Cell was a disappointment in the PS3 though. It was such an underperformer that the SPEs couldn’t be used for their intended role as a GPU and got shanghied into being used like SMT cores, while a whole separate GPU chip had to be added into the design at the last minute to compensate. The Cell’s weak CPU performance, strange programming model, and weak developer tooling was a problem for pretty much the whole generation.
The much more traditional XB360 tri-core plus GPU design was widely considered the better design and Sony reverted to a similarly traditional design the next generation as well.
It was a highly heterogeneous design, and yes the world has adapted to those much better now than they did in 2007 (and maybe it would have been more successful with better tooling), but there are also prior heterogeneous designs like Saturn that were also widely considered a mess apart from rare cases where developers were skilled enough to effectively utilize them. The tasks where it succeeded (like the oft-referenced air force supercomputer) tended to be embarrassingly parallel workloads (matrix math, password cracking, etc) that didn't have a high degree of dependencies in their processing workflow. That is pretty much the lowest-hanging fruit for a heterogeneous throughput-oriented system, if it couldn't do even that it would be fully useless, but merely doing that doesn't make it good.
I guess maybe think of it like a GPGPU - but a GPGPU would have been a terrible choice to use as your control CPU for a console even if it did allow higher degrees of processing in some narrow use-cases. Just not how that type of programming is usually done. Maybe at best you can say it was mis-understood and mis-applied by Sony but it really just wasn't as great a chip for most things as people thought at the time. Wasn't a great GPU, wasn't a great CPU, wasn't good enough to be the first "APU" to bridge to having both one one chip as originally intended.
It clearly was designed for HPC/supercomputing situations where someone understood the architecture. Probably still was held back by the tooling and documentation anyway. I know it was in some supercomputers, but there probably is a reason that 10 years later nobody is designing follow-on processors or inspired by that particular architecture.
People at the time jerked off heavily about it, both in HPC and the PS3, and I get that it powered a lot of people's fond childhood memories, but I think in hindsight it pretty much was a failed design and an architectural dead end. You can say that maybe APUs took some inspiration from it but the cost-savings of an integrated CPU+GPU are pretty much an obvious goal, and later processors used a much more standard CPU and GPU that happen to live on the same chip and have some coherent memory access protocols. They share nothing in common with the Cell's SPE/ring concept. Nobody ever used that in a design ever again, for a reason.
Sony wanted high-performace integrated graphics a full generation before such things were really feasible. I think that's ultimately why the CELL was such a failure - it was trying to do way too many things too ineffectively.
Also, building a GPU out of multiple tiny CPU cores on a ring bus just seems incredibly wrong. Even back then, the way you built a GPU was to have one instruction stream controlling many ALUs and register files. Intel made the same mistake with Larabee too - I guess everyone just assumed all the fixed-function bits would disappear eventually? (Meanwhile, ray tracing means even more fixed-function units...)
Cell's problem was the end of Dennard scaling hit them as hard as it hit the Pentium 4. It was designed in an era that was expecting to be able to get 5Ghz+ easily out of (albeit premium) consumer electronics chips.
By the time they figured out that they weren't going to be able to cash the metaphorical checks the process side had been writing up until that point, it was too late to go back to square one. The design unfortunately had to be pigeon holed out of it's designed niche, and they had to strap that anemic Nvidia GPU to the side of it just to have a shippable product. Turned out really well given all those factors working against it.
There's a good On the Metal podcast with Jonathan Blow as the guest, where he talks about his experience with the Cell processor. He also states that in his experience, the idea didn't really pan out.
The much more traditional XB360 tri-core plus GPU design was widely considered the better design and Sony reverted to a similarly traditional design the next generation as well.
It was a highly heterogeneous design, and yes the world has adapted to those much better now than they did in 2007 (and maybe it would have been more successful with better tooling), but there are also prior heterogeneous designs like Saturn that were also widely considered a mess apart from rare cases where developers were skilled enough to effectively utilize them. The tasks where it succeeded (like the oft-referenced air force supercomputer) tended to be embarrassingly parallel workloads (matrix math, password cracking, etc) that didn't have a high degree of dependencies in their processing workflow. That is pretty much the lowest-hanging fruit for a heterogeneous throughput-oriented system, if it couldn't do even that it would be fully useless, but merely doing that doesn't make it good.
I guess maybe think of it like a GPGPU - but a GPGPU would have been a terrible choice to use as your control CPU for a console even if it did allow higher degrees of processing in some narrow use-cases. Just not how that type of programming is usually done. Maybe at best you can say it was mis-understood and mis-applied by Sony but it really just wasn't as great a chip for most things as people thought at the time. Wasn't a great GPU, wasn't a great CPU, wasn't good enough to be the first "APU" to bridge to having both one one chip as originally intended.
It clearly was designed for HPC/supercomputing situations where someone understood the architecture. Probably still was held back by the tooling and documentation anyway. I know it was in some supercomputers, but there probably is a reason that 10 years later nobody is designing follow-on processors or inspired by that particular architecture.
People at the time jerked off heavily about it, both in HPC and the PS3, and I get that it powered a lot of people's fond childhood memories, but I think in hindsight it pretty much was a failed design and an architectural dead end. You can say that maybe APUs took some inspiration from it but the cost-savings of an integrated CPU+GPU are pretty much an obvious goal, and later processors used a much more standard CPU and GPU that happen to live on the same chip and have some coherent memory access protocols. They share nothing in common with the Cell's SPE/ring concept. Nobody ever used that in a design ever again, for a reason.