To view parent comment, click here.
To read all comments associated with this story, please click here.
This sort of gets back to my thesis, hardware currently has a Memory Wall that continues to grow, that makes software that manipulates large data sets run much slower with all those cache & TLB misses. I'd bet the kind of app that would simulate HAL would have an enormouse working set of maybe 100s gigabytes of "distributed knowledge" (what else does 10^10 neurons do). What the current crop of cpus are good at is running media codecs and the like that does alot of DSP on small tiles of pixel data. I am sure knowledge software is the complete inverse of that.
A 64 bit address aspace should help on the size of the knowledge base but the Memory Wall esp to hard disk stops the software from connecting it all up. I bet that distributed multiprocessors on a chip with real support for concurrency will be much better suited for AI software. Years ago Transputers produced lots of AI conference papers.






Member since:
2005-07-06
"perhaps the hardware we have is not really suitable to build such software."
Well, for instance hardware 64bit support has been available for quite a while, and it is getting better very fast. But I don't see many 64bit apps.
Even less I see apps which can take advantage of 4 or more CPU cores.