cancel
Showing results for 
Search instead for 
Did you mean: 

Price, Performance, Functionality

Scholar
Scholar
0 1 42.7K

I attended the ITRS quarterly meeting, as the regular Xilinx representative was unable to attend. ITRS* is the International Technology Roadmap for Semiconductors, a set of documents and the “keepers of Moore’s Law” since 1998.

 

Gordon Moore** stated that twice as many transistors could be placed in the same area every 18 months as the technology improved. This cycle of improvement began in 1965, and has been with us ever since.

 

Unfortunately, this geometrically increasing functionality for the same cost (or area) is totally unique to integrated circuits. And (as with all good things, especially those that are too good to be true forever), there must be an end.

 

How Much Longer?

 

The industry has recognized that we are reaching the physical limits of the fabrication of integrated circuits as we know them. The horizon for the shift to something else (as yet, unknown) is after the 22nm technology node.

 

Right now, as we have seen with Intel and AMD, they can go no faster, but they can put two, three, or more cores on the same chip. So, performance doesn’t improve, but functionality doubles, and for the same cost. The magic of getting everything is now reduced to getting at best two out of three.

 

Traditionally, FPGA devices have benefited strongly, as we have used the best and newest technology as soon as possible. We have provided more, for less money, even before ASIC or ASSP vendors.

However, we too face the dilemma: do we provide “best two out of three,” and what is “best” for each business sector?

 

Navigating Uncharted Waters

 

This new ITRS Roadmap represents a major shift, as our customers have adopted Moore’s Law into their business plans.

 

Why does this matter? Well, they have enjoyed 20 years of being a Xilinx customer, with costs per function going down with every node, performance increasing, and a doubling of functionality in the largest part. What if no one can supply what their business plan requires? Who will answer for their failure?

 

What To Do?

 

First, designers are always thinking of ways to obtain improvements that do not rely on the technology node alone. So we continue to do that.  We do get the doubling of the number of devices, at the same costs. That density improvement alone is still worth a great deal.

 

What Can You Do?

 

As our customer, you can help us navigate this unknown territory, by letting us know which of the three process node related improvements (price, performance, functionality) are absolutely required, and which ones are not.

 

For example, do you just need more LUTs to do your job, while speed and power are okay the way they are? Or, do you need less power at the same price, with the same functionality?

 

We like to think of ourselves as a clever bunch, so we remain committed to providing the improvements that our customers have always enjoyed. We just may not be able to provide all of the benefits that have historically been provided by every new process node technology.

 

Austin Lesea

*International Technology Roadmap for Semiconductors
http://en.wikipedia.org/wiki/International_Technology_Roadmap_for_Semiconductors

 

**Moore’s Law
http://en.wikipedia.org/wiki/Moores_law

Message Edited by austin.lesea on 05-30-2008 02:14 PM
1 Comment
Explorer
Explorer
Although Moore's Law as stated only applies to transistors, Ray Kurzweil points out in his book _The Age of Spiritual Machines_ that the exponential curve actually extrapolates smoothly for four technologies before integrated circuits, all the way back to mechanical computation circa 1890 or so. This is also covered in his paper "The Law of Accelerating Returns", available from his web site:

http://www.kurzweilai.net/articles/art0134.html

Kurzweil predicts that even when current semiconductor technology stops scaling, there will be other technologies that continue the curve forward into the future.

Note that Moore's law only indirectly promised increasing circuit speeds, and not on any specific curve. It's not entirely surprising that at some point in transistor scaling the speed would hit diminishing returns.