The proposition of the Cell chip architecture being put forth by IBM, Sony and Toshiba depends on kaleidoscopic alignment of component technologies in a way that actually shows us something new. But, without development tools, Cell won't sell.
From the other room, I overheard a conversation between two of my teenage sons: "What if you could plug your brain into a supercomputer ..." said one, "... and have it design a game from your mind?" finished the other. In nearly perfect unison, they agreed: "That would be so cool!"
I wondered if I should tell them about the Cell architecture
not just the advanced microprocessor but also the concept for a pervasive distributed platform whose building block has been called a "supercomputer on a chip." Thats the claim being made by the interesting trio of IBM, Sony and Toshiba.
To look at the prospects of something such as Cell requires neither a microscope nor a telescope but rather a kaleidoscope. The key to Cells success wont be found in the submicroscopic wizardry that packs 234 million transistors into 221 square millimeters, soon to shrink further from 90-nanometer to 65-nm technology. It wont be found in the long-range trends of bandwidth that will make it affordable for Cell machines to talk with one another on a worldwide scale.
As my former colleague Dave Methvin once observed, "Its not whether machines can talk to each other; its whether they have anything to say." Distributed processing has to be more than a worldwide multigigahertz gabble of "Are you there?" "Yes, Im here; are you there?" And so on. The Cell proposition, over and above the raw performance of the processor and network alone, therefore depends on kaleidoscopic alignment of component technologiesboth hard and softin a way that actually shows us something new.
I dont want to understate Cells raw strengths. One thing that makes Cells success more plausible is its high off-chip bandwidth. Distributed-processing skeptics have been scoring points with arguments such as, "Why would I send data to other processors to share the work? By the time I move the data off the chip and back again, Id have done better to do the job myself on a single, fast multicore chip." Past chip architectures have mustered a weak response. Cells bandwidth to off-chip memory, on the order of 100GB per second, puts more of the burden of proof on the skeptics.
Click here to read Peter Coffees examination of why theres less division than there once was between high-performance and data center computing.
So will Cell machines be able to share immersive virtual-world environments with real-time multiplayer interactions? Or handle supply chain optimization and other enterprise tasks on a massive scale? Undoubtedly. The question, though, is whether Cell-based programmers will soon see the kind of programming tools that will let enterprise developers describe their bolder visions in executable terms. Cells diverse types of processor core, like those of Sonys Emotion Engine, offer a powerful palette to the game designerbut are novel territory even for early adopters of enterprise grids and developers of the tools to use them.
Without breakthrough development tools, Cell will just be science fiction.