Everybody is excited about Intel’s upcoming Xeon with 64-bit extensions, code-named Nocona, and the Pentium 4 with 64-bit extensions, code-named Prescott. IBM’s endorsement hit my inbox before the Intel press release did. Microsoft was so jazzed that it sent Steve Ballmer -- actually, a video of Steve Ballmer -- to the announcement at Intel Developer Forum. I’m certain that when I dial my modem from the one unwired hotel in San Francisco, my mail folders will inflate (slowly) with huzzahs from HP, Dell, and who knows who else.
Intel gets an “attaboy” from me, too. It can’t have been easy for Intel to dig its heels into the bedrock on the unsuitability of the x86 architecture for 64-bit work, only to wrench its shoes out of the muck when AMD did the unsuitable. There are questions about what Intel knew and when it knew it, or rather, how long this whole Yamhill (the rumored name of this project) thing has been going on under the cloak of denial. As I write this, San Francisco is abuzz over whether Intel licensed x86-64 technology from AMD. And what on earth will Intel do to differentiate Nocona and Prescott from Opteron and Athlon 64?
Intel is counting on observers looking down from a high altitude, where AMD’s 64-bit x86 is just like Intel’s 64-bit x86, except that AMD is not Intel. That ought to not be relevant, but unfortunately it is. Intel used its clout to turn Windows into a pace car that held Opteron back until Intel had its x86-64 ducks in a row. They’re formidable ducks: chip sets, device drivers, and development tools. In the last category, Intel has butt-kicking capacity that AMD and Apple can’t hope to match.
Therefore, Intel has squeezed the life out of its erstwhile competitor. How sad.
Oh, put your Kleenex away. Remember, the name of the game is not the number of bits, the clock speed, or the size of the cache. It’s throughput. Intel likes to put a lot of silicon between its CPUs, memory, and I/O bus. AMD’s design focuses on direct (“glueless”) links among up to eight CPUs in a multiprocessing system. Opteron’s on-CPU memory controllers can give each processor a dedicated bank of RAM so that it spends less time waiting for a single pool of memory that’s shared among all CPUs. Opteron’s memory and inter-CPU buses operate independently, thereby removing the stoplights inherent in, shall we say, familiar designs that stack up I/O requests in single file. Before you go sticking forks in AMD, wait for some throughput numbers. As Intel has said, just slapping 64-bit instructions on a 32-bit processor doesn’t buy you much. It’s a good thing AMD didn’t do that.
AMD threw a little party on the first night of IDF -- no, it was not on the IDF show floor -- to fete a new wrinkle in its strategy. AMD’s processors have the reputation of running fast but running hot. AMD announced upcoming reduced-power versions of its Opteron CPUs (dubbed Opteron HE and EE), due to ship this spring. Lowering power requirements and heat dissipation makes my heart skip a beat; I’ve felt like a lone voice in the wilderness on this issue for years. I particularly like this twist: AMD claims to have turned down the heat without sacrificing performance or compatibility with existing Opteron chips. I think I have another “attaboy” here in my desk. Actually, my praise for Intel and AMD is provisional. There are specs, and there is reality. I’ll wait to see how this plays out when spring has sprung.