Last year I spent a lot of time looking at where our industry would be in the year 2020 (see 2-6-08 blog: "Processor Architectures " Where will we will be in 20/20?") and came up with some interesting predictions on everything from where architectures would be to what software would look like, SoCs and even multicore designs, etc. I will be adding some of those predictions to my blog in the coming weeks.
However, in performing that exercise, one of the amazing results was the thought that in the year 2020, a billion transistors would be able to be integrated onto one piece of silicon for about a production cost of one US dollar. Or to put it in slang, transistors would be "a buck a billion."
Wow, a buck a billion, what would our customers be able to do with that and what might this mean to us? I can think of several implications:
- What am I going to do with billions of transistors?
- How does this change the way I do IC design?
- How does this affect performance and power dissipation?
What can I do with a billion transistors?
When I give lectures on innovation at universities and conferences, I generally end with this statement:
"Technology will not be the limiting aspect of innovation. What will limit us is our inability to figure out how to take full advantage of the technology."
What I am doing when I say this is to actually pose a challenge to systems innovators to stretch their imaginations to a point where they have nothing limiting their ideasparticularly technology. And, yes, I'm still waiting for the brilliant new product ideas that will need a billion transistors. But please keep in mind that we like to create devices that sell for $10.00 to $100.00. So no one should limit themselves to only one billion transistors but tens of billions of transistors.
It seems that the only obvious use is mass storage. I certainly hope this is not the ultimate innovation for billions of transistors.
How will this change the way we create ICs?
Several years ago we announced a new technology node that would allow us to integrate 100 million transistors on one device. I was asked to put some reality to what this meant. I summarized with two ideas:
- At that time a PC with all of its processors and memory added up to about 100 million transistors. So, a complete PC on a chip was a good mental picture of the magnitude of transistors we were considering.
- The design of a 100 million transistor device could not be done from scratch. The math was simple, if the average time to complete the design, per transistor, were one staff hour it would take a total of 100 million staff hours. Since there are roughly 2000 staff hours per staff year, this comes to fifty thousand staff years to complete a design from scratch. Other methods would need to be invented in order to design a device with that many transistors.
One method, for example, is one I call "sloppy design." With all of the transistors we can be sloppy in our design. We can use extra transistors to make it simple. We can use microprocessors with a lot of memory rather that a custom devices which efficiently uses all of the transistors. Programmability will continue to be our primary way of being sloppy.
How about performance and power dissipation?
It is this sloppy design concept that has given, and will continue to give us, higher performance and lower power dissipation:
- With more transistors I can increase the performance. We call it multi-processing including the use of accelerators along with microprocessors.
- With more transistors we can lower our power dissipation. We do this by either turning things off when not being used (I call this the father's solution to low power), or use multiple processors at lower clock speeds and lower voltages to do the same task.
So, what will we do with billions of transistors?
- We'll drive sloppy design to an art form (but we'll call it something more technically brilliant).
- We'll drive higher performance and lower power dissipation using more transistors.
- And, with any luck, we'll figure out how to do things we could never do before.
What are your thoughts? What would you do with a billion transistors for a buck?