Page 1 of 1

Performance Analysis and Validation

Posted: Wed Feb 18, 2004 5:26 pm
by Jonathan
The Intel Technology Journal article I contributed to is available online now. The full title is Performance Analysis and Validation of the IntelĀ® PentiumĀ® 4 Processor on 90nm Technology. It describes the current state of performance analysis here at Intel, which is what I do.

I've worked a little bit on most of what is described in the article. However, the part I wrote is in the section Post-Silicon Performance Validation and is titled "Optimal CPU Performance Feature Tuning". Tuning Prescott for optimal performance is what I spent most of the summer and fall working on in 2003. We laid the groundwork for doing this in the spring.

We used a genetic optimization algorithm and the SPEC CPU2000 benchmark suite to improve the performance of Prescott. Our main source of improvement was in Hyperthreading performance, but we did get gains in single-threaded workloads as well (see chart).

I'm pretty pleased with how this turned out. For the 65 nm processor I'm working on right now, I hope to decrease the number of weeks it takes to do this tuning and get our optimizations in the first stepping to be sold.

Posted: Wed Feb 18, 2004 8:14 pm
by Zekmyr
It's interesting to hear some of the more industry-related side of this stuff. The professor I work for is director of the California NanoSystems Institute, so we hear a lot of stuff about developments in nanolithography and research into making smaller wires and stuff. I think I see a seminar with Moore's law in it at least once a week. (damn people just need to stop showing that slide)

Anyway, kind of curious what the general opinion of molecular electronics is over at Intel. The research group I work in is one of the leaders in the field, and depending on who you ask, the only collaboration that actually has something working. We work with HP a lot, but the device work done in HP labs is crap, the good stuff is done at CalTech by our collaborators, the Heath group.

I know we've had a few visitors from Intel in the past, but the general feeling seems to be they're not interested in molecular electronics at all, at least not at this point. I think there was some work going on with carbon nanotubes over there, but that's all I recall. Do you hear much about molecular electronics and quantum computing?


Zekmyr

Posted: Wed Feb 18, 2004 9:45 pm
by Jonathan
So, I definitely do not speak for Intel. I work for Intel, but my opinions are not necessarily the opinions of Intel.

In my opinion, quantum computing is really going to take off in about 20 years. This will coincide with the hard physical limits of semiconductor technology. I am not personally aware of any efforts within Intel to commercialize devices with obvious quantum mechanical characteristics. I personally have tried to keep up with advances in this area, but not on a professional basis.

As far as I know about Intel, yes, they have demonstrated some sort of carbon nanotube device which they may use in a process at some point. I think perhaps Intel's event horizon is still a little bit shy of molecular electronics, but that's just my opinion. I work in the design group, after all. Furthermore, I'm an architecture engineer, not a circuit guy. What do I know about devices? In my little neck of the woods, we think about the process about to start volume production and maybe the process after that.

Anyway, I think Intel may have ideas about what they're going to do for the next ten years or so. That should get them to the 10 nm range, roughly. Beyond that, I imagine they'll start focussing on molecular devices more.

But don't take my word for it. Like I said, I'm not Intel. I don't speak for them, and my opinions are not necessarily Intel's opinions.

Posted: Wed Feb 18, 2004 10:53 pm
by Zekmyr
Hehe.. of course I realize you don't speak for Intel, I was just curious on a personal level what you've heard and thought. Something you said sounds strange to me though, am I right to assume that you think quantum computing will be commercialized before molecular electronics? Because in the academic community, the opposite opinion is currently the dominant one. Most people think that molecular electronics will be commercialized in the next 5-10 years (maybe sooner for specialized uses), and that quantum computing is like 20-30 years down the road.

Of course opinions have changed about the way molecular electronics will work also. It used to be people thought we'd be building all molecular computers, and stuff, which is just silly. Now it's become more realistic, where we plan to incorporate, for example, molecular memory components into CMOS architectures that will provide the interface to the macroscopic world. We've actually demonstrated several working molecular memory devices. I think the one the device people are currently working on is 1024 bit crossbar memory array (i forget the density, but i think it's around 10-20 nm pitch).

We've actually had a lot of interaction with circuit designers and people like that from HP (Phil Kuekes, maybe you've heard of him? he seems to be pretty famous from doing some fault tolerant computing stuff). It's interesting when they show us all these things they've designed ways to do with just using crossbar junctions with molecules in them. To be honest though, most of it is over my head, since I'm just a simple chemist. =P

As for the nanotube devices, most of the stuff going on in the academic world is purely academic (go figure). Most nanotube devices aren't very practical due to the problems with synthesizing and purifying nanotubes. Of course, even as we speak people (including us) are working to overcome these problems. It's quite possible you could start seeing nanotube-based transistors and sensors being commercialized in the near future, but that's mostly just my personal opinion.

Zekmyr

P.S.- sorry for hijacking your thread, just wanted to kill some time at work and was interested to hear your opinions.

Posted: Wed Feb 18, 2004 11:10 pm
by Jonathan
That's not what I said regarding timeframes. I said ten years out, maybe for molecular devices and twenty for quantum. About in line with what you're saying.

I don't think Intel has plans to include nanotubes in any of the processes on their roadmap yet.

Posted: Thu Feb 19, 2004 12:00 am
by Jonathan
Heh. I produced the data for this chart, though I am uncredited. Woo.

Posted: Thu Feb 19, 2004 4:58 am
by quantus
Dwindlehop wrote:Heh. I produced the data for this chart, though I am uncredited. Woo.
Like I've been told numerous times, make sure the axes and legend are legible. Also, what is HWP?

Posted: Thu Feb 19, 2004 5:14 am
by quantus
I don't remember the specific quote, but roughly what the intel exec said about technology from labs is that it'll be about 7 years at least from when they have a fully functional prototype until it is commercialized. Unless a lab gets extremely lucky in the next year or two, molecular computers are a ways off - longer than than the 10 years you guess.

Molecular memory elements are a bit different from molecular computing imho. There's a big difference between memory and logic. Rarely is logic in the nice even arrays like memory. However, regularity is starting to be designed because even with current technologies, it's getting pretty hard to keep the 'what you design is what you get' reality with a high enough yield (probability) to make money.

Anyone have any comments on what they think the likely path to commecialization will be? Personally, I'm thinking something like an add-in card or networked machine that you batch jobs off to for your tough computations since it'll be a waste to throw away all the current computing and software that's been created. This would allow a smoother transition. Otherwise, we'd be starting from scratch with writing OSes and such all over again. Besides, I don't think I need a molecular computer when typing up an email or letter.

Posted: Thu Feb 19, 2004 7:17 am
by Zekmyr
quantus wrote:I don't remember the specific quote, but roughly what the intel exec said about technology from labs is that it'll be about 7 years at least from when they have a fully functional prototype until it is commercialized. Unless a lab gets extremely lucky in the next year or two, molecular computers are a ways off - longer than than the 10 years you guess.
10 years isn't my guess, it's a pretty common prediction from people in both academia and industry that actually work in the field of molecular electronics. 7 years is maybe what it takes a big company like Intel to market things, but small startup companies get things from lab to product in much less time. And depending on what you mean by prototype, several working molecular devices (mostly memory and sensors) have already been demonstrated. Obviously not packaged and optimized for any sort of application at this point.
quantus wrote:Molecular memory elements are a bit different from molecular computing imho. There's a big difference between memory and logic. Rarely is logic in the nice even arrays like memory. However, regularity is starting to be designed because even with current technologies, it's getting pretty hard to keep the 'what you design is what you get' reality with a high enough yield (probability) to make money.
That's the advantage to using the crossbar architecture that our molecules go into. It's a fault tolerant architecture, so basically you just make the device, test it to see what junctions work, and compile the code down onto the structure you have. I actually saw a pretty cool demonstration of a simulation program some guy is working on to do just that. He can basically take C++ code and compile it onto any system architecture by programming in basic elements of the architecture.

Another interesting thing about the crossbar structure is that you can do memory and logic without really changing the fundamental architecture. You basically have patches of crossbars that do logic, memory and routing.
quantus wrote:Anyone have any comments on what they think the likely path to commecialization will be?
Actually, that's a pretty easy answer. Sensors. Disposable sensing units and things like that. Mostly because everything can be made small and cheap and you can use the molecules to do the sensing and computing. There's a lot of effort working toward that goal now (DARPA is throwing a lot of money at it currently).

That will be the niche that molecular electronics uses to estabilish itself, then my personal guess would be it would work it's way into new types of memory (like SD cards and stuff), and maybe ultimately into specialized processors for specific applications. I doubt it will become a component of general use processors anytime soon, if ever.


Zekmyr

Posted: Thu Feb 19, 2004 7:24 am
by quantus
Zekmyr wrote:That's the advantage to using the crossbar architecture that our molecules go into. It's a fault tolerant architecture, so basically you just make the device, test it to see what junctions work, and compile the code down onto the structure you have. I actually saw a pretty cool demonstration of a simulation program some guy is working on to do just that. He can basically take C++ code and compile it onto any system architecture by programming in basic elements of the architecture.

Another interesting thing about the crossbar structure is that you can do memory and logic without really changing the fundamental architecture. You basically have patches of crossbars that do logic, memory and routing.
Got any papers that describe these processes lying around that could be linked to or maybe sent privately? I could send a private message with my email address.

Posted: Thu Feb 19, 2004 7:31 am
by Zekmyr
I'd have to dig around. A lot of the device architecture work is unpublished because it's not patented yet. I probably have some general stuff I could send you from work tomorrow. PM me your email when you get a chance and I'll send you what I can find.

Zekmyr

Posted: Thu Feb 19, 2004 5:14 pm
by Jonathan
I didn't make the chart, just the data. HWP stands for hardware prefetcher. Pentium 4s have big beefy ones.