Intel's "Tera-Scale Computing" demo @ IDF (funny)

[pxc]

If you haven't heard about it before, it's Intel concept for "many core" CPUs in the future. Think dozens or even hundreds on a die. Intel's page about it is here: http://www.intel.com/technology/techresearch/terascale/

I saw a link this morning to theinq, where they have pictures of the platform Intel demonstrated it on at IDF today. It's running in an FPGA to test the concept, but I thought it was funny that it ran on an ancient 430HX motherboard (socket 7, > 10 years old) and it did run WinXP at a clock speed under 2MHz.

http://uk.theinquirer.net/?article=34623

 

[brucedeluxe169]

I'm pretty unimpressed by that whole many-core concept.....

I just wish we really were running 15ghz Northwood style CPUs right now... you know those things would be crazy fast!

*damn you physics!*

 

[pxc]

Software lags hardware by years sometimes. Of course it's going to take a while to fully utilize just a 4 core CPU in many more software cases, but it will happen.

Sun's Niagara is already showing what a modest multi-core design (8 x 4 way SMT cores) can do today, in terms of performance and relatively low power consumption (72W).

If Intel is able to make Tera-Scale cores based on a Core 2 derivative as planned, it will be very impressive.

 

[Poncho]

I just wish we really were running 15ghz Northwood style CPUs right now... you know those things would be crazy fast!

*damn you physics!*

Honestly... I would like to see a 45nm Netburst based CPU. I bet you could get 6+ghz out of it with some work.... and the smackage would come. It's a shame that they had all the leakage problems with it.... since had they hit their goals of 10ghz+ it would be INSANELY fast. Oh well... onto better things I guess. LOL

 

[TType85]

I can't wait for the 4+ core processors hit the desktop. Using my C2D 6300 @ 2.55ghz and the newest Nero Recode I can go to Cinema AVC at 28-30fps.

I have 350 movies that I want to convert (mp4 video, ac3 sound, subtitles). Last night I did 5, reclaiming 30-50% of the space they were taking with no real noticable loss in quality on my 32" LCD. I was down to under 200GB free out of 2TB, so the faster I can re-encode the better!

Bring on the 4-cores!

 

[eno-on]

multi core computers are where its at. It's important to be able to process MANY strings at the same time. Our massively parallel brain is why we can recognise patterns almost instantly. Face recognician, being able to take in the contents of a room at a glance, being able to pick one voice out of a crowd of peopel talking and understand what they say. It's due to massive parallelism. Current computers cant do it well, because they cannot process millions of bits of information simultaneously.
1000 processors running at 1mhz will be exponentially better at these things than 1 processor running at 1000mhz.

 

[pxc]

Info on another terascale research product: http://www.eet.com/news/latest/showArticle.jhtml?articleID=193005741

One of the first prototypes chips in Intel's program is the Tflops/s processor. Containing 80 simple cores and operating at 3.1 GHz, the goal of this experimental chip is to test interconnect strategies for moving terabytes of data from core to core and between cores and memory, according to Intel.

Unlike existing chip designs where hundreds of millions of transistors are arranged, this chip's design consists of 80 tiles laid out in an 8- by 10-block array, according to Intel.

Each tile includes a small core, or compute element, with a simple instruction set for processing floating-point data, but is not Intel x86-based processor compatible. The tile also includes a router connecting the core to an on-chip network that links all the cores to each other and gives them access to memory.

 

[Lyquist]

This is really cool. I like reading about advances in technology.

 

[Brent_Justice]

80 core CPUs, geez

 

[eno-on]

80 core CPUs, geez

These will increase exponentially, too, once the process gets tidied up.
Which means, of course, performance will increase exponentially.
Once we get past the hurdles of keeping large multicore units talking freely with each other, we'll see exponential growth in the number, and density, of 'cores' on a 'cpu'.
At that point, we probably WILL have computers-on-a-chip.
Cpus arent good at rendering 3d video games, because they arent parallel. GPUs are.
With massivley parallel 'programmable' cores (meaning they can change types of tasks on the fly, which they will be able to do), we will have our CPUs doing just about everything other than storage.

 

[Slider19]

Honestly... I would like to see a 45nm Netburst based CPU. I bet you could get 6+ghz out of it with some work.... and the smackage would come. It's a shame that they had all the leakage problems with it.... since had they hit their goals of 10ghz+ it would be INSANELY fast. Oh well... onto better things I guess. LOL

I would rather see more work done per clock then an insanely high clock speed.

It's impressive to see 2MHz cores run XP.

 

[pxc]

This is really cool. I like reading about advances in technology.

The slides from presentations are available: http://www28.cplan.com/cv125/sessio...et_audience=+&idate=+&isession_id=&iabstract= There are several interesting topics covered.

l/p: idf/fall2006

 

[Arcygenical]

multi core computers are where its at. It's important to be able to process MANY strings at the same time. Our massively parallel brain is why we can recognise patterns almost instantly. Face recognician, being able to take in the contents of a room at a glance, being able to pick one voice out of a crowd of peopel talking and understand what they say. It's due to massive parallelism. Current computers cant do it well, because they cannot process millions of bits of information simultaneously.
1000 processors running at 1mhz will be exponentially better at these things than 1 processor running at 1000mhz.

Only if they can properly combine their efforts to focus on one task. I say this because, as great as our brains are at determining changes in patterns, we lack the ability to perform complex math equations at any useful speed.

I would have to (tentatively) disagree with your statement for one simple reason, environment. Our environment is directly responsible for our massive parallelism. Being able to complete a really difficult physics equation doesn't serve as much of a purpose to us as picking a possible predator out of a camouflauged landscape, or instantly determining whether we find a member of the opposite sex to be attractive.

The same goes for computers. Facial or pattern recognition might be very important to certain applications, but for the most part, processors are simply translating digital data into events, times, spaces or visualizations that we can interpret. The amount of computer power required to translate a digital media file into audio that our ears can understand is, actually phenominal. The majority of computer tasks that we do today do just fine with one, or two cores (and will run especially well with 8-16 cores) as our minds can only focus on one or two main things (which is quite funny considering how parallel we really are). Eventually we're going to get a huge reduction in the benefit/cost curve, as only certain applications will benefit from such a parallel cpu-space.

If the "1000 core CPU" Could effectively break up digital data into 1000 equal parts (or however many are required), interpret, concatencate the outputs and display/produce analog output data, then yes, your theoretical processor would be substantially faster. To do this, Intel would need to make sure that incoming data is equally acessible to each core, without impeding the performance of other cores and dedicate a large portion of the execution centres to reducing the inevitable overhead.

Eventually a line must be drawn between sheer number, and power of each individual, cores on the Die. The sweet spot between millions of execution centres and a hundred or so more powerful cores, would produce, as you said, the "perfect processor".

What's amazing to me is how quickly the industry sprang into action. We've been using single core processors for YEARS, then within the last... 1 and a half years, we've progressed onto almost 4 cores/die. By next year we should be closer to 8 cores. That's an 800% increase in 2.5 years... Or about 10% of the total time CPU's have been "mainstream."

**EDIT** I guess that dev isn't so worried about ESD

 

[Poncho]

I would rather see more work done per clock then an insanely high clock speed.

It's impressive to see 2MHz cores run XP.

If you look at the direcetion and the original goals of Netburst.... a P4 @ 10ghz would have killed anything on the market. Netburst scales INCREDIBLY well... the problems came from leakage. Solve that... and you'd have a great CPU. I agree it's a dead technology... but it had a lot of potential.

 

[eno-on]

Only if they can properly combine their efforts to focus on one task. I say this because, as great as our brains are at determining changes in patterns, we lack the ability to perform complex math equations at any useful speed.

I would have to (tentatively) disagree with your statement for one simple reason, environment. Our environment is directly responsible for our massive parallelism. Being able to complete a really difficult physics equation doesn't serve as much of a purpose to us as picking a possible predator out of a camouflauged landscape, or instantly determining whether we find a member of the opposite sex to be attractive.

The same goes for computers. Facial or pattern recognition might be very important to certain applications, but for the most part, processors are simply translating digital data into events, times, spaces or visualizations that we can interpret. The amount of computer power required to translate a digital media file into audio that our ears can understand is, actually phenominal. The majority of computer tasks that we do today do just fine with one, or two cores (and will run especially well with 8-16 cores) as our minds can only focus on one or two main things (which is quite funny considering how parallel we really are). Eventually we're going to get a huge reduction in the benefit/cost curve, as only certain applications will benefit from such a parallel cpu-space.

If the "1000 core CPU" Could effectively break up digital data into 1000 equal parts (or however many are required), interpret, concatencate the outputs and display/produce analog output data, then yes, your theoretical processor would be substantially faster. To do this, Intel would need to make sure that incoming data is equally acessible to each core, without impeding the performance of other cores and dedicate a large portion of the execution centres to reducing the inevitable overhead.

Eventually a line must be drawn between sheer number, and power of each individual, cores on the Die. The sweet spot between millions of execution centres and a hundred or so more powerful cores, would produce, as you said, the "perfect processor".

What's amazing to me is how quickly the industry sprang into action. We've been using single core processors for YEARS, then within the last... 1 and a half years, we've progressed onto almost 4 cores/die. By next year we should be closer to 8 cores. That's an 800% increase in 2.5 years... Or about 10% of the total time CPU's have been "mainstream."

**EDIT** I guess that dev isn't so worried about ESD

It WILL be important for us to have computers that can recognize the things we can.
Why? Because they WILL be able to.

They will be able to enhance our own awareness of things. A processor that is massivley parallel, as well as many times better at doing math, will be able to recognise patterns in smaller amounts of time than our own brains at that point.

What does that mean? It means we will be better at predicting devastating storms, better at seeing financial trends, better at designing more intricate machinery and electronics, better at designing highways, freeways, and city street traffic lights. Better at predicitng earthquakes, socio-economic changes, insurgencies... I mean, the list goes on.

All of these things have PATTERNS that lead up to the event. We don't have computers figure these things out for us, because they suck at it.

These things require massivley parallel processing power, which our brains do. And yes, while we can only focus on one or two things at a time very well, there are still LOTS of things going on that influence our ability to recognise patterns, all happening at the same time. We dont NEED to focus on those individual tiny things. Our parallel brain is taking care of it for us while we focus on the immediate.

Of course the software will be required that can "effectively break up digital data into 1000 equal parts".

The lovely things about these future multi processors is that they will be "universal" cores. They will be able to configure themselves to take on different kinds of tasks. Also, if a certain kind of task requires something more complex than a single core can handle, they will work in conjunction with another to make the circuit that piece of software requires. They will be able to emulate x86 or whatever at basically a hardware level (although I don't assume they'd want to.)

It will also be an important step in designing true AI's.

It's gonna be awesome, and it's gonna be sooner than most people think.

 

[alan.p]

If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?
- Seymour Cray

That's about the first thing that comes to mind when I read this. Plus some problems just don't scale to multiple processors that well. I'm not slamming the basic idea, just pointing out that heavy duty singlecore still has it's place.

 

[eno-on]

If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?
- Seymour Cray

That's about the first thing that comes to mind when I read this. Plus some problems just don't scale to multiple processors that well. I'm not slamming the basic idea, just pointing out that heavy duty singlecore still has it's place.

I'd use the chickens, just for the sake of absurdity.

With the way these things will be, the 1024 chickens will be able to act just like the 2 oxen if the need arises. Thats the beauty of these generic, universal cores. They can line up to do whatever you want em to.

 

[swatbat]

It's impressive to see 2MHz cores run XP.

Yea that is pretty interesting. Would be real funny is they had like quake or something running on it.

 

[Slider19]

If you look at the direcetion and the original goals of Netburst.... a P4 @ 10ghz would have killed anything on the market. Netburst scales INCREDIBLY well... the problems came from leakage. Solve that... and you'd have a great CPU. I agree it's a dead technology... but it had a lot of potential.

You can also have a great CPU like the Conroe which has a much lower clock speed, with far greater performance. I don't see any advantage for a high clocked CPU. 10GHz P4s would have also likely killed the power grid.

And its just a number, what should matter is performance. I wonder at what clockspeed would a Conroe have to run to beat a 10GHz P4. Any ideas?

 

[InternalStorm]

You can also have a great CPU like the Conroe which has a much lower clock speed, with far greater performance. I don't see any advantage for a high clocked CPU. 10GHz P4s would have also likely killed the power grid.

And its just a number, what should matter is performance. I wonder at what clockspeed would a Conroe have to run to beat a 10GHz P4. Any ideas?

A wild guess would be 5GHz... I wouldn't say any more, maybe even less. It's just that you can't say a Conroe is better at everything compared to a Netburst CPU, as they both have their strengths I suppose... higher clockspeed can do some things better than lower clock speed.

 

[Poncho]

You can also have a great CPU like the Conroe which has a much lower clock speed, with far greater performance. I don't see any advantage for a high clocked CPU. 10GHz P4s would have also likely killed the power grid.

And its just a number, what should matter is performance. I wonder at what clockspeed would a Conroe have to run to beat a 10GHz P4. Any ideas?

Mypoint<--------- -------------->You



You're missing what I am saying. Yes... I know Conroe is better, yes I know it uses less power. What I am saying is that the original GOALS of netburst, because of it's very nature and ability to scale well, was going to be 10ghz+. When they had thought this up... they didn't realize the leakage would be so bad, thus the increased power consumption. A 10ghz P4, that didn't have leakage problems and was @ 80 Watts wouldn't have killed any power grid.... just the competition. You are right... it's just a number, but with netburst this number WAS performance. It's clockspeed had a greater impact on performance than AMD and Conroe. I would have like to seen netburst work the way it was planned... I think it would have been impressive. That's all.

 

[Slider19]

OK, I get your point now, but personally speaking I just never liked the netburst architecture much and its heavy emphasis on only pure GHz scaling. If we can get similar performance from Conroe or K8L, I'll be content.

 

[savantu]

I would rather see more work done per clock then an insanely high clock speed.

There is a limit on how much ILP you can extract from x86 code.And Core march is reaching that level already.The gain from 3 to 4 way decoding is less than 5% for 99,9% of the time and it added a great deal of complexity.

3 options left to increase performance: either change the ISA , pump up the frequency or improve TLP ( n core madness )