From: James Rogers (firstname.lastname@example.org)
Date: Mon Apr 09 2001 - 17:58:55 MDT
At 05:18 PM 4/9/2001 -0500, Jimmy Wales wrote:
>One potential problem with Moravec's estimates (indeed, any estimates
>based on MIPS) is that even if we achieve human-level MIPs in 5-20
>years, we've only got one piece of the puzzle. One other major component
>is having enough memory (RAM). And so we need to look at Moore's Law
>type considerations for RAM/$$$$ as well.
RAM/$ isn't the real problem, nor is addressability (a reasonable AI
implementation could probably be done in a full 64-bit address
space). The problem is memory bandwidth, which is already becoming a
serious problem for ordinary computer applications.
Let's suppose you have a computer that has enough memory to store a human
level AI and some roughly equivalent level of computational power, both of
which are very easy to foresee in the near future because they are growing
quickly technologically. The problem is that we are a very long ways away
technologically from being able to feed a memory that large into a
processor that fast; improvements in memory bandwidth have been very slow
in coming and it doesn't look like this situation will change.
If you hang out on the scientific computing lists, you already see this
limitation being hit pretty hard. For problems that have substantial data
sets, the computational speed is limited to the speed of the memory bus,
not the theoretical computational throughput of the processor. For these
apps, stream benchmarks (memory throughput) are the hardware benchmarks of
choice, not FLOPS or MHz. (Note that for these reasons the Apple G4
processor is considered a joke in the high performance computing circles
compared to Intel and AMD processors, despite Apple's marketing campaign
suggesting that they give "supercomputer" performance).
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT