Talk:Cell (processor)

Talk page Archive-1 7 June 2006, 117KB

Differences between CEll/ Multi-CPU Systems

It would be interesting to read what differences there are between those, both from a coder and consumer point of view. Intel is gradually introducing multi-cernel CPUs with more and more cores so aren't these becoming competing concepts?

Trusted Computing / DRM

I'm extremely surprised that this article makes no mention of the Trusted Computing / DRM system built into the Cell Processor. Some people think it is a good thing and some people think it is bad, but either way it is extremely noteworthy. Just to cite a single link, IBM has a technical document on it here. IBM themselves explicitly discuss Digitial Rights Management there, which should forestall any controversy over applying the term DRM. I may try to add this to the article myself in the future, but I don't have time right now... and to be honest I know I'd have to work pretty hard to produce a suitably Neutral POV writeup. I'm not here to grind an ax, I came here LOOKING for information and was befuddled by its absence. Alsee 11:22, 31 October 2006 (UTC)

Open Source Hardware

I read at many places, that Cell will be open source hardware:

• Is it only the interface specification to the hardware, which is open?
• Or is the whole hardware design (including verilog source code etc.) public available?
• Or is even the hardware design licensed under a licence which fulfills the 4 freedoms of software/OS Definition?

If anybody knows sth. about it => Please add it!

Major Rework Proposed

In anticipation of this, I first moved the existing talk page to Achive-1.

At this point I don't have any fixed ideas about how to best proceed. I'm merely aware that the existing article has a number of defects that need to be somehow addressed, and I have certain intentions to add new content that won't help matters while the article remains in its present form.

I don't wish to tackle all the issues all at once. First, a general survey of what I see as the big fish.

citation system

The good side is that there are lots of citations, references and links. The bad side is that the article has sprouted no end of unsourced statements. In my own contributions I wasn't careful to expose my citations, I tended to hide them inline as HTML comments until I could attack this problem more systematically.

I've looked at all the Wikipedia citation methods, and none of them are issue free. My own opinion is that the new cite.php method which inlines the citations is a bad direction. I feel the overall framework could be salvaged with a version 2 rewrite, more along the lines of one of the bioinformatics citation systems I once saw, which it already partially resembles.

The existing ref|note system works fine for shorter articles, but not so great for articles of the present length, and the numbering system sometimes conflicts with other things such as footnotes, if those are employed.

Given a choice, for the more technical material, where I'm primarily citing IBM design engineers, I would lean toward the Harvard referencing style, which I find the most easy to verify and maintain while reading the text in article format; I also feel it handles a large body of references much better than the ref|note system.

One possibility is that this article should be split, leaving the main article to keep the existing style while freeing the more technical sub-articles to adopt Harvard style.

comprehension

Cell is a complex topic. I tend to feel a core article is required that tries to remain light on the jargon (EIB, coherent interconnect, DMA controller, etc.) while presenting a servicable overview.

Note that the Cell microprocessor is conceptually distinct from the Cell BE architecture. This follows the model X86 architecture as distinct from derivative devices such as Intel 80486 or Pentium. A first cut at improving comprehensibility would be to slice off some of the technical description of Cell as a architecture, since that is not what most casual readers are intending to pursue: they mostly wish to find out about Cell the commercial phenomena.

I sense a third distinct article will be required to cover Cell software development methologies. An example of this is the "porting VMX to SPU" text that I added to the main article somewhat prematurely. There's so much more that could be said on this subject esp. pertaining the unnatural supervisory relationship between the PPE and SPE cores, the coherent DMA controllers, and how to cope with that mess.

subdivision

I already opened this subject under comprehension because I feel that is the primary driving force in shaping the main article. Finding natural cleavage points to simplify life for the expositors (myself among them) is secondary.

From my present vantage point, I would probably begin by introducing a split into these three titles:

• Cell microprocessor - the overview and device specific material
• Cell Broadband Engine Architecture - heavy on engineering
• Cell BE development - heavy on software development

I chose those titles not to confuse people looking for Cell biology.

I'll leave it off here for tonight. Does anyone have any observations about whether this will work or not work? More sections? Fewer sections? Different emphasis? Exactly where to cut? Please make your views known. MaxEnt 07:54, 7 June 2006 (UTC)

controversy of the day

• broken Cell / RSX refuted

Not even worth mentioning since it's just a case of stupidity or willful ignorance on the part of some hasty reporters. This has nothing to do with the Cell microprocessor or architecture. -- uberpenguin @ 2006-06-08 00:07Z

I come here exactly for the same reason. As far as I know, those numbers were presented by Sony. How exactly they managed to broke their measurements bests me.
I am not really concerned about the slower triangle transform rate (NV4x is not triangle limited by using 2 tris per pixel) but the memory access performance is something awful.
What's the real deal with this issue?
MaxDZ8 talk 06:40, 19 July 2006 (UTC)

The "Local Memory" in the Sony slides is the RSX's video memory, the bandwidth is the bandwidth for reading back from the video memory into the Cell. AFAIK that wasn't even possible at all with the Emotion Engine and isn't something that is usefull in computer games. If you want to make a screenshot, you can just write to main memory. --Silvestre Zabala 19:46, 19 July 2006 (UTC)==

(Sorry for the late reply, I don't really understand how did I miss this).
It's still something quite ugly if you want to GPGPU something and stream it to CPU for further processing. It looks quite useful to me on next-gen games, unless they have very high GPGPU capabilities and this would mean RSX to be USM 4+ (or they just have a very good shader compiler). MaxDZ8 talk 08:03, 19 September 2006 (UTC)

English usage of "belies"

I don't quite understand the intent of this phrase in the Overview: "The name [Cell BE] belies its intended use, namely as a component in current and future digital distribution systems;"

The American Heritage dictionary defines "belies" as follows:

  1. To give a false representation to; misrepresent.
  2. To show to be false; contradict.

What is the contradiction between the concept of "cell" and the concept of "digital distribution"? I wonder if the author of this phrase actually has something like "underlies" or "gives a hint about" in mind. At any rate, I think the statement as it is will not parse for many non-native speakers of English, and should be rephrased to not use the word "belies".

Fbkintanar 03:25, 20 July 2006 (UTC) Cebu City, PH

belies

How about "The name of the Cell implies the smallest structural part of a larger organism, as a component in current and future digital distribution systems;" I don't want to change this article as it's really good considering the complexity and points of view involved. --Notagoodname 21:32, 3 September 2006 (UTC)

subarticle navbar mockup

First, I familiarized myself with Wikipedia:Summary style and considered some other issues in choosing good names, and also found one about how to break up a long article in sub-articles.

Then I started to mock up a possible approach. Created category:Cell BE architecture and then modeled a navbar after the one in Isaac Newton, and mocked up some text in the podlet articles.

The sw-dev podlet could at some point split in a separate Linux on Cell article, but I think it would be best for now to keep all the development concerns in one place until it gets large enough, if that happens, to spill over.

I'll be starting soon to move large blocks of text out of the main article into the corresponding podlets. Any suggestions about where to draw the scalpel are welcome. MaxEnt 23:29, 9 June 2006 (UTC)

fat reduction and restructure

• first reduction complete

The main purpose thus far was to reduce the overall size of the main article and align it better with the topics designation for subarticle. Much remains to be done. MaxEnt 06:41, 10 June 2006 (UTC)

MIMD

I was tracking down other WP articles that might have pertinent content.

• Shader
• GPGPU
• Stream processing
• MIMD

Added a small screed to the MIMD article of the sort that got me in a little trouble here for being too colorful. I'm determined to educate, even at the cost of having to say something. Based on what I wrote there, expect the MIMD references in the present article to diminish and go into the west, as Galadriel once spoke. MaxEnt 09:11, 10 June 2006 (UTC)

Should the "Other MIMD-on-a-chip processors include" part of the "See Also" the "cell microprocessor" article be moved to MIMD? Leave behind only a link something like

* "The Cell microprocessor is one of several MIMD-on-a-chip processors"

? --70.189.73.224 04:28, 24 September 2006 (UTC)

Heavy copyedits history and overview

It's hard to dial-in to the major fault lines without cleaning up the text as it already exists. Note that the MIMD passage that bothered me so much was actually in the stream processing article, not here, as I realised again later. MaxEnt 00:39, 12 June 2006 (UTC)

I created a somewhat odd inline-footnote in the history section about the often-confused distinction between cores and threads. I think it works well to get that out in the open in a passage most readers are likely to at least skim. Appreciating the highly-threaded nature of the Cell is an matter of core (ouch) comprehension for the rest of the article. MaxEnt 00:39, 12 June 2006 (UTC)

Copyedit lede

I liked the content I introduced into the lede a week ago, but the wording was a little heavy and the resulting text too blocky to read comfortably. I split some sentences and paragraphs, added more intralinks, and fixed a bad parallel sentence construct by introducing the word favors.

After responding to some initial criticism that my stronger vocabulary (modest, ground breaking, prowess) was too POV, I no longer feel there is any inbalance here. My initial use of ground breaking was indeed excessive, but now that I've demoted and recast it as breaks ground and anchored the use to obtained many patents it no longer functions as a free-radical to convey an adulatory impression. Likewise I feel that my depiction of the PowerPC core as modest is properly balanced in describing the floating point capability as prowess. The people who hate Cell (Carmack) point at the crappy PPE; the people who adore Cell (supercomputing centers) point at the staggering floating-point subsystem; these two words together effectively flag and bookend a wide spectrum of opinion. Likewise exotic is an effective flag for the jaded architect who only wants to know "what does Cell have that I've never seen before?" If anyone still thinks the lede is too POV, you'll have to speak up again. MaxEnt 02:16, 12 June 2006 (UTC)

major edit extended through the weekend

It's going a little slower than I hoped as I wasn't able to find as many hours over the last couple of days as intended. Some of the material is being reworked in my personal MediaWiki (the conehead has landed) and won't be visible here until I'm ready to merge the new material. User:W.marsh came along and removed my major-edit pilons 48 hours before my reserved period expired I guess because he equates churn with progress, and there wasn't much churn visible here for a couple of days. If anyone wants to become involved with this restructuring effort I'm more than happy to negotiate sub-tasks, but please don't remove the work-in-progress pilons unless you plan to pitch in; they aren't impeded anyone who wishes to contribute, just those who wish to contribute without checking in before-hand with those of us who are already active. MaxEnt 09:35, 15 June 2006 (UTC)

Thursday

It seems I have to make frequent reports or the killer T cells come along and abscond with my pilons.

Content for software development

Today I'm trying to work up some content for the software development podlet. I'm busy extracting glossary terms from this series of articles:

• http://www.research.ibm.com/journal/sj/451/eichenberger.pdf
• http://cag.csail.mit.edu/crg/papers/eichenberger05cell.pdf
• http://arstechnica.com/news.ars/post/20060225-6265.html
• http://domino.research.ibm.com/comm/research_projects.nsf/pages/cellcompiler.refs.html/$FILE/slide-pact05.pdf

No new content has landed yet. Just one grouse about why this process takes so long. IBM regards the PPE and SPE processors as two distinct processor types, but refuses to provide consistent formal names. The PPE ISA is usually termed a PowerPC ISA, or more precisely a PowerPC 970 compatible ISA, while the SPE ISA has no formal name at all. MaxEnt 12:25, 15 June 2006 (UTC)

From GreenbDS -- we've worked to clean up the Power branding so it's more consistent. If you can update this page based on http://www.power.org/brand_center/home we appreciate it. The SPU ISA is still just the "SPU ISA". Since SPUs never stand alone, it probably does not pay to have some sort of branding associated with that aside from the Cell branding itself. But your grouse is gratefully accepted!

Some wrong numbers?

There is a text "At 3.2 GHz, each SPE gives a theoretical 256 GFLOPS of single precision performance. The PPE's VMX128 (AltiVec) unit is fully pipelined for double precision floating point and can complete two double precision operations per clock cycle, which translates to 6.4 GFLOPS at 3.2 GHz; or eight single precision operations per clock cycle, which translates to 256 GFLOPS at 3.2 GHz[7]." Well, this would give you 2.3 TFLOPS, while Cell spec. says 256 GFLOPS total... Also, 3.2 x 8 = 25.6, not 256.

From GreenbDS -- please see http://www-128.ibm.com/developerworks/forums/dw_thread.jsp?forum=739&thread=136461&cat=46 The 200+ GFLOPS is accurate for the whole chip, not for "each SPU."

VMX?

Which VMX on the VMX page should this link to? --72.128.33.115 22:56, 18 September 2006 (UTC)

AltiVec -- mattb @ 2006-09-19T00:34Z

Presently this is described as VMX128 which is incorrect for Cell.

All of the technical documents I've seen have referred to it as VMX128, a VMX unit with 128 bit wide vector registers and some extra odds and ends. Do you have a source to back up your assertion? -- mattb @ 2006-10-10T17:17Z

VMX128 is VMX with 128 registers (as implemented on Xenon), vs. "standard" VMX which has only 32 vector registers (which is what the PPE implements). Both have 128 bit wide registers.

You're correct. -- mattb @ 2006-10-16T15:29Z

IBM Blade

Disclaimer: I work for IBM.

IBM has announced that general availability of an IBM dual-Cell blade (called "QS20") will commence on September 29. Here are the specs: http://www-03.ibm.com/technology/splash/qs20/

Also -- in terms of the lenghy list of links on the main page -- let me encourage you to point people to a portal to many of them: http://www.ibm.com/developerworks/power/cell.

I apologize that I am new to posting comments here. For tracking purposes, please call me "GreenbDS".

branch prediction

"The SPU branch architecture does not include dynamic branch prediction, but instead relies on compiler-generated branch prediction using "prepare-to-branch" instructions to redirect instruction prefetch to branch targets". [1].

also "Two SIMD instructions can be issued per cycle: one compute instruction and one memory operation."

also, what is the power consumption? the berkeley article estimates 40 watts. --24.7.86.143 17:19, 20 October 2006 (UTC)

news

IBM have mentioned they may/will be making 'cells' with optimised double precision instructions on the SPU's. "JK: We used that first tape out to get the initial software up and running. There were modifications we did to the chip over time. The design center is still active and participating. Our roadmap shows we are continuing down the cost reduction path. We have a 65 nanometer part. We are continuing the cost reductions. We have another vector where we are going after more performance. We have talked about enhanced double-precision chips. Architecturally we have double precision but we will fully exploit that capability from a performance point of view. That will be useful in high-performance computing and open another set of markets" from http://blogs.mercurynews.com/aei/2006/10/the_playstation.html.

Perhaps in the future it will be neccessary to make the difference between the cell used in the PS3 and other cells - hope this is useful

400 million

What's with the first paragraph? It states that the project cost has been $400 million but gives no citation to this.

I aded this link. http://www-5.ibm.com/de/pressroom/cebit/pressreleases/060307_1en.html -- Henriok 10:18, 12 December 2006 (UTC)

About the Successors of Cell

Wouldn´t it be better to add a section about this? While it´s wonderful to know how it is good right now, but I just can´t see any future research on this architecture, like I see with all the others. Wouldn´t it be nice to add a section about this? —The preceding unsigned comment was added by MTd2 (talk • contribs) 13:01, 12 December 2006 (UTC).

65nm

To whom it may concern. http://news.spong.com/article/11413?cb=721 says 65nm cells have been made.87.102.4.89 16:33, 3 January 2007 (UTC)

Commercialization Section: Expansion Request

Please expand the following paragraph within the Commercialization section:

This Cell configuration will have one Power processing element (PPE) on the core, with eight physical SPEs in silicon. In the PlayStation 3 one SPE is locked-out during the test process—a practice which helps to improve manufacturing yields—leaving seven SPEs operational in PS3 software.

How does this process improve manufacturing yields? In what way is it 'locked out'?--RedPoptarts 23:41, 4 January 2007 (UTC)

When choosing CPUs for the PS3 Sony can allow one SPU to be broken. They can chose second rate processors, something that neither Mercury nor IBM does when they choose Cell CPUs for their products. I think that the disabling of SPUs is made inside the processors or the chipset. There are features that can detect and disable faulty circuitry. Either that or a combination with the operating system. I think that the operating system locks out or reserves an SPU for itselv leaving just 6 SPUs to the game developers. I have read this somewhere but a quick googling revealed nothing. Thre might need some more digging. Either way.. this article needs some rewriting. As it stands now, it's almost like its a future product, but it's been commercialized for about a year by now. -- Henriok 08:22, 5 January 2007 (UTC)

Cell using the Windows Vista?

Hi everyone, i want to know if the Cell proccesor would be capable of working with windows vista or future microsoft operative systems one day, do u think that would be possible if the cell is heavily commercialized and if it´s used with new servers? I believe this can happend, but i´d like to know ur opinion. And one more thing, how has yellow dog linux been working on the PS3? is that a nice OS to work with?

There's no technical reason why Microsoft shouldn't be able to engineer Vista to work on computers based on Cell processors, but there are certainly A LOT of political and market reasons why that won't ever happen. Short: Microsoft won't spend resources backing a direct competitor's product (the Sony PlayStation 3), and they won't spend resources backing such diverse niche products where the Cell is going to end up. Cell is bound for set top boxes, embedded computers, supercomputers and game systems, where Windows have virtually no precense at all. I don't see Cell processors breaking into Micorosfts core market of desktop and server computers anytime soon, or at all.

Yellow Dog Linux is certainly running on PS3s and is running good. It's as nice to work with as any Linux I'd say, but for traditional home and work use, I'd recommend a traditional cheap x86 box with some other Linux distro like Ubuntu. Cell processor shine when you can bring the use of the SPUs, and surfing, mailing, wordprocesing and the like, you'll only use 10% of the system's capacity. Are you doing 3D rendering, scientific computation of dynamic flows, or gene splicing, you might want tro try out a PS3 cluster with YDL though. -- Henriok 11:16, 5 February 2007 (UTC)