real world people tend to pay attention to what is local.
By definition, if agents need to know about agents and terrain in proximity, it is a local problem. One does not need an exact solution for occlusion/visibility tests, and agents are permitted to have blind spots, to "miss details".
Real world people focus on what is salient and do not have perfect vision like AI, wherein an agent will notice you and all other agents as long as a single pixel of your body is visible. This is completely unrealistic and defeats games where snipers try to use cover or camoflage since the AI doesn't care.
With just 128k of RAM, you can fit 10k hull vertices for visibility checks, which I think is overkill.
mckmas8808
Legend
Arwin said:
Yes. The interview was done at E3.
Kolgar said:
I am not sure how many threads they are on, but back at GDC they said they were still working on the renderer working on its own thread. I am not sure we can read that as on only one core (I guess they had said in the fall they were on only 1 core still, but that was 6+ months ago). But honestly none of this tells us much. Even if a game runs on 3 cores (or 8 on the PS3) that tells us little about utilization and effeciency.
Btw, nice to see ya back Kolgar
Technically 7 on the PS3, isn't it? What with one SPE disabled and one dedicated to the OS and underlying software...?Acert93 said:Even if a game runs on 3 cores (or 8 on the PS3) that tells us little about utilization and effeciency.
Btw, nice to see ya back Kolgar
DemoCoder said:
It's not "miss details" the biggest problem. The biggest problem is more like "seeing through walls."
And if you limited your agents to a relatively small detection range, your AI will very limited because people know they won't be seen at a certain distance, while they can still see these agents.
DemoCoder said:
Considering that you have to fit everything (including the program itself and all temporary variables) inside the LS (256KB), I doubt that you can freely use 128KB of RAM for the vertices.
And remember when the agents are moving, you have to constantly streaming new data into the LS to remain "local." This can also stress the memory bandwidth unnecessarily compared to a shared memory model.
Last edited by a moderator:
As Acert said, at GDC they said they were still on 1 core, working on the second core to double framerates. And in other interviews Rein/Cliffy/Sweeney said they were working hard on finishing the new mult-threaded renderer. Don't know why they are using the term renderer when they are talking about using the X360 CPU, I guess the GPU won't do all the graphical rendering for EPIC, or maybe they are just talking about physics rendering.Kolgar said:
You might want to google on 'multithreaded', 'multicore', 'EPIC' and 'renderer' in 1 search for some interviews.
I also read in interviews that came out of GDC that Kameo was for example using 1 complete core for decompressing heavily compressed data from the DVD, and that PGR3 was using one core completely for audio. The decompressing-thing make sense, since D. Braben said he expected problems with the lack off DVD-capacity. Looks like they need to compress heavily to cram all the hi-def gamecontent on the 7GB DVD9's, and therefor need a lot of CPU power to decompess it in the game, especially with streaming.
Seems to me that there's still an awful lot to do on maximising performance on the 360-CPU too, cuz when I think about these things it looks like a big trade-off for such a relative 'small' game-aspect. Using 2 cores of the 3 for decompressing and audio leaves very little room for extra's, and that might explain the relative poor physics performance in 360 games so far.
Carmack itself described in an interview how he is thinking of developing the next major id Software game on the 360 as basically having 2 big threads, one for the game engine, one for the renderer running on 2 Xbox 360 cores with the 3rd core simply dedicated to applying optimizations and physics back to the renderer or directly to the renderer outputs running on the graphics card.
Btw, still wonder if the guy's working on a daily basis on the PS3 with his team to know what he is talking about, looks like he's still mainly focussed on PC, 360 and handheld stuff. He has a very pronounced opinion in every interview about the Cell, but is he actually working on a game or engine to really know what he is talking about, on an official PS3 devkit, or is he just making general guesses. If that's the case, his opinion looses significant weight. I prefer the expert opinion from devs that are digging deeply into the system, making progress he may not have been thinking about.
Last edited by a moderator:
mckmas8808
Legend
TrollFace said:
I'm wondering that too. I wonder what a conversation would be like between John Carmack and nAo?:smile:
mckmas8808 said:
No offense to nAo but I'm gonna guess that John Carmack forgot more about programming than he'll ever know. On the sidenote I really don't understand why everyone thinks that cell is so revolutionary when all the evidence is pointing towards it not being a great gaming CPU. Xbox had like 5x theoretical power of PS2 did the games look THAT much better?
Similar threads
- Poll
