Monday, February 9, 2009

Larrabee In PS4? So What’s the Future of IBM’s Cell?

My, oh my. Intel "bought" use of Larrabee in the Sony PS4, says says The Inquirer. Sony tells TechRadar that's fiction. Other news sources repeat both statements over and over. Land sakes, who are we to believe? And, if Larrabee's in, what's the future of IBM's Cell, now in PS3? Let's look at the sources, and then consider possible results.

The Inquirer's information came from people manning the booth at the Consumer Electronics Show, including (and possibly limited to) a now famous but anonymous "nice Sony engineering lady" who is probably now wishing she were under a rock. TechRadar called up Sony Computer Entertainment Europe, and a "rep" said this was "quite possibly the best work of fiction I've read since Lord of the Rings."

If you call up SCEE and get someone you only call a "rep," it was a PR person. Of course a PR person will say it's fiction. For one thing, it really is fiction until legally binding contracts are signed. Even after that, it's fiction in PR terms until everybody has agreed on and staged the proper event to announce it. At that point, exactly the script agreed to is fact; anything else is fiction. Also, that PR person was rather far from the likely centers of negotiation, and so likely followed his current script with a very clear conscience.

However, engineers don't necessarily know what's really going on, either. They know what they've heard, but many an engineer's thoughts and analyses have been tossed into the Butter Dimension by an executive decision that seemingly came out of nowhere, influenced by things far beyond engineering's ken. Like, maybe, a really good round of golf. Or a whole lot of money. I've heard "You don't know the whole story" many more times than I'm comfortable with.

The "whole lot of money" angle was brought up in The Inquirer, and Jon Stokes over on Ars Technica gives a lot of good reasons why that may be extremely relevant, even if it throws out a lot of coding done on the PS3 for Cell. Stokes also said that Intel has, not surprisingly, been working that angle with Sony for quite a while. So our engineering lady may be reacting to the latest round and jumping the gun a bit.

I'm inclined to believe "jumping the gun," if for no other reason that anything prior to the legal signing ceremony really is jumping the gun. But I'm also inclined to believe a tipping point may well have been reached, for Stokes' and other reasons. So let's assume Intel Larrabee is the GPU for the PS4.

What does this mean for IBM's Cell processor – formally, the Cell Broadband Engine – rather famously used in the PlayStation 3? IBM might still participate as a provider of main processors for consoles; IBM's PowerPC is used in Nintendo Wii and Microsoft Xbox 360. But Cell was, and is, special.

In addition to Cell's use in the PS3 – where it never was the complete GPU; an Nvidia chip does the back end processing (that's a split Sony wanted, I heard) – it gained fame in HPC circles. Attached as an accelerator to conventional systems in Los Alamos' RoadRunner within QS22 blades, Cell was the reason Roadrunner garnered headlines for topping the 2008 TOP500 HPC chart and for being the first system ever to exceed 1 PetaFLOPS. (To those without the official decoder ring: A PetaFLOPS, and yes that S is part of it, is a thousand million million arithmetic operations every second, done on, roughly, fractional numbers as opposed to integers.)

Cell's raw massive FLOPage was what made that possible. You can do a lot with 155.5 GFLOPS on a single chip (single precision Linpack benchmark) – oops, that's the first release. The one in RoadRunner is the new and improved PowerXCell, which peaks at 204.8 GFLOPS single precision, 102.4 GFLOPS double, and you double that for a QS22 blade, each of which has two PowerXCell chips. (Peak quoted because I couldn't find per-chip or -blade Linpack numbers. Sorry.) (Decoder ring: GFLOPS = GigaFLOPS is a mere thousand million per second.)

Will IBM continue development of Cell without Playstation volumes? HPC volumes are a nice frosting on the console-volume cake for IBM, but hardly a full meal for a fab. Further development will be required to stay competitive, since a widely-leaked Intel Larrabee presentation, given in universities, indicates (with a little arithmetic) that an initial 16-processor Larrabee is supposed to reach 640 GFLOPS single precision peak, half that for double. That's supposed to see the light of day in late '09 or early 2010, and maybe 24-way, with 32-way and 48-way versions following, providing proportionally larger peak numbers (that's 1,920 GFLOPS for 48-way); and possibly even more with a process shrink. Those are really big numbers. And Cray is planning on using it, so there will be big RoadRunner-ish non-Cell future systems, maybe with a PS4-derived development system.

On the other hand, the newer PowerXCell isn't used in the PS3, and development always knew it wouldn't be a high-volume product. Will IBM continue to fund it, perhaps using some of their pile of profits from Services and Software? (In IBM, Hardware has lagged those two for many years in the money department.)

My guess is that it will depend on how much good will high-level executives think they get out of the front-page column-inches that result from topping supercomputer lists and breaking arbitrary performance barriers. That is a market force definitely not to be sneezed at. Here's an example from my personal experience:

Back in the late 90s, I was asked by an IBM account rep to give a presentation on IBM's HPC cluster (RS/6000 SP) to the CIO of a large, very conservative, IBM client in the US Midwest. A high-end HPC system pitch to that client? A client who said they liked being behind the leading edge? I was skeptical, but travel expense was on the account rep, so I went.

About five minutes into the presentation, the CIO asked a question causing me to point out, clearly, that the system consisted of utterly bog-standard IBM UNIX (POWER) systems, in a rack, with a fast network connecting them as an option. Utterly standard. Regular old OS. Regular old hardware. That's what he wanted to know. It turned out that's why I was there: To have someone not on commission say that with some credibility. They were consolidating IT locations because it was hard to find competent UNIX sysadmins in the boonies of the US Midwest, and the account rep noticed that buying several systems as a nominal HPC cluster was simply a cheaper and denser way to buy a whole bunch of UNIX systems at once. The CIO, not unreasonably, wanted to be sure he wasn't investing in some strange pile of gear only a Ph.D. could operate.

I could have left right then and there, after five minutes, mission accomplished. What I did was noodle through the rest of the presentation, have a nice chat about this and that, later have a drink with the account rep, and go home. Gave one for the company, heigh ho.

So, they bought one. Or maybe they bought five, or ten. I forget. But the story doesn't end there.

A few months later, there was all over the news a multi-day historic battle at the end of which the IBM Deep Blue system, which was built on and publicized as an RS/6000 SP, defeated Gary Kasparov, the reigning chess champion. So what? So I get a call from the account rep a few days later, practically oozing glee, saying he now owned that CIO.

It seems that the executives at that company had been walking around during the match, and quite a while after, saying to each other, in so many words, "Hey, don't we have one of those systems that defeated Kasparov? Didn't our CIO just buy it? And we're using it for our operations?" Nudge. Wink. Smile.

They had a player in a world-class game that they heard about day after day. It was "their" system playing Kasparov. They were fans, and their team was winning. They never got involved in stuff like that, and, though they'd probably never admit it out loud, they all thought it was just too cool. They would have been high-fiving that CIO at every turn, if anybody at that level in that company ever gave a high five, and the CIO was strutting around like a pigeon with his chest puffed out. He now had massive cred – the primary currency of both the street and the executive corridors – and it was all because he bought that system from IBM, on the advice of that account rep. You know he wasn't going elsewhere anytime soon.

Good will can work, big time.

So maybe Cell will still thrive even if it's not in the PS4. But it will require someone to look a few quarters beyond this afternoon's bottom line, and that's an increasingly difficult sell.



(P.S.: Some of you may have seen some stuttering in this blog's feed around this article. I originally jumped out after seeing the rumor, but hadn't seen the denial. Igor Ostrovsky almost instantly left a comment that pointed to the Sony denial, after which I obviously had to do some rewriting. That involved taking down the old post and replacing it with this one, a process that dropped Igor's comment. Thanks again, Igor.)


Anonymous said...

Great read with a wonderful anecdote.

Sam Johnston said...

Interesting read. Regardless of the future of the cell processor I think there's some interesting things to learn from the architecture, particularly with regards to cloud computing.


Post a Comment

Thanks for commenting!

Note: Only a member of this blog may post a comment.