Toshiba Demonstrates Cell Microprocessor

Cybro writes "Toshiba has demonstrated some cool applications for the Cell Microprocessor. They also revealed that they have written their own OS for the new processor. However the article on TechOn does not reveal the license of the OS."

Comments

Background For Those Unfamiliar With Cell

[tquinlan]

http://arstechnica.com/articles/paedia/cpu/cell-1. ars
http://arstechnica.com/articles/paedia/cpu/cell-2. ars
http://www.blachford.info/computer/Cells/Cell0.htm l

ARTICLE TEXT:

[Anonymous Coward]

Toshiba Demonstrates Cell Microprocessor Simultaneously Decoding 48 MPEG-2 Streams
Apr 25, 2005 14:15

Toshiba demonstrated that its Cell microprocessor, jointly developed with the Sony Group and IBM Corp., can simultaneously decode 48 SDTV format MPEG-2 streams. At the COOL Chips VIII event held in Yokohama from April 20 to 22, 2005, the company showed a film demonstrating the decoding process.

In the film, 48 MPEG-2 streams stored on a HDD were read, decoded and projected onto a 1,920 x 1,080 resolution display divided into 8 x 6 cells, each of which showed a different video in each cell. The company expects the technology to be used to display thumbnails for a video list. Of the eight synergistic processor elements (SPE) used in the Cell, six are used for decoding 48 MPEG-2 streams and one is used for scaling the screen. The remaining SPE can be used for a completely different processing function.

In the demonstration, Toshiba used an operating system environment it had developed to increase the efficiency of Cell software development. One of the environment's key features is that application software developers can program software without considering which threads will be allotted to each of the different SPEs, because the environment allows the automatically scheduling software to SPEs.

Not much info

[Ironsides]

It doesn't say anywhere what the bitrates of the originating SD streams were. That is a biggie in terms of processing power. MPEG2 can run from 1.5Mbps (crap) to 50 Mbps (I Frame only, dam good) and higher. Give me more info and I might be impressed.

Re:Not much info

[benwaggoner]

Actually, the I-frame only is quite a bit easier to decode, since it doesn't require motion estimation, doesn't reference multiple frames, etcetera. It's almost identical to a series of JPEGs at 29.97 frames per second.

In normal IBBP MPEG-2, a given B-frame will reference frames either direction of it, requiring two frames be buffered in memory, and a lot of moving around of data from those two frames.

WMP?

[oiarbovnb]

I swear the screenshot looks like it is WMP... I can't believe that windows is going to work on the CELL. Anybody have any info on what OS will be supported (other than the Toshiba OS?)

Re:WMP?

[Anonymous Coward]

It said "the company showed a film demonstrating the decoding process"

Re:WMP?

[vectorian798]

You are right that it looks EXACTLY like WMP in full-screen mode. However, it seems unlikely that they ported it over to the weird OS they have going. Most likely they decode the streams and send raw data streams over a fast gigabit ethernet over to a comp with a media player that just chugs it through onto the screen. Or something like that anyways.

Re:WMP?

[Jozer99]

Possilby, they did not allow cameras into the demo. They did provide a video, and the picture is simply an amerature screenshot of WMP playing said video.

Re:WMP?

[jnaujok]

They say in the article that this is a "video of a wide-screen monitor". In other words, the Cell processor ran this video to a big LCD display, and someone filmed it with a camera. Then the film of that display was played back at the demo.

This isn't that uncommon, in order to hide the fact that the ball of "wire-spaghetti" sitting on the bench below the monitor is the prototype system. It might also hide the liquid nitrogen pump they were using to cool it...

In other words, they wanted to hide something, which could be as innocuous as a messy lab environment.

I am still

[Anonymous Coward]

totally baffled how one could write something for the cell that we would traditionally call an "OS". At least, a time sharing OS. Who gets to use the SPE/APU/SPUs, and when? The attatched memory on the SPEs is nontrivial to swap to memory, and it seems absurd to think that it would just be done offhand with a context switch. Yet, context switches must happen. So are SPEs merely given to processes, who get to keep them, so that the main processor is switching betwen processes normally in a preemptive style but the SPEs stay under the control of single processes?

Or is the Cell OS Toshiba's using here non-multitasking or cooperative multitasking? Or what?

Re:I am still

[adam31]

I am also baffled. But the SPE has a memory size of only 256kb and 25 GByte/s bandwidth to main memory, so it's not like a context switch is out of the question. Also I'm pretty sure that each SPE can run 2 threads SMP (or is that just the Power chip?)

It probably is a combination of the method you describe (where a SPE is dedicated to a thread) and traditional pre-emptive... And it probably boils down to the more processor-intensive threads get their own SPE, while a couple SPEs are dedicated to context-switching threads.

To me, that part is not as baffling as how programs are going to be written in the first place. A thread will have to communicate all the memory accesses it will need to make to prefetch all that data. On the PS2, we use environments like this, where that memory would be double- or triple-buffered so we can simultaneously stream, process, and write-back... but that's hand-written individually for the hardest-core processing pieces, where memory accesses are predictable and sequential. How to do this in a general way to make the process easier on application developers?

That's never really been a concern for Sony...

Re:I am still

[_generica]

No,

Each SPE can only run one thread.

The POWER core _can_ run two threads, but it is recommended to only run one.

Re:what does it do?

[A Dafa Disciple]

Video Games

Huh?

[garcia]

Well, in this picture I see a movie file being played (on what seems to be WMP) showing the cells on the screen.

Now, I wasn't there, nor was the article really in depth by any means, but it would seem to me that this was nothing more than a movie demonstration and nothing live.

I'm not quite so impressed. Maybe we should start linking to real content from the front page (i.e. in-depth accounts and not some blogger's one page summary with a blurry photo of a movie file being played on a projection screen).

Re:what does it do?

[beef3k]

No, none whatsoever.

The PlayStation 3 will be nothing but a DVD player capable of playing back 150 movies at the same time.

Which is kind of sad really, I would have hoped for more.

perhaps more interesting

[Anonymous Coward]

Ken Kutaragi Talks about Cell

http://techon.nikkeibp.co.jp/english/NEWS_EN/20050 407/103542/

Re:perhaps more interesting

[birge]

Agreed. I found that the better article, as well. Especially interesting was this bit:

Q: Cell has 8 embedded "SPE" CPU cores. What is the basis for this number?

A: Because it's a power of two, that's all there is to it. It's an aesthetic. In the world of computers, the power of two is the fundamental principle - there's no other way. Actually, in the course of development, there's this one occasion when we had an all-night, intense discussion in a U.S. hotel. The IBM team proposed to make it six. But my answer was simple - "the power of two." As a result of insisting on this aesthetic, the chip size ended up being 221mm2, which actually was not desirable for manufacturing.

The interesting thing to me here

[Anonymous Coward]

Isn't even *exactly* how impressive a multitasking feat it was, but that if I'm reading the article right, the 48-stream decoding thing was done *entirely by the auxillary processor units*. That is, the "SPE"s. The main [PPE] processor in the Cell was apparently not really doing anything at the time. This seems to bode wel for the usability of the SPEs.

Re:The interesting thing to me here

[marcansoft]

That is *exactly* what they are there for. The real power of the Cell comes from the SPEs, not from the main PPC which is mainly there to control them and move data around. The SPEs do the number crunching.

This does prove though that each SPE can handle 8 MPEG-2 streams with no problem (equivalent; probably they are used in a pipeline fashion so really saying one "does" 8 streams isn't too real, but they can do 48 streams as a whole. The typical thing is one for each major step of decoding.)

Re:The interesting thing to me here

[imsabbel]

Have you looked closer at cell?
those "auxillary processor units" make up 75% of the die and have 95% of the computing power of cell.

Mebbe it can power

[KingBahamut]

this............
http://www.lod.org/Projects/Other//index.htm

That Cell microprocessor is pretty good...

[TripMaster Monkey]

...but the smart money's on the Goku miccroprocessor.

^_^

Business Idea

[3770]

Here is a business idea for some small to mid size hardware company.

The CELL processor is cool and the geeks love it and it is based on the POWER architecture. Surely, it'll run Linux.

Build a machine with the CELL. Don't follow any standards (well, use PCI and PCI express Serial ATA and USB 2.0 and stuff like that). But just make sure that you are first out the door with a box.

And make it cheap. It must be possible to make it cheap since it will be sold in the PS3.

I bet that there'll be a lot of enthusiasts that will buy it and be early adopters which will help you work out the bugs.

And then, a year after your first release you'll have a computer that is very fast for its price and a system which is source code compatible with the largest source code library in the world.

Well, I know I'd consider buying one.

Re:Business Idea

[LWATCDR]

"and make it cheap. It must be possible to make it cheap since it will be sold in the PS3."

Nope you will have to develop a chipset for it and that is not cheap. It is unlikely that Sony will hand over the supporting chipset for the PS3 to anyone. Cheap comes from big numbers and there is no such thing as a lot of enthusiast. At least when we are talking about the number it would take to make this cheap.
The only two I can see doing this are IBM and Apple.
A Cell based Mac mini as the ultimate gaming platform would be interesting. I could see this as part of IBMs plan of world domination. They dump the Intel line on china and then push for the Power/CELL to replace it. A Cell based workstation running Linux or even Windows talking to a Cell/Power based server. All with IBMs blessing. IBM goes back to being the master of it's own destiny with no real need to make nice with Microsoft or Intel.

The Perfect OS for the Cell Processor

[MOBE2001]

I got the perfect OS for the Cell Processor. I just need funding. :-D

The COSA Operatin System

See also the link below.

RTFA please: (to those talking about Windows, WMP)

[llamalicious]

"... the company showed a film demonstrating the decoding process.

In the film, 48 MPEG-2 streams stored on a HDD were read, decoded and projected onto a 1,920 x 1,080 resolution display divided into 8 x 6 cells, each of which showed a different video in each cell. The company expects the technology ..."

If you bothered to read the article contents, you'll see that they simply showed a video of the process actually working.

This is a far cry from a live tech demo, but if they can really pull it off, definitely shows the power of a Cell.

Maybe ToshibaTVs can do 720P correctly with this.

[PornMaster]

Wouldn't it be nice to combine this article and the previous one and have a Cell to make your 720p TV display 1080i content properly? :)

Threads and units

[Spy der Mann]

From TA:

In the demonstration, Toshiba used an operating system environment it had developed to increase the efficiency of Cell software development. One of the environment's key features is that application software developers can program software without considering which threads will be allotted to each of the different SPEs, because the environment allows the automatically scheduling software to SPEs.

Now *THIS* is the interesting part on their OS. Because the SPEs have different kinds. When I looked at the cell architecture, I thought: "Programming for this thing is going to be a MAJOR MESS!"

Thumbs up for Toshiba on figuring this out AND doing something about it.

OS for Cell and Cell Application

[vectorian798]

For all those talking about what kind of OS Cell will run:
ZDNet Article on Cell

The article is dated in some ways (like when it says 16 cores...I believe it is half that right?) but it does point out some interesting things, for example, like the fact that there will be a Cell SDK and a end-user OS aimed at embedded devices and the like.

If you google around for 'cell forums' you will come across interesting discussions where they point out that linux will be ported very quickly to cell and that IBM has hinted at possible uses for Cell as a workstation. Also, Cell is OS NEUTRAL meaning that it does not have any particular hardware functionality that makes one OS run any faster.

Overall, I would say that since market penetration is needed, you can't just say "Here is our OS and our SDK, use it from now on". The trick will be of course, to assuage the existing target audience who use today's OS's.

Also, note that the Cell is not a processor bred entirely for the PS3 or anything like that - it will be embedded in devices such as PVR's, TV's, music players, and in all likelyhood, it will even find its way to the desktop - with its potential it is likely to also find some niche in supercomputing since it will be cheap (if 4 whole cells can be thrown into a game console why not?)

Probably not an "OS" in the common sense

[Lemming Mark]

[ disclaimer: this is speculation but it's informed speculation - hopefully useful ]

It's worth bearing in mind this is unlikely to be an OS in the common sense. I'd rate it very unlikely that this OS supports such niceties as filesystems, network IO stacks, protected processes, etc - or that it ever will.

Rather, it's likely to be a shim (albeit a clever one) for insulating the developers of embedded-style applications from the real hardware. I wouldn't be surprised if this Toshiba OS is actually a "library operating system" which is linked into the application itself.

Don't think of it as an OS in the Linux sense, more as a toolkit / library for Cell programmers. Exactly how a "conventional" OS will run on the Cell is not clear to me but it seems certain that it can support a Linux-style OS well - otherwise it'd scupper Cell's World Domination plans ;-)

Related links on the Cool Chips VIII

[News for nerds]

Cool Chips website with the program
http://www.coolchips.org/

Another report for the conference (in Japanese, with pics)
http://pcweb.mycom.co.jp/articles/2005/04/28/coolc hips1/
http://pcweb.mycom.co.jp/articles/2005/04/28/coolc hips1/001.html

Maybe

[News for nerds]

this image can explain a lot for you?
http://pcweb.mycom.co.jp/photo/articles/2005/04/28 /coolchips1/images/016l.jpg

Comment removed

[account_deleted]

Comment removed based on user account deletion

PS3 SDK?

[dioscaido]

Has there been any info leaked about the PS3 SDK? Programming a game on a massively parallel platform like the CELL can't be easy, especially for a console industry used to a fairly traditional hardware environment.

In many respects, the Saturn failed because the SDK was just too hard to work with, as did the N64 (although that also had the cartridge limitation to further pull it down).

Given that it seems like the PS3 will surely trounce the Xbox360 in HW capabilities, I wonder whether ease of development will have the final say on who has the better gaming platform.

Microsoft responded

[2bluemike]

by releasing a new picture of the Xbox 360. Bill Gates was heard to remark, 'Quick, buy out Toshiba before they ruin everything!'

There is no license.

[node 3]

You don't need a license to use your own software. A license comes into play when you distribute your software.

So when can we start buying them?

[nurb432]

All this stuff is nice, but until i can actually get one and mess with it on my bench its still just 'ooh, thats cool'.

Lets see some silicon!

Re:This is WAY cool

[mikael]

Because either way, you are going to have to define new process structures to represent each auxiliary processor units as well as the PowerPC CPU and recompile the kernel. For such real-time processing you want to keep the data structures to the absolute minimum and not have any 'fluff' left over from previous CPU architectures. Writing a kernel from scratch is the best way to achieve this.

Much like theTAOS OS did.

You've gotta be joking..

[LilGuy]

Synergistic Processor Elements? They aren't seriously going to start calling it that... please lord no.

I refuse to use some old buzzword from the 90s' megacorporations to describe my computer hardware. Forget about it.

Embarrassingly parallel tasks?

[CTho9305]

Simultaneous MPEG2 decoding, as shown here, is what computer architects call an embarrassingly-parallel problem. The easiest way to speed it up is just add more processors - with 8 processing units, the Cell is a great fit.

However, the really interesting problems are the ones that don't scale linearly in performance with the number of processors - these are the tasks for which the Cell processor will probably be running with 7 idle units and 1 active. These are also the tasks where we need actually new architectures; supercomputers like BlueGene will tear their way through extremely parallel problems.

One very cool approach to handling less parallel workloads (or even "sequential" workloads - like the majority of programs people usually run on PCs) is speculative threading - taking a sequential program, breaking it up into chunks, and running those chunks in parallel. Of course, when you do this, you have to make sure that the later work doesn't depend on the earlier operations, and check for violations of "sequential execution semantics" (programs expecting sequential execution semantics are ones that expect their instructions to execute in order - basically any program you'd write today). The Stanford Hydra project is an example that uses this technique; Wisconsin Multiscalar Group takes an approach that requires modified binaries to do something similar.

One thing people fail to mention when they talk about the supposedly-amazing performance of the Cell processor is its floating point precision: first, it only attains it's >200GFLOPS with single precision numbers (not accurate enough for many scientific applications), and second, it doesn't follow IEEE754 rounding requirements. The rounding policy in IEEE754 floats is specifically designed so that as you perform more and more calculations, the error doesn't grow rapidly. Cutting corners lets you calculate faster but even less accurate numbers. Basically, to get the high FLOPS ratings, Cell sacrifices precision in both the number of bits used, and the accuracy of the data in those bits.

Additional reading: Info on transputers

[Morgaine]

You write: Transputer != microprocessor

You really shouldn't comment on things you don't know anything about.

Here is some info on the transputer family, and links to data sheets on devices in each of the four main families. The T212, T414, and T805 became the most popular. And yes, they're all microprocessors, ie. a little integrated circuit CPU which you plug into a motherboard just like you do a Pentium, and with all the normal features of a normal microprocessor plus a few others of their own, like the 4 on-chip comms links. I've got a couple of T414's upstairs sitting on the shelf.

http://homepage.ntlworld.com/kryten_droid/inmos/im s_transputers.htm -- An intro to transputers

http://www.classiccmp.org/transputer/documentation /inmos/2186.pdf - 16-bit IMS T225 transputer (T200 famiily)

http://homepage.ntlworld.com/kryten_droid/inmos/im s_t414.htm - 32-bit IMS T414 transputer (T400 family)

http://www.classiccmp.org/transputer/t805.htm - 32-bit IMS T805 f/p transputer (T800 family)

http://www.classiccmp.org/transputer/documentation /inmos/4260.pdf 32-bit IMS T9000 virtual-channel transputer

These Inmos microprocessors were right down the middle of where Kutaragi wants to take the Cell, with lots of interdevice communications being handled directly by the hardware. Inmos even made graphics output chips which were often driven by multiple transputers in parallel, so graphics demos were really common on the transputer scene.

Interestingly, after being passed around between various European parties once Inmos ran out of money, the rights to the transputer were eventually sold off to some Japanese megacorp, iirc.