Explorations

Intel Quad-Core Coming in 2006

Intel has announced that they are releasing their quad-core server and desktop chip lines this year.  This is good news for Intel, which has recently been ceding high-end computing to AMD.  But the big winner will likely be Apple, who may take advantage of this chip in its XServe servers.  Or depending on the timing of the chip's availability, perhaps we could see a dual quad-core Mac desktop this year?  They're due for a refresh, and a dual-CPU quad-core desktop would reset the bar for high-end desktop performance.

Intel Bests AMD on Virtualization

Intel has announced the release of support for virtualization in its x86 line.  What's important about this release is that it should remove the last roadblocks to offering efficient virtualization on the x86 platform.  (The canonical list of requirements for virtualization support in an architecture were laid out by Gerald J. Popek and Robert P. Goldberg in 1974.  For more details, read the original paper.)  With the release of its updated chips, Intel has made the x86 line Popek-Goldberg-compliant.

This is a big deal for two reasons.  First, this will make it easier to implement virtualization software.  Existing general-purpose x86 virtualization software such as VMWare and Virtual PC uses a complex software-based emulation and dynamic recompilation process in order to circumvent the x86's lack of native virtualization support.  This is very difficult to do well, which is one reason there are no good open source direct competitors to VMWare.  Having the proper tools to build a virtualized environment will make it easier for open source developers, and will set the stage for FOSS alternatives in this space.  (I understand that the Xen team is interested in taking advantage of the updates to the architecture.)  This will in turn force the commercial options to improve while driving down entry-level price points.

The second major shift that will be brought on by the widespread adoption of virtualization will result from the change in perceptions that will result.  The concept of what a PC is has been largely wedded to the concept that it's the thing behind the keyboard.  But as virtualization becomes pervasive, we'll start to see a lot more delivery models for desktop software.  New niches will be created in the areas of scaling, management, and exploitation of large virtualized systems.  New generations of ultra-thin clients will be created.  Perhaps users of the $100 PC will even be able to leverage more desktop power on demand.

Widespread virtualization is a paradigm-shifter that it's worth keeping an eye on.

IBM Unveils Dual-Core PowerPC Chips

IBM has announced its first dual-core PowerPC chips.  This really sounds germane to the recent Apple switch to Intel unless you analyze the news strictly through the prism of building a better Mac laptop to fill the hole where a G5 PowerBook would have been.  In particular, pay attention to all the power-saving features they engineered into the chip:

The frequency and voltage of both cores can be scaled downward to reduce the power during periods of reduced workload. For further power savings, each core can be independently placed in a power-saving state called doze, while the other core continues operation. Finally, one of the cores can be completely de-powered during periods of less stringent performance requirements.

Whew, it's a wonder they had any transistors left for doing things like multiplication!  So clearly IBM sees power as an issue.  But the announcement leaves in the dark the question of the power consumption when the chip is running both cores at full speed.  Since even the 970FX burns around 100W at top speed, I'd say that's a hard lower bound.  And the 970MP essentially has two 970FX's!  Add the fact that IBM isn't even pitching this for mobile use, and this adds another reason in support ot Apple's decision to switch to Intel. 

IBM is chasing raw performance, while Apple is chasing performance per Watt-$.  Which makes sense given their vastly differing customer bases.  So Apple is making a bet that the Pentium M is the platform that will deliver the most bang for the Watt-$, as Intel shifts its focus from the Pentium 4 to the Pentium M (if they indeed do so). 

What's Multicore Good For?

One additional note on Sun's vision for their multicore chips, as put forth by Schwartz.  Sun appears to think that chip multithreading (CMT) is good for them, because it plays to their strengths on the back-end where everything is massively threaded.  Internet requests, SMS messages, etc. are all measured in the tens of millions to tens of billions.  Which means that Sun's strongholds of server-based applications should really be able to leverage CMT right out of the box in ways that desktop software can't.

What else can be done with this type of concurrency?  I'm thinking we'll start to see more applications that leverage the rapidly-growing pool of unused CPU cycles for the benefit of individual users.  This is in contrast to the systems, such as SETI@Home and other distributed-computing applications, which benefit a central entity.  I'm thinking about applications more in the vein of the background indexing of the Google Desktop Search application, which directly benefits the user of the machine.  What other things will people find to do with this untapped resource? 

Tim Bray has written a concise primer on some of the challenges that we'll have to solve to really start to take advantage of the new CMT architectures, at least from a pure coding standpoint.  I think what will be more interesting to see will be the answer to the question of what new types of applications require these type of massively parallel chips?

I'm thinking there will be a big breakthrough in desktop virtualization technology that will benefit consumers.  What else is on the horizon?  New applications still have to get around the software as a commodity thing first.

Moore's Still Kicking

Recently, there have been a set of commentators arguing that Moore's law is slowing down or even dying.  I've been asked to weigh in on this by some friends in the investment community, so I'm going to jot down the situation as I see it. 

First, a brief refresher on Moore's law.  Contrary to popular belief, Moore's law doesn't address clock speeds or other outward trappings of CPU speed.  Rather, Moore's law is simply an (exceedingly resilient) observation that the number of transistors that can be put on an integrated circuit (IC) grows at an exponential rate.  Since transistors are the basic building blocks of digital circuits, this has meant more complexity for the microprocessors at the heart of modern CPU chips (more on this complexity in a moment).  In practice, this has also meant that CPU performance increased exponentially as well. 

This is all well and good, except that modern processor architectures that leverage this exponential growth have found themselves at a kink in the growth curve.  Then can continue to put more transistors on an IC, but the increased density creates tons of heat.  This heat makes it difficult to do useful things with these processors, like build laptops and servers that run at sub-atomic temperatures.

That said, engineers are still able to continue the relentless growth in transistor counts.  And being the practical people they are (okay, the practical people the can be), they haven't lost sight of their true goal: increased performance.  When a customer buys a computer, all he cares about is that he can get 30 frames per second in Doom 3.  He doesn't know or care how many transistors it has or what frequency  the chip runs at.  Objective performance is the ultimate goal of processor improvement.

Which brings us back to complexity.  Due to the nature of processor design, more complicated chips make less efficient use of the combination of transistors plus heat dissipation than simpler designs.  So the simple answer is to simplify the processors.  But since you still have the benefit of the shrinkage predicted by Moore, you have lots of chip space left over after you place the simpler processor on you chip.  So what do you do with that extra real estate?  Put in another processor.  Or a few more, if you like.  These mini-processors are called cores, because they still need some scaffolding to make them behave like software expects.

What will happen is that the basic architecture of the chips will change to allow for more simpler components on a single chip.  So now you have Intel and AMD announcing dual-core chips.  Sun is announcing a really slick 8-core processor chip, where each core can process four simultaneous streams of data.  To your software, this monster will look like 32 CPUs, and I expect you'll see performance close to that neighborhood.  (Note: I believe the fact that Sun is doing 8 core chips while Intel and AMD are doing dual-core chips is an artifact of the difference in the architectures of the two systems: SPARC on the one hand and x86 on the other.  x86 is a more complicated architecture, so even a simplified x86 core needs more transistors than a single SPARC core.)  The announcement by Sun is big; I can't wait to see how this thing performs on real applications.

So Moore's law survives, computers get faster, and video games will look more like The Incredibles.  What's not to like?