Of the interconnect technologies Intel is working on, which ones are likely to have the biggest impact on corporate computing?
The five that I'm really excited about are optical, high-speed Ethernet, 3GIO (third-generation I/O, the replacement for PCI architecture), Serial ATA (Advanced Technology Attachment) and Infiniband. These redefine connectivity in the enterprise and in the data centre in a fundamental and dramatic way.
What other technologies are likely to have a significant impact?
Our goal with the Itanium processor family is to rearchitect the data center of the future. Today, that's filled with proprietary RISC-based machines, and our job is to move those to standard building blocks.
Will the IA-64 processor commoditise the 64-bit server market as the Pentium has done in the IA-32 space?
[With IA-64], we're trying to deliver a building block for big-iron machines. It's not that those other ones are bad; they're all different, they're all incompatible. They're all forcing investment in areas that no longer are the things that IT cares about.
Intel has announced the 2GHz Pentium 4 processor. How fast can you go?
We're on path to deliver multibillion transistor chips in the next decade, operating in excess of 25GHz. We're going to keep pushing away at clock speed. However, clock speed alone will become less and less a determinant of performance as we look forward to things like hyperthreading.
How will that work?
Hyperthreading is the idea of doing more than one thing at once. In today's applications, when you finish one set of instructions, you go on to the next set of instructions and so on. In the future, we [will] have the transistor budgets and the technologies to have one microprocessor doing more than one thing at a time.
This is reasonably well established in servers or high-end computing. We want to bring this into the mainstream of computing. You'll see the first implementations start to emerge next year.
What optical technologies are you developing that will affect IT?
Optical redefined long-haul networking over the last decade. Over the next decade, [it] will have dramatic implications for metropolitan and campus data centres. We want to get to the point where we're building direct optical interfaces onto our silicon component.
That's a long-term vision, but where an optical interface component might cost $10,000 today, over the next decade, I want to make it cost a penny. I don't see [Intel building] optical transistors and things like that in the near future, if ever.
Will applications need to change to leverage hyperthreading?
Maybe. If Microsoft builds in the ability to have an operating system and a networking stack and the printing daemon running in parallel, and it's all part of the operating system, then the application could benefit from hyperthreading without requiring any modification to the application itself.
Imagine if I were running one of these cool, new dynamic runtime applications like a Java or a C# application and the garbage collector, the [just-in-time compiler] and the dynamic runtime environment [were] separate threads running in parallel with the application. So my new C# application wouldn't need to be restructured for multithreading, but the underlying dynamic runtime environment would be.
The third example would be to put these capabilities directly into the compiler so that the compilers automatically generate those threads. So the application doesn't need to be rewritten; it just needs to be recompiled. However, to get the greatest benefit from hyperthreading, yes, you would need to rewrite the application. But you're going to benefit from this well in advance of requiring this restructuring of the applications themselves, by any of the first three paths.
What will a typical server look like in three to five years?
I see us building four- or eight-way machines that are clustered together to build really big machines. So I have a rack [of] 16 4u (1u equals 1.75 in.) slices, and each slice is an eight-way Itanium, and each of those is using hyperthreading.
I might have four threads running, so within a slice, I could have as many as 32 threads of execution going on. In a rack, I have 16 of these, and they're all clustered together using technologies like Infiniband. That's the mainframe of the future And that will blow away the price performance of anything that the alternative approaches will offer.
What will the desktop look like?
Key technologies will allow us to repartition the form factor of the PC things like 3GIO. And when you've done that, you can all of a sudden start separating. I can show you a system of the future where my computer is actually under the desk, and I have all of my I/O capabilities on top of the desk or even integrated into the LCD monitor.
I see form factors continuing to decrease, power efficiency becoming more critical, flat-panel displays becoming dominant and technologies like speech and audio being good enough that they are meaningfully deployed. The move to speech, pen and vision computing [will] really explode.