Let me try to explain. Heretofore, the vast computing power of the Internet -- and servers in individual enterprises, too, but the 'Net makes for a more accessible example -- has come about because of the sheer number of servers out there. (Many of them are cheap boxes with reasonably powered, but not gold-plated, processors.) In such a situation, to double the overall computing power in the universe, you'd have to double the number of servers.
Except that would be hard to do, because a) this number of preexisting boxes is so large; and b) people are running up against the power and cooling barrier. As in, the electricity to run all this stuff is starting to cost serious money. (This is the biggest factor in the current trend toward consolidation.)
Hey, doesn't this sound like a large-scale version of what happened in 2004, when the chip industry had to make a "left-hand turn" away from increasingly power-hungry single-core processors, toward more manageable multicore designs? Indeed, it does!
So back to my jumping-through-hoops analogy. My point is, now with the new AMD Istanbuls and Intel Nehalem-EXs, we're going to have so many physical cores (and a higher multiple of logical cores) that you could theoretically double all the available computer power out there by adding. . . let's say only 10 percent more servers (assuming they were these high powered boxes).
So I see this analogy is a bit strained. But I think this argument dovetails with the case I made a few months ago that, as high-end chips get more powerful, a technology trickle-down effect is at work. So commodity servers, in turn, are getting more powerful too.
Of course, as I argued last week, the biggest challenge in attempting to harness all this new-found computing horsepower will be cracking the software nut. We've never been very good at decomposing apps across multiple threads. Mostly, more logical and physical cores are used as separate server resources.
Which is fine, but I think there's kind of an aspect of industry self-delusion going on, where we've used buzzwords like threads and virtualization for so long that we think we're at the end of the road. We're really just at the beginning of a new era. As we were when RISC emerged and began to edge CISC aside. (See how I neatly brought my post 'round back to the beginning?)
OK, now on to Istanbul. Check out Margaret Lewis's blog at AMD, where she talks about the HyperTransport Assist technology, which provides cache-level optimization in Istanbul.
As AMD's John Fruehe explains in another AMD post, HT Assist increases memory and I/O performance by reducing the overhead of cache lookups.
Here are some Istanbul slides (click the slide below to see more):
AMD's six-core 'Istanbul' Opteron server processor.(Click picture to enlarge and to see three more shots.)
Intel's Nehalem-EX eight-core server processor.(Click picture to enlarge and to see three more shots.)
Follow me on Twitter: (@awolfe58)