The future of Cell ?

In an article today, a report noted that Sony may be trying to sell off its fabs that make the Cell BE processor. This is, to quote the article

In a bid to streamline its business and get rid of some assets to free up some valuable cash,

Remember, there are 3 partners working on this: IBM, Sony, and Toshiba. Sony is using these in the PS3, which by all accounts are not selling well. PS3’s are badly crippled for computing (rumor has it that this is the case as a result of an agreement with IBM), and are having a hard time against Wii and others. Xbox is having a hard time against Wii.
If Sony is looking to get out of the Cell fab business, might they be looking to divest themselves of the PS3 as well? It certainly would make it easier to sell it off or close it down if they aren’t making the chips themselves anymore.
Cell-BE has huge potential for HPC. Yes, it is single precision. Yes, it is not easy to program. It is a design of CPU+8 APUs (or SPU in their parlance). And when freed of the constraining PS3 architecture, it provides absolutely insane levels of performance, for the right codes.
Our partner TerraSoft Solutions, has been working on this stuff.
The problem, and believe me, this is a problem, is that there are two providers of Cell-BE hardware out there. First of course, is IBM, who are not exactly interested in creating a competitor for their Blue Gene systems out of lots of Cell-BE units. They are sadly, not altogether supportive of efforts to try to expand the ecosystem, preferring to allow it to happen organically, and not drive this. Second is Mercury Computing. Unfortunately, Mercury’s line are PCI cards for almost $10k USD, or 1U units for about $13k USD. I am not aware of any other vendor’s offerings in Cell-BE hardware.
Now compare the pricing and interest with that of nVidia. The nVidia C870 product will come out significantly less than the Cell-BE PCI card. The nVidia side-car unit will come out less as well than the PCI card. The 1U unit will come out higher, though have 4 GPUs vs 2 Cell units. If you compare (yeah, it is specious) peak GFLOPs, the 1U nVidia unit will be ballpark 2TF, while the 1U Mercury unit will be ballpark 0.5TF, at very similar pricing.
My thoughts are that APUs and accelerators are here to stay. The winners will be governed in part by part availability, as well as pricing. Not to mention ease of programming.
We know Cell-BE can’t be too expensive, it fits into a sub $1k USD unit. It is a shame that there is not another provider of Cell-BE accelerator PCI units out there, and that there is no VC interest in this stuff.
I know, lets stop calling them clusters of HPC accelerators, and call them ajaxified web 3.0 social networks (of fast processing elements).Yeah, thats the ticket.
More seriously, there is a big hole in the market for non-GPU accelerator systems. We know about where the price point needs to be (and nVidia is hitting it). Since Sony looks like it wants out, what does the future hold for Cell-BE? Even if we could get some of the monied classes to want to fund a company building accelerators, this sale would be sure to raise eyebrows on any company wanting to build Cell-BE based accelerators. I don’t see IBM doing it any time soon, nor Mercury.

1 thought on “The future of Cell ?”

  1. As John notes on InsideHPC, Sony is denying anything other than a sale for cash.
    I am still worried about the PS3 though, as it is the major consumer of Cell-BE’s. Without the volumes of PS-3 (even if they aren’t as good as WII or xbox), Cell-BE pricing would likely be quite a bit higher.
    Nothing like volume to drive economies of scale.

Comments are closed.