Toshiba Cell demo @ CEATEC JAPAN 2005

[randycat99]

I'm curious (yield-wise) how many 7-SPE Cells occur to achieve x amount of 8-SPE Cells for server product. The server product is certainly going to be charged at a premium, so that should make all the tiers below essentially "on-the-cheap", since a good amount of them are going to be simply by-products of trying to acquire the top tier product...

 

[MfA]

Wow. Thats a pretty cool way to implement Security on a hardware level. I wonder how it works though? If the system detects its being compromised it isolates that core?

I assume code would be uploaded to the SPU during initialization ... after which you could only stream data to it and gets result.

For instance you could have a bit of ROM with a routine which puts the SPU in isolation mode, and then decrypts _crypted_ code and data stored in local memory ... after which it would accept streamed data, and output the results. Would be a pretty good way of implementing DRM in fact.

 

[one]

Though this is a Cell demo, do we have anything to compare it to on non-Cell systems? In other words, what is it about Cell that makes it able to do this and how much is it having to do?

Though the magic mirror demo lacks implementation details, for 48 MPEG2 streams decoding demo, according to
http://techon.nikkeibp.co.jp/english/NEWS_EN/20051005/109352/?ST=english
they are 48 MPEG-2 MP@ML streams. Also according to
http://www.watch.impress.co.jp/av/docs/20051005/ceatec11.htm
1 stream is 720x480 @ 4Mbps, so you can conclude 1 stream is 720x480 @ 4Mbps / 30fps. 6 SPEs decode them and 1 SPE scales them into thumbnails.

 

[one]

A Famitsu article on the demo
http://www.famitsu.com/game/news/2005/10/05/103,1128498917,44271,0,0.html
It maps 500 3D coordinate points on a face then scans them from a face more than 100 times / sec. Also applicable not only to a face but also to a whole body. The page also reports about 3D display, digital pen and paper, and humanoid robots.

 

[macabre]

Mercury sytems has a new video on their site with cell performance in medical imaging:

http://www.mc.com/cell/demo.cfm

 

[Titanio]

Mercury sytems has a new video on their site with cell performance in medical imaging:

http://www.mc.com/cell/demo.cfm

Cheers! No hard data, but it's really nice to have a visualisation of the performance improvement like that. Very fast!

 

[BlueTsunami]

Mercury sytems has a new video on their site with cell performance in medical imaging:

http://www.mc.com/cell/demo.cfm

That was pretty cool to see that. The visual example of how fast Cell was (for that specific task) was very interesting to see.

 

[_phil_]

nice.it also contains a true picture of the true cell chip (near the end).Nice to clean up these boards from at least a threat...

 

[Shifty Geezer]

Good link. Just watching it now. The first half is all marketting spiel, but they've got to a realworld example of Cell doing work starting around 2:45. Highlights :

Apparently rendering on imaging 'slice' of data requires 'several gigaflops' and takes a 'typical processor found in a high-end server' 2 seconds to reconstruct one slice. A full scan takes 5 minutes. This process is shown slowly taking place in the background. Then Cell is show 'on exactly the same problem' and completes it in 3 seconds, maybe less.

"Mercury thinks of Cell as a multicomputer on a chip."

No specific details of the high-end PC chip or the Cell configuration such dual Cell, (single Cell? He says the Cell chip when giving the example) or clockspeeds. But still that's a realworld improvement of several 100 times, and in seemingly little time in software development. Nice to see people getting to grips with Cell's architecture and getting real perforamnce from it, especially compared to existing chips. This is the first Cell demo that shows what a PC counterpart is like, and in the case how much faster Cell is.

 

[mckmas8808]

This is really nice to see. Look like Mercury is really to push their products to use CELL technology. And really how can you turn it down when it is (in that example) 150 times faster when running the same solution?

 

[liverkick]

Neat stuff.

Hmmm.

*lightbulb*

New marketing campaign for PS3. "Powered by the technology that SAVES LIVES!".

On another salient tangent, that voiceover for the Mercurary Systems corporate ad in the beginning sounded disturbingly just like the ones in Doom 3. I think it was the same exact chick actually. I was half expecting a portal to hell to open up in the cat scan machine.

OK, Im done raising the level of discussion in this thread.

 

[mckmas8808]

Neat stuff.

Hmmm.

*lightbulb*

New marketing campaign for PS3. "Powered by the technology that SAVES LIVES!".

Yeah good slogan, too bad it's kind of cheap though. Seems a tad politcally incorrect doesn't it?

 

[liverkick]

Yeah good slogan, too bad it's kind of cheap though. Seems a tad politcally incorrect doesn't it?

It was a joke mckmas.

Although the depraved souls at SCEE marketing could probably make it work.

 

[drpepper]

Damn it, I hate WMVs! It keeps timing out and "buffering" when I it the 52 second mark. Why can't MS make their streaming stable on Mac platforms?

 

[mckmas8808]

Damn it, I hate WMVs! It keeps timing out and "buffering" when I it the 52 second mark. Why can't MS make their streaming stable on Mac platforms?

You have to make it work man. It is too cool not to see. (Just messing wit cha)

 

[silhouette]

I have not seen the video, but if it is MRI (it is from what it sounds), CELL is exactly designed for this kind of applications. In MRI, the output of the sensor is actually a 2D FFT of the actual image. So, to get the actual image, you need to apply a 2D inverse-FFT to the output of the sensor. And, I do not have to say that 2D FFT's are highly computationally complex. If your image is -say- NxN, you have to make N*N FFT's each of whose complexity is N*log2(N). So, you do the math.. For high-res images, the complexity becomes enourmous.

Now, since you do the FFT's first row-wise, then column-wise, you can partition the data first into 8 segments, load each of these segments into a different SPE and do the row-wise FFT computation simulatenously in all 8 cores, move the data around the SPEs for column-wise operation, do the column-wise FFT operation, finally combine output data in each SPE to form the final image... In addition, FFT is also very well suited for SIMD architectures as well.

It is not suprising that CELL accelerates this operation significantly.

 

[BlueTsunami]

I have not seen the video, but if it is MRI (it is from what it sounds), CELL is exactly designed for this kind of applications. In MRI, the output of the sensor is actually a 2D FFT of the actual image. So, to get the actual image, you need to apply a 2D inverse-FFT to the output of the sensor. And, I do not have to say that 2D FFT's are highly computationally complex. If your image is -say- NxN, you have to make N*N FFT's each of whose complexity is N*log2(N). So, you do the math.. For high-res images, the complexity becomes enourmous.

Now, since you do the FFT's first row-wise, then column-wise, you can partition the data first into 8 segments, load each of these segments into a different SPE and do the row-wise FFT computation simulatenously in all 8 cores, move the data around the SPEs for column-wise operation, do the column-wise FFT operation, finally combine output data in each SPE to form the final image... In addition, FFT is also very well suited for SIMD architectures as well.

It is not suprising that CELL accelerates this operation significantly.

Thats exactly what their doing and the end result was a (spinal column?) being rendered very very fast compared to the server cpu counterpart. If the advancement is truely as they showed then its quite a big one.

As Shifty stated though, i'm wondering what class of Cell their using (not really class but if its two Cells, a 3.2GHz Cell etc).

 

[Carl B]

Thats exactly what their doing and the end result was a (spinal column?) being rendered very very fast compared to the server cpu counterpart. If the advancement is truely as they showed then its quite a big one.

As Shifty stated though, i'm wondering what class of Cell their using (not really class but if its two Cells, a 3.2GHz Cell etc).

If it's their present blade showpiece, then it'd be a dual Cell system at 3 GHz. Ideally what we saw was a single Cell vs a single competing chip, or if the dual processor blade was used, a comparable dual processor setup on the part of the competition was used as well.

The Cell clearly has value and utility in the imaging fields, no matter how stacked the decks may or may not have been, what I saw was very impressive. I'd be interested to know what the competing 'mystery chip' was though.

 

[Vince]

Thats exactly what their doing and the end result was a (spinal column?) being rendered very very fast compared to the server cpu counterpart.

It's not exactly the same as this was CT which uses a FBP, a discrete form of the Radon transform, and not an MRI. But the basics relating the the 3D visualization speed-up are the same and one would expect a similar speed-up in generalized MRI and fMRI uses.

 

[mckmas8808]

It's starting to become easy to see how and where CELL can be used in an everyday environment. If this is what Ken Kutargari and IBM meant then I'm excited to see who and where else the CELL will be used.