AIPU?

[digitalwanderer]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

I thought AI was the huge CPU killer in most games. :?

 

[_xxx_]

Only if it can be SLI'd, we want hardcore!

 

[anaqer]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

Offload to what? How do you accelerate AI calculations?
What kind of special dedicated hardware does AI need exactly?
Likely it must do branching fast++, but beyond that...?

 

[MuFu]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

Offload to what? How do you accelerate AI calculations?
What kind of special dedicated hardware does AI need exactly?
Likely it must do branching fast++, but beyond that...?

t3H n3ur4L n3t\/\/0rXX0rZ!!!!!!

 

[Frank]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

I thought AI was the huge CPU killer in most games. :?

Multi-core CPUs

 

[Killer-Kris]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

I thought AI was the huge CPU killer in most games. :?

Multi-core CPUs

That'll help but you'll need to stick to more traditional cores like P-M and Athlon, and not go to the extremely simplified and massively multi-core chips like niagra and cell, IMHO.

 

[Killer-Kris]

With all the happy PPU talk I was curious if a dedicated card to offload some of the AI calculations were feasible.

Offload to what? How do you accelerate AI calculations?
What kind of special dedicated hardware does AI need exactly?
Likely it must do branching fast++, but beyond that...?

Fast random access to world data?

 

[anaqer]

Fast random access to world data?

Faster than CPU-to-RAM?
Wouldn't that require an entire uber-bus dedicated to AI in games?
I don't think that'll ever fly.

 

[Killer-Kris]

Fast random access to world data?

Faster than CPU-to-RAM?
Wouldn't that require an entire uber-bus dedicated to AI in games?
I don't think that'll ever fly.

I suppose I ment a way to hide those latencies. Much like current processors do with OOOE, data prefetching, large caches, etc.

Essentially processors like the P-M and Athlon tend to make rather excellent AI processors to begin with. They have pretty much all the traits listed so far, low latency random memory access, fast branching/prediction, data forwarding, short pipelines, and will shortly be multi-core.

 

[anaqer]

Essentially processors like the P-M and Athlon tend to make rather excellent AI processors to begin with.

Yup, that's kinda what I was thinking with #3. Now that we even get multicore Real Soon (TM), coming up with a dedicated AI accelerator just sounds like lots of work for very little extra performance and/or functionality.