First Cell demo (48 MPEG 2 Videos)

[mckmas8808]

From http://www.ga-forum.com/showthread.php?t=44520

Well, might not be what we're hoping for - no game demos - but Toshiba provided a first demo of Cell at the Cool Chips conference that took place last week.

Basically, using a 8-SPE chip, they had 48 standard definition MPEG2 movies decoding and running simultaneously.

To quote one at Beyond3d:

"Toshiba's demo is
1. Load 48 SDTV-resolution MPEG2 streams from HDD simultaneously then decode them with 6 SPEs
2. Another SPE resizes them to thumbnails, then displays them tiled on a 1920x1080 screen. (The remaining 1 SPE is idle throughout the demo) "

Basically, each SPE decodes 8 streams simultaneously. What's perhaps even more interesting is that they did this using software that allowed the programmers to code this without worrying what SPE does what - their software automatically split everything up for them (though this is perhaps easier to do with an application like this than in general).

Pics:

They didn't disclose the clockspeed of the chip doing this. Anyone know how this might compare to a regular PC CPU?

 

[rendezvous]

Why posting this when it so obviously is here allready!?

http://www.beyond3d.com/forum/viewtopic.php?p=505128&highlight=sdtvresolution#505128

 

[a688]

Obviously the cell is faster than a standard PC doing this but thats like asking how fast a dump truck can do the 1/4 mile compared to a sport bike. Doing 48 streams sounds impressive but its only standard definition. How many fewer streams would it be doing on each of the different HD resolutions? In regards to the coding, streaming "static" video from a HDD is a lot different than a game so you couldn't directly compare the ability for the software to automatically split the workload over the SPEs. Granted it works but how well does it work.

 

[The GameMaster]

I already knew the Cell architecture would be extremely adept at encoding/decoding of compressed data, and it should also be extremely adept at distributed computing tasks. I do not see where it was stated that they was using a 8 SPE version of the Cell CPU, and looking at the charts that they provided it only looks like a 4 SPE version (SPE0,1,2,3) and I was not able to find out what clock speed they was using in that demostration. Like I said though this kind of demostration is something I expected as this certainly plays to the strengths of the Cell CPU, the Cell CPU will excell at any streaming application and distributed computing application, but it's performance in general applications and gaming is still very much so questionable. I still await those PS3 specs and benchmarks!

Furthermore I do find it cute that they was using Windows Media Player for this demostration.

The GameMaster...

 

[Shifty Geezer]

I do not see where it was stated that they was using a 8 SPE version of the Cell CPU

Toshiba's demo is
1. Load 48 SDTV-resolution MPEG2 streams from HDD simultaneously then decode them with 6 SPEs
2. Another SPE resizes them to thumbnails, then displays them tiled on a 1920x1080 screen. (The remaining 1 SPE is idle throughout the demo)

8 SPE's

 

[function]

Why posting this when it so obviously is here allready!?

http://www.beyond3d.com/forum/viewtopic.php?p=505128&highlight=sdtvresolution#505128

The text he's copied from the thread at Gaming-Age even says:

"To quote one at Beyond3d:"

 

[mckmas8808]

my bad didn't look and search. That was pretty deep in the thread though, so I should be given a break.

 

[mistan]

The cell is doing this in software mode right?
Because i heard a couple of ppl could run like 7-8 streams before noticing any slowdown, but that would be because they have hardware acceleration specifically for MPEG's right?
I tried to run 1 stream in software mode and my computer crawled..lol
unusable...
So 48 is nice!!
this is a correct assumption right?

 

[ERP]

You should be able to run considerably more than 1 Mpeg2 stream on any PC with a decent processor. I have no idea how you make a reasonable comparison without more knowledge of the test.

Yes this is all in software.

Unless the streams are interleaved in some fashion I think the more difficult part of this is reading the data from somewhere fast enough to be able to decode it.

 

[London Geezer]

*Imagines 48 pr0n vids playing all at the same time on the screen*

gives a new meaning to "adult oriented platform"...

 

[Johnny_Physics]

*Imagines 48 pr0n vids playing all at the same time on the screen*

gives a new meaning to "adult oriented platform"...

How did you reach that conclusion?
Have you taken Struck's Law into consideration? It states something along the lines of how one unit of pornography grabbed from the internet will (roughly) double in complexity(size/decoding time) every 18 months. Calculating a computers pornography throughput is tricky business and not something for amateurs.

 

[London Geezer]

Well anyway, seeing this makes me think:

- 8SPUs Cell actually works. If PS3 is to be released in more than 12 months, i expect this to be cheap enough to be in there, to say the least.

- This just goes to show that processing power has progressed far beyond people's expectations, so much so that everything else in the system is far too slow to keep up with it. Needless to say, optical drives are just way too slow to make this test actually useful (although this test is pretty useless for 99% of real world applications, who would want to watch 48 videos all at once?), and even the fastest memory will struggle too keep up, trying to feed these processors.

- The system will only be as fast as its slowest component.

 

[Shifty Geezer]

As a test subject, to show what Cell's good at, this makes sense. We now know chucking Cell into TVs, production studio mixing workstations, blah blah blah, will provide the media grunt promised. I don't know who the target audience was and why. Maybe it was as much a demo of Toshiba's auto-threading than the processing capabilities of Cell?

It's a plus for Cell, but meaningless for PS3.

 

[London Geezer]

As a test subject, to show what Cell's good at, this makes sense. We now know chucking Cell into TVs, production studio mixing workstations, blah blah blah, will provide the media grunt promised. I don't know who the target audience was and why. Maybe it was as much a demo of Toshiba's auto-threading than the processing capabilities of Cell?

It's a plus for Cell, but meaningless for PS3.

Definately not useless for Cell, and not useless at all for some, i mean i'd love a chip that lets me encode a few MPEG2 files at the same time very fast (though this test was a decoding test). Currently, if i want to encode DIVx files to DVD files, takes all night to complete 7 1.5-2hour movies on my A64 3200+.

 

[Pozer]

That's not such a big deal Marty Mcfly's future version of himself was able to do the same thing in Back to The Future part 2, till needles conned him and Mr. Fujitsu fired both of them.

 

[randycat99]

Doing 48 streams sounds impressive but its only standard definition. How many fewer streams would it be doing on each of the different HD resolutions?

That would probably knock it down to 1/4...so, 12 HD streams is still pretty friggen impressive. Anybody try playing those wmv HD samples on the MS website? How many of those can your PC play simultaneously? Now swap out your fancy, shmancy videocard for some plain vanilla SiS chipset (can't have that speedy DCT acceleration throw off the test, right?). Things slowing down pretty quick, now, eh?

In regards to the coding, streaming "static" video from a HDD is a lot different than a game so you couldn't directly compare the ability for the software to automatically split the workload over the SPEs. Granted it works but how well does it work.

It's just a demonstration of brute processing power. In that context, it seems to be delivering quite a wallop. It's one step past vaporware, one step past hypeware, now, AFAIC.

Obviously the cell is faster than a standard PC doing this but thats like asking how fast a dump truck can do the 1/4 mile compared to a sport bike.

Except, this "dump truck" appears to be leaving the sport bike in the dust, on top of just being a fast dump truck.

 

[Farid]

To quote one at Beyond3d:

 

[PC-Engine]

That would probably knock it down to 1/4...so, 12 HD streams is still pretty friggen impressive. Anybody try playing those wmv HD samples on the MS website? How many of those can your PC play simultaneously? Now swap out your fancy, shmancy videocard for some plain vanilla SiS chipset (can't have that speedy DCT acceleration throw off the test, right?). Things slowing down pretty quick, now, eh?

Decoding MPEG2 at HD resolution isn't the same as decoding VC-1 at HD resolutions...

Decoding MPEG2 at HD resolutions just requires higher bitrates. Decoding VC-1 at HD resolutions requires much more processing power as your HDWMV demos show. You only need a Pentium II 300MHz to decode DVDs with audio. A single threading 3.5Hz P4 will be able to do at least 12 streams with audio like aaaaa00 mentioned in the other CELL thread. A dual core P4 with HT at 3.5GHz could probably do around 40 DVD streams.

 

[randycat99]

Now that seems a bit farfetched, but I do grant you the point between mpeg2 and mpeg4. There's probably a balance between there somewhere, anyway. Processing twice the data in mpeg2 vs. x amount of data that is y more intensive to process in mpeg4...either way you go, 12 HD feeds with scaling, tiling, and compositing is still no meager feat.

 

[one]

A single threading 3.5Hz P4 will be able to do at least 12 streams with audio like aaaaa00 mentioned in the other CELL thread.

Unless it's verified that it's scalable linearly to 12 streams with no hiccups in all movie frames, it'd be safe not to assume like that...

BTW, in this Cell demo, there is an OS overhead, just like on Windows. The PPE in Cell runs OS services to secure responsiveness while SPEs are working, which is a big difference from Pentium4.