or to get the monster to run through the fire and react.

Most games with exceptional AI have precomputed data about each level and that often makes the difference between the NPCs behaving like lemmings/cannon fodder or otherwise. However if the lookup/pathfinding tables were able to be computed in real time the behaviours would look less scripted and the AI better able to respond to a dynamic world. For example a character wants to get from point A to B and has his/her trust BFG they could blow a hole in a wall to get there, however the AI would just ignore this new pathway or be omniscient and shoot you before you could even see them. Red Faction anyone?

This is where i suppose cell could come in. Calculate volumes in a map as they are created and offer the AI the chance to both react to the new "doorways" and make their own.

The way they did it in Half-Life 2 was to give the level a node map; the AI would calculate the best route from one place to another, and on the way there, work out if it should be hiding behind a barrel to miss gunfire, or what it can do to best create a better route (certain node map points triggered a script event).

They accomplished one-on-one fights between characters by simply having them intersect on the node map; it would then check their positions and look for the closest tween in the animation, and run that one. So you didn't have a totally scripted encounter, but in fact the vast majority was. At the end of the day it looked pretty damn good, and if you think you may only play the game 3 times all the way through.. it only needs to look good for those 3 times.

I think dedicated physics and AI cards (like AEGIA and these other guys) are a gimmick, running such calculations in a spare socket (like AMD want to publicise http://en.wikipedia.org/wiki/Torrenza) is going to be dead before it becomes alive. But there is always room for a dedicated, MULTIPURPOSE solution, which I think is Cell.

I just wonder why if they think they can sell a $300 Physics card, a $300 AI card, $300 pair of graphics cards, why they can't just make a single graphics card that does all that, and have a $300 Cell processor stuck in a slot, which will last people forever. Maybe it's because there is no scope in upgrade markets

_________________
Matt Sealey

The main business contingent of ageia was the API, which is dominated by havok atm. Unfortunately Havok is yet to be supported under linux (which is a pity) which is a major sticking point for game developers who are experienced with the havok api when it comes to porting a game.

ODE is a step in the right direction for physics for games under linux/osx/windows because it is open but it does lack plenty of features which are finding their way into commercial games, ie user manipulation of in game objects (aka "the gravity gun")

But are they mil-spec?

Cell PCIe card is a nice idea I mentioned to Mercury in their Webcast live Q&A session 14 months ago.
I wonder if IBM or someone else will be doing a professional multipurpose gfxcard with Cell just like Sun put their MAJC on such a card.

Maybe the 65nm shrink will be the sign for Genesi to get into the game.

Neko, no offense meant but i don't find the connection between your post, and the one from me you quoted. Except maybe talking about the price?
If so.
Mercury Card Costs 7999
Playstation3 "20GB" version is 499 (not 599 sorry)that was meant with the "big company that will bring out something with 1 Cell and so on..."

So in my opinion an Cell add on card shouldn't be much more expensive than a whole Game console with an integrated Cell. Even if other companys bring out strange Physics card for about 300 bucks.

<Editet for clarification of some crappy sentence i wrote>
Cheers

And usefulness of the card. You made a small comment about the PS3 without mentioning it - ask yourself what are Sony going to use Cell for? Of course they have licensed Aegia's physics API to convert it to Cell and get more game support. So why do Aegia need to make a dedicated PhysX chip, when they could put a Cell on a PCI card and sell that?

Also for $300. And it would kill the AI accelerator guys off in a single swoop, allow Playstation quality games on a PC.. oh but of course isn't that the problem? That Sony will be very disappointed if you could buy a $250 or $300 add-in card for your PC, that suddenly allowed PC games to completely triumph in image, audio and gameplay experience?

You just have to wonder why nobody has done it yet. I expect that it is entirely down to software support; Genesi could provide a Cell card in a couple of months if the wind is in the right direction. But what are you going to run on it? We don't have any games coming out. Spending the time TALKING to games developers, would take eternities.

Maybe this is why the $7999 Mercury card is $7999 - because Mercury are expecting to have to offer such especially customized support per customer.

_________________
Matt Sealey

Now i understand what your point is. But i did not mention gaming at all . What about writing an enhanced version of ffmpeg by Genesi bundled with the card, or ask Apple for an updated version of FinalCut Pro, maybe another Software Vendor showing them what Cell could do. Ask the guys from gromacs.org if they will provide Cell enhanced Software (Gromacs seems to run on Cell already see Cure@PS3). Maybe it's not usefull for me as an end user, but for Genesi selling units to Universitys/Scientific Companys? They surely will pay more than i do, maybe 7998 bucks per card?

Bye

Please find us 50 Universities who want to buy 50 cards each and we'll make them for them

One thing we threw around the office was it would have been cool if Elephant's Dream was rendered using POWER5+ or Cell (it WAS rendered with Mac G5's however, and XServes). Too late now.

I think it would be more useful to find a solution everyone can buy (every one of a million people buying a $300 card is much more profitable than every one of 20 people buying a $8000 card) and everyone can find a use for. Be it rendering, or Seti@Home, etc., this requires a lot of software support, or it is useless bit of silicon wasting power in a slot.

You're right it could be any software, but should we talk to every software developer on the planet about this? We run into the same problem we do with the EFIKA and Pegasos; Nobody can develop for Cell without a Cell, although there is a simulator, it's nothing like the real thing (you can develop software for Pegasos using psim/gdb and the Intel/Solaris versions of SimG4+ too, but nobody does, because it is a pain..)

One thing we could guarantee; you could probably boot Cell Linux on any card we'd make. Running it in parallel to your ordinary PC or Mac system might be something someone can find some fun in..

_________________
Matt Sealey

Sorry can't do it for you, i had only this crazy idea i wanted to tell you. Thought that you maybe could make use of it...
Oh, the damn hen/egg problem appears again. Developing could be done on YOUR (Genesi's)Cell Board. You don't need all software developers on earth on board. *If* (i hate this word) Rendering in BlenderCE (CellEnhanced-should i protect this as a trade mark?) will go 40 times faster than before, companies will come to you and ask for support, because they probably can not afford to be sooo slow. And even if not, people will know that they have to wait one minute in BlenderCE and 40 Minutes in Cinema 4D, so which one will they use, and whose Cell accelerator card will they buy?

Ok the part before is maybe some "simple thinking" by me. So last one from me, as you and BBRV are the ones who probably know what the market want's.

Doing a little brainstorming. I get 3 (open sourced) Apps Cell will probably be good for.
Gromacs
Blender
ffmpeg

Doing resarch on it if it's viable to port and produce real hardware is up to Genesi.

C'ya

We'd have to design and produce the board before we talked to any developers. What if nobody saw any software apart from 2 or 3 minor apps that would benefit from it? What if nobody in a game company wanted to re-support Cell for the PC?

OR we could buy every developer a Mercury Cell board and get them to port the software while we design it. That means we would spend $8000 for every application we want, maybe 3 boards per application if it is a big project like Blender (I don't think you can take a project with 1000s of developers and give ONE guy a board and expect a fast result).

Cell also has a major learning curve. We have trouble explaining the niceties of AltiVec and general SIMD optimization to some developers. There is ALWAYS a need for teaching and training sessions (we are still working on this, and have a University willing to do it!)

It would need some major money behind it. Maybe IBM think it is a good idea. If they can envision selling a million Cell units on top of Sony's PS3 sales (6 million by March 2007 if you believe them), maybe they would be so cheap, that everyone could have one.

_________________
Matt Sealey

good point, the key aspect is price... (always has been always will be except for extreme tech-heads). But applications for cell have to come sooner than later to speed up the assimilation of code and skills into the cell architecture. The PS3 might help on this point though.

As far as hardware (i know the cell will be expensive, new tech always is) a PCI card would not be the best solution because there would not be enough bandwidth between cell and the mainboard. So extra ports on the cell board could be a solution or it could be integrated onto the mainboard itself.

Any way i look at it i just don't see how they can push Cell onto the PC landscape without sponsoring the hardware to developers or pouring money into creating a killer app which makes it almost mandatory to have the cell hardware.

As to SIMD i still have troubles working it all out, even though i recently got a book about programming vectorised code.

I already asked for a shell on a Cell system and I did by best to get the toolchain (that seems working quite fine now) and the emulator (that now is running but that's a pain since using it in delay exact mode is quite slow and using it in loose mode can cause a large and vary degree of issues you cannot track with ease)

I don't have much time in this month to proceed any further and I'm quite in the "wait and see" mood about Cell/PPC hardware release...

You are right on the integrate on mainboard part, but designing a whole new Mainboard is more difficult i think, and people had to buy a new workstation... Bandwith limitation on PCI Express are there, but if a card has enough on board ram it shouldn't be too bad.

lu_zero if you have something working please keep us updated!

Bye

http://likerabbits.blogspot.com/2006/09/hard-cell.html

I made a blog post.

In summary; Cell is expensive. But only because the most expensive solutions use large amounts of XDR RAM and Cells which are the cream of the bin sort when they finish production.

How many SPEs does a developer need anyway? And Sony has proved XDR can be affordable in limited quantity.
Can they make do with DDR PC2-8000 for the RAM?

At some point you have to make a compromise between a $800 card (or cheaper) and an $18000 blade server. One Cell, with 4 or 5 SPE units on board, adequate amount of adequate speed RAM, etc.

Remember the Cell has a lot of processing power so it is also a HOT chip to run. The 2.8GHz Mercury card soaks 210W, which is 3 times the amount a PCI Express slot can support, for the chip, 8 SPEs, and 5GB of RAM.

You could substantially accelerate gaming, image processing, POVRay/Blender, and things like encryption etc. with a lot less.

I don't think a port of ffmpeg really will benefit Cell - only because, you are going to be using a chip capable of realtime rendered high definition graphics to decode an MPEG stream. That seems a little TAME.

2 or 3 ~100W Cell boards, with 256MB DDR2 RAM each, is an instant render farm. How fast could you render the entirity of Elephants Dream on such a thing, let alone just play an 800MB AVI back?

_________________
Matt Sealey

>And Sony has proved XDR can be affordable in limited quantity.

Afair XDR RAM was about 5x the price of regular RAM?

>How fast could you render the entirity of Elephants Dream on such a thing,

Wasn't it rendered on a G5 farm?

Can't you just use a cluster of PS3 with Linux on it?