IBM has revealed that the Cell processor line is an evolutionary dead-end. Some of the ideas behind it will live on, but the Cell family itself will not. I guess we now know what probably won't be in the PlayStation 4.
In an interview with Heise.de, IBM's VP of Deep Computing, David Turek, confirmed that the Cell processor has reached the end of the line. Turek then put a more positive spin on the news by stating the obvious truth that heterogeneous multiprocessors, of which Cell was the first mass-market example of, are here to stay, so insofar as IBM continues to produce such chips, Cell's basic concepts and ideas will live on in the company's product line.
By many metrics, Cell has been a success for IBM, even if it didn't live up to much of the hype that preceded its launch. In conjunction with x86 chips supplied by AMD processor, Cell has helped IBM's RoadRunner take and, until recently, keep the top slot in the Top 500 Supercomputer List. And, as this older HPCWire article on rumors of Cell's demise points out, Cell-based systems have been among the most power-efficient in the industry. IBM was able to build the specialized coprocessor in high enough volumes to keep its price down because the company was successful in selling it to Sony as a game console chip; the degree to which Cell, which gives the PS3 higher peak theoretical performance than the Xbox 360, has worked out for Sony is debatable. But IBM's decision to call it quits on the line confirms suspicions that Cell's overall commercial success has been limited, and there will be a number of cheaper, higher-performance, more widely supported alternatives to the processor starting in 2010.
IBM is certainly correct that at least part of the future belongs to heterogeneous, single-chip multiprocessors (CMPs)—these are chips that contain at least two different types of cores, typically a mix of general-purpose cores and more specialized cores. In this respect, the heterogeneous CMP party really kicks off in 2010, when both Intel and AMD will launch processors that feature a CPU and GPU on the same die. These CPU/GPU combos will eventually give way, at least on the Intel side, to processors that look substantially like Cell.
At some point, Intel will integrate Larrabee, which is a GPU made up of multiple small x86 cores, onto the same die as a regular x86 core. The result will be at least one large, general-purpose processor core and some number of small general-purpose cores that consist heavily of SIMD floating-point hardware.
The differences between Intel's future CPU/Larrabee hybrid and IBM's Cell may seem small, but they're critical. Cell's smaller floating-point cores are not general-purpose—they're specialized and they implement their own instruction set. These small cores also don't have cache coherency and a real virtual memory implementation. Rather, they have "local store" pools of programmer-managed memory that make them a huge pain to program for. So in terms of programmability, the difference between Cell and a CPU/Larrabee hybrid is night and day.
OK, it's not exactly night-and-day; it's more like "night and early evening," because there are still significant hurdles to actually designing software that takes full advantage of higher levels of parallelism. So while it may be easier to write code for a heterogeneous x86 CMP, the higher-level challenge of breaking an application up into parallelizable tasks still remains.
As for comparisons between Cell and what AMD is doing, this is more difficult, because AMD isn't dropping many hints about what it has in store for the final stages of its CPU/GPU "Fusion" plan. I've suggested that they won't go the CPU/Larrabee route described above, but that remains to be seen.
Speaking of Larrabee, the oft-floated rumor that Intel's upcoming GPU will form the brains of Sony's PlayStation 4 hinges in part on the idea that Sony and IBM will abandon the Cell, and not double down on it. Now that we know IBM is ditching Cell, does this mean Larrabee is a shoo-in? Not quite. This removes one of the objections to Sony using Larrabee, but an even bigger objection remains: even in 2011, which is the earliest that we could expect a PS4, Larrabee still may not be as good at graphics as discrete GPUs from NVIDIA and AMD/ATI. There's also an outside possibility that Sony could just use one or two of the last version of Cell in its console, and let Toshiba do the fabrication.
Source: ars technica
