I've learned that Intel is wedging notebook CPUs into places they ought not go. The Core Duo CPU, the very same one that's now in iMac, MacBook Pro and (cover your ears) Mac mini, is being branded as a Xeon and sold for blade servers. That's -- well, that's just not natural, that's what.
To me, low-power servers are natural to the point of inevitability. Intel can't stamp Xeon logos on notebook chips fast enough to suit me. I remember as clearly as if it was the last time I boasted of it here, how I wrote in my column for Windows NT Systems (R.I.P.) that my approximation of the ideal server was a stack of laptops interconnected via an infinitely fast and scalable bus. It was lucky for me that coaxial Ethernet, in vogue at the time, made an infinitely slow bus, otherwise I'd have had to prove my cockamamy theory.
I believe in that vision now as much as I did then, and I can see what I imagined actually taking shape. I know I had nothing to do with making it happen, but I've got to admit, as I watch computing go green (and virtual, which is another of my long-standing, perfectly practical passions), I feel like an excited kid.
Intel's about five years too late to claim first place in the race to bring the pieces of my low-power-server vision together. AMD solved the problem of the infinitely fast bus. AMD's HyperTransport is precisely what I had in mind. It moves data as fast as CPUs can generate it; it scales as CPUs, memory, and peripherals get faster; and it is as easy to create a HyperTransport fabric as it is to use HyperTransport as a point-to-point link. The beauty part of HyperTransport is that AMD not only uses it to link discrete CPUs in a multiprocessor system, but the very same HyperTransport technology links cores inside multicore AMD64 processors. When they externalize that bus, I will achieve nirvana.
Now Intel gets its due by ponying up for Part Two. The Core microarchitecture is based on Pentium M, which is based on Pentium 3, which powered the last cool and quiet dual-processor x86 servers ever built with Intel parts. Just as HyperTransport fits my specification for an infinitely fast bus, Intel can craft Core into a server platform that can really, truly, run on batteries. And it isn't just the CPU. I'm using a MacBook Pro to write this story. It has two 32-bit x86 cores, 2MB of shared cache, 2GB of DDR2 memory, internal and external PCIe (PCI Express) buses, and Gigabit Ethernet. My Apple-brand mobile Xeon server blade runs for about four hours on a charge. Two grand for a server blade, much less one you can pull out and take home with you, is a bargain.
That stack-of-notebooks theory might have direct implications after all. But the stick in the wicket is Intel's bus. For Intel, the data/memory kind of bus has much in common with a beep-beep kind of bus.
AMD's got no bus worries. AMD has 35- and 55-Watt Opteron CPUs; an impressive dual-core, low-power Turion notebook design; and a chipset of its own, Turion, which is shipping now as a 64-bit part and will be going dual-core very soon. You see, the whole time Intel was working on Pentium 4, AMD was working on its version of a clean-room-developed Pentium III. Intel had to go back, but AMD never left.
My hat's off to Intel for branding notebook technology with Xeon. But when AMD calls for our midyear briefing, I'll be there.