IBM has announced that supporters of IBM's "Cell" family of PowerXCell multicore processors have only another six months to buy more blade servers to build up their systems. January 6, 2012 will be the last day that the BladeCenter QS22 blade servers, the last of three different Cell-based machines, will be available from Big …
rip pos arch
The only surprise is that it took so long to die. this after all is the architecture that falls so flat on its face for general computing that it finally pushed apple over to Intel. lol Sony was dumb enough to believe the ps3 wouldnt need a gpu and had to delay launch adding gpu after getting first silicon. PowerPC wont die but at least it is largely gone from general computing.
@asdf, missing out
"PowerPC wont die but at least it is largely gone from general computing."
But not gone from other fields. Sectors like telecoms, military computing, etc. tend to care less about architectural backwards compatibility and more about performance, power consumption, etc. By being willing to switch around a bit they can exploit whatever is best at the moment. PowerPC is in fine form in the telecoms world, but has slipped a bit in the high performance embedded world.
Whereas 'general computing' has been stuck in the Intel rut for decades now. The trouble is that the battery powered and server farm sectors of 'general computing' have already chosen ARM or are threatening to do so. Why is that relevant? Well, it signifies a greater willingness on the part of the vendors to look beyond the world of x86 for the performance that sells. Doing that once means they have to keep doing it should something better come along in order to retain a competitive edge. It's entirely possible that PowerPC will be the chip of choice, and it might not be too hard for some vendor to go for it. The trouble is that such an endeavour will always be commercially driven; an offer of cheaper Intel chips might be just as commcercially advantageous as switching to another architecture to get a performance edge.
ARM is having a quite interesting impact on the market. They own the mobile market, and they may end up owning the laptop market too (MS porting Windows, Apple talking about an ARM laptop). They may also end up owning a large chunk of the server market too if vendors see useful performance per Watt figures for ARM servers. So where would that leave the great big hulking chips that Intel and AMD are churning out? With a somewhat smaller market I would imagine, apart from the power Desktop users, and there's not really many of them. So will Intel/AMD keep developing these very powerful chips if the ARM architecture starts taking over the server market too? Possibly not, or not at the same pace.
So where might users who do actually need fast general purpose compute performance turn? With Freescale seemingly tarting up the PowerPC line with some recent anouncements and the embedded market there to support it, there might yet be a commercial rational to move high performance computing over to PowerPC. Adobe may yet have to dust off their old Photoshop source code. I can remember in the early days of Intel Macs Photoshop was slower than on the G5s because there was no altivec unit to exploit for all those image processing functions.
Regardless, it seems likely that the end users are going to have to get used to underlying architectures changing more than once every 20 years. We should be grateful. It'll mean more performance (hopefully) and less power consumption, and who cares what instruction set lies beneath?
Fog of War
Pity. I've had my eye on Cell for years, but the roadmap uncertainty has been quite off-putting. Maybe Freescale's newly announce multicore PowerPCs will make up the difference. Regardless, Cell was certainly a programming challenge and was not one that any old programmer can achieve maximum performance in their first afternoon. Perhaps that's the real reason why IBM have backed away from it. I gather that there are some in the games industry who have got to grips with it (and all that horsepower presumably makes a difference), so maybe Sony will continue with Cell. Who knows.
Yes, CELL was certainly a programming challenge, even for IBM.
For instance, IBM researchers did heavy optimizing (loop unrolling, asm, etc) in String Pattern Matching benchmarks, and still you need 13 of the CELL@3.2GHz to match one Niagara T2email@example.comGHz. The Niagara did a plain implementation of the Aho-Corasick algorithm in pure C with no optimizations.
So the performance of the CELL might be stellar. But in reality it was quite slow and even researchers had serious problems to extract any performance.
I feel dizzy...
Cell is a back-eddy of Power2/Power3. It never got the proper 64-bit changes that IBM did with Power 4. It accepted core-shrinks and GHz accelerations, but no one used it for general processing in the last decade.
Apple went to Intel in 2005 because IBM wasn't developing Cell as Apple wished. Apple would have faced a conversion of their 64-bit programming to go from Cell to Power 5 (current at the time), but IBM wouldn't commit to mobile CPUs/chipsets that Apple needed to compete with Windows vendors. Apple did commit to the then-still-evolving x64 setup. It was 2006 before Intel's Core 2 did x64 with a mobile chip, but it has gone well since.
Power 7 is just radically different from Cell. Comparing Niagara to Cell is possible, but NOBODY runs Cell against Niagara. You cannot get AIX to run on Cell, for example.