IntelDeveloperForum2005Keynote

This is a summary of JustinRattner's keynote speech at the Intel Developer Forum for 2005. It is summarized by LionKimbro.

Intel has put info on it on the [WWW] Platform 2015 page. (Is [WWW] this the link?) (factcheck.)

Major Themes of "Platform 2015"

The keynote considers computing between the years 2010-2015, which appears to be the "wiggle room," given 5-year design cycles.

Justin lists some of the major themes of development for this time range:

The last two, Virtualization, and Parallelism, received most attention.

He also mentions, with less emphasis:

Specific technologies are noted as well:

The basic idea is that Intel looks at what people want to do with their chips, and then thinks about how to service those needs. It was nice for the demo, because he actually shows people demonstrating what they want to do, asks what they need, and then talks about what Intel's doing to meet that.

Video Analysis

So, we see an example of VideoAnalysis. We see a jittery photo capture from a camera held in the hand. (If you want to see this, skip to 18' 10 into the video.)

Two things are noted: Low capture resolution, and the shakiness of the image.

The analysis is able to create a version of the image that is not shaky, and has higher resolution than the original. How does it do this? It does it by asserting that the image from moment to moment is of the same thing, and then using the different samples over the same image, is able to figure out how all the "in-between" pixels fit together. (This is something we should consider for [WWW] PaperTalk - multiple scans of the same dot pattern can give us a higher-resolution image, should the scan-side, rather than the print-side, be the limiting factor.)

Justin asks, "What do you need to do this in real-time?" The answer: "1000x performance."

This is all foreshadowing the need for parallelism.

It is noted that these techniques can be applied to existing video. (See VideoAnalysis for more on this theme.)

3D Graphics

Then we see [WWW] ILM representative [WWW] Steve Sullivan, and we see some cool 3DGraphics and renderings.

"These worlds are inherently parallel," he says, if I scrawled it down right. He is referring to the myriad things they are involved in. They want to be able to do everything real-time, not having to deal with long batch processing jobs that work overnight in a server farm.

(See 3DGraphics for more on this theme.)

Computer Virtualization

Virtualization here refers to running virtual computers on your computer. Not just running two independent processes within the same computer, but actually making a good pretense that there are actually 2 computers on the same computer.

Justin poses the question: "What can we do at Intel to help this along?"

This is a very important technology, after all. With it, you can do things like running a program backwards through time, which makes debugging 100x easier. It doesn't just speed up debugging, it actually makes whole new types of debugging plausible, which means that we can make radically different, and much more intuitive, kinds of program designs. (See Virtualization for more on this theme.)

He talked about something called [WWW] Vanderpool Technology. It seems to be hardware level support for switching between virtual machines. This is so you can do things like: Share a graphics card between 4 different virtual machines. The goal is full platform support for virtualization. That is, if you have a piece of hardware supporting the protocols, it should be able to participate in any number of configurations.

We see a demonstration of the concept in software, we are promised hardware support in the future.

See also: [WWW] Intel Virtualization Technology

Parallel Processing

Justin talks a little about progress in parallel processing.

Justin estimates that by 2015, we will have tens (10x) to hundreds (100x) of processors on a single chip.

However, he makes it very clear, that for this to work, there needs to be progress on the programmer's side as well.

He brings up successes in the 1990's with domain specific programming languages and compilers. (In particular, "Baker." But, I can't find anything about it on the Internet- perhaps I wrote the name down wrong?)

An architecture is outlined:

We see a simulation of the run-time working, it doesn't look immensely complicated. He's going to talk about memory bandwidth next.

Part of the goal of the developers forum is to be inclusive to the industry; Justin asks the audience to work on programming languages and compilers. He re-emphasises that without programming model changes, it's not going to work. I may be remembering wrong, but I believe he mentioned that Intel was contributing to the research for the programming changes as well.

Memory Bandwidth

If you have 100's of processors on a die, you run into memory bandwidth problems. That is, it's hard to get data onto and off of the chip.

Usually, you have 100-1000 pins for just one single chip. The problem is, it's really hard to place 1,000,000 pins around the chip. It's just physically very difficult.

So, they're looking at other models. They're talking about stacking up wafers, so that you can route data into the chip from above and below, not just on the sides like we're used to.

This way, you can get 1,000,000+ pins, and support your, say, 100 processors on the chip.

I didn't understand the difference between "wafer stacking" and "die stacking;" Justin said they were investigating both.

Photonics, Optic Signalling

Near the end, Justin said he was amazed about the "continuous wave silicon laser" the company had made.

This is another thing I'm not sure I understand. I understand that they've made a laser out of silicon, and that it's extremely tiny.

I'm just not entirely sure how they intend to use it. Would individual pins of a chip connect to one of these tiny lasers, and then electrons convert to light, fire away, and then be turned back into electrons again? My ignorance tells me that anytime you convert from one type of signal to another, it takes some time. Will they be able to convert electrons to light, fire the light, and turn the light back into electrons, faster than just sending signals along wire?

Or is this not what they have in mind? He mentioned that they had other parts turned into light systems, so perhaps they mean to have the whole thing operating on light in the future?

This wasn't very clear to me.

He mentioned, though, that it wasn't just chips that could make use of these things. He mentioned that they should be useful in chemical and biological analysis, and in medical procedures as well.

Summary

He summarized by saying:

...all this stuff is what they have planned for 2015.

Discussion

Phew! Took an hour to watch the video, and a bit of time to write that up as well. :) -- LionKimbro 2005-03-13 08:30:41

last edited 2006-04-18 15:12:51 by 63