Looking forward to getting your hands on a Cell workstation? We were too — until we talked to Linux for Cell kernel maintainer Arnd Bergmann. Learn about programming for Cell in general and for Linux for Cell in particular, and why the “workstation” that has been described in the press is more of a prototype Blade board, and a bit of a misunderstanding, in this Q & A session.
..and this is why Apple won’t be using the Cell processor in their future Macs.
It is a 64bit PPC chip like some others with a special addon for single precission float processing. That addon is indead powerful but not very useful for tasks that do not use single precision floats or that can not get split into 8 distinct chunks that can be executed in parallel.
This includes most office and server tasks and lots of scientific computing (which often relies on double precission numbers — the cell can do those too, but *way* slower than single precisson calculations). Everything else will just run on the PPC64 core. Tasks that “fit” the cell chip need to be handcrafted for the architecture: Do not expect those to spring up by the dozend as soon as the chip becomes available.
So far I only see game engines using the cell-addon… they need a lot of number crunching for the graphics (easy to make that run in parallel) and they can be sloppy about doing it, so single precission is enough. Well, that were the requirements the chip was designed for, no wonder it excels there:-)
Looking forward to getting your hands on a Cell workstation? We were too — until we talked to Linux for Cell kernel maintainer Arnd Bergmann.
Everything is just as it was to be expected. Why disapointment? The same fact goes for all other CPUs and it always did for every new feature it was introduced. It just souds like you were having preschool wet dreams about Cell.
Hoping that someone would wave his magic wand and optimizations would start working on their own? Nope, everything goes according to expectations. First implementations will arrise in the longterm operations as processing pictures, video or sound. These operations are not even a bit tied to your UI speedup expectations.
p.s. Actualy after Sony published price $399 for PS3 I’m looking even more to get hands on one of those.
The disappointment is not in the Cell processor – it is in that it’ll still be a while before anyone can get a “workstation” based on it. Sony/Toshiba might have one in the fall. IBM’s plans for Cell appear to center around their blade series.
The Cell processor will find a large niche in high performance vector and floating point calculations, but isn’t really a desktop CPU. That doesn’t mean IBM doesn’t have a desktop CPU available. It does – the 970FX (and dual-core is rumored to be on the way).
The disappointment is not in the Cell processor – it is in that it’ll still be a while before anyone can get a “workstation” based on it. Sony/Toshiba might have one in the fall. IBM’s plans for Cell appear to center around their blade series.
So…the fall isn’t soon enough, and a blade is not useable?
So…the fall isn’t soon enough, and a blade is not useable?
Blades aren’t good workstations. Good servers, but just to get one blade you have to have a chassis, power supply (220v), and an ethernet module (the chassis has an interconnect plane for other blades, but no outside ethernet connectivity without additional hardware). There is no KVM capability in the JS20 (970) blade, so it’s either serial or ethernet. It takes about $6k to get a single blade going. After that you can add blades at the cost of a blade until you run out of slots. Once the cost of the chassis is amortized over a full chassis filled with blades, the bang for the buck is pretty good.
There’s also the point that some media outlets were portraying the Cell server that was demo’d as a workstation. It wasn’t – it was blade based. No one has a workstation ready to ship _now_. It might take a while, which is disappointing. Cell itself is not.
Funny how they tried self-hosting, i.e. compiling on the Cell, but then went back to cross-compiling on their workstations.
Probably it was disappointingly slow, because branch-intensive compiler code is what the Cell is least suited for.