Simple Questions Simple Answers - Page 261

You can always put RAM on the CPU add-in card, as you do a graphics card. That's what they're doing.


At a high level, you can think of processors and computing this way:

• The more generalized the hardware and the more different kinds of instructions it can do, the more expensive / complicated / large die area / flexible / less efficient (power) / easy to program (more efficient for the programmer in terms of man-hours, feasibility) it is.
  
• The more specialized the hardware and more fixed-function it is, the less expensive / less complicated / less die area / less flexible / more efficient / harder to program (particularly, more complicated tasks) it is.

CPUs are the most generalized, with GPUs more specialized than that. Because the GPU "cores" are simpler from not having to handle as many possible instructions and so on, they take less space per unit and you can stick more of them in a chip (of equal size) as CPU cores. Modern GPUs have hundreds or thousands of execution units inside, while some behemoth server CPUs have up to 16 (or so?) CPU cores in a chip.

The problem is that less types of computing jobs are suitable for running efficiently on GPUs than on the more general-purpose CPUs. All kinds of graphics calculations and some kinds of large scientific matrix calculations, etc. naturally work well on GPUs, particularly because they can work on data in parallel where the outcome of one operation is not needed by another piece.

There are a lot of fixed-function DSPs, FPGAs, etc. for very specialized tasks that pretty much just do one thing, so even further away from the general-purpose CPU. Sometimes people think about this difference as a software vs. hardware thing. Software = flexible, slower. i.e. runs on those general-purpose CPUs. Hardware = less flexible, faster.

I think some of the discussion could have been about the integrated GPUs that are found in most CPUs these days. The idea is that one day, all sorts of tasks can be offloaded to integrated GPU cores, much more so than they are today. And in a very seemless fashion, passing data back and forth between the integrated GPU portions and the CPU portions. And eventually, with everything right next to each other, sharing data easily (recall that passing data between CPU and a graphics card takes a relatively large amount of time in terms of CPU cycles, so co-location is of interest). That's particularly where AMD wants to go, because their x86 CPU cores are worse than Intel's, but their GPU cores are better than Intel's and certainly more established, at least for now.

As processes shrink, you can get more transistors per area (more or less, per money), so if you can find a way to make use of those extra transistors, you can get more performance.

+ Show Spoiler [oh yeah] +