IBM is pulling out of Cell development
http://www.playstationuniversity.com/ibm-cancels-cell-processor-development-1295/
http://www.playstationuniversity.com/ibm-cancels-cell-processor-development-1295/
End of Cell for IBM
- Thread starter Butta
- Start date
Competition has caught up with CELLs advantage: Execution unit density.
In the near future we will see x86 cores with AVX with its 256 bit SIMD extensions, - and Larrabee with its 512 bit SIMD units. The raw flop density of these cores will be close, - or better than CELL's SPEs, - and with a regular programming model.
The four way SIMD seems borne out of being a good fit for pixel and vertex computations with ARGB and XYZW composants. But with the world + dog moving to a SOA programming paradigm, there is no reason for why vectors-units shouldn't be wider.
The concept of a local store might persevere. It essentially has register semantics, requiring explicit loads and stores (through DMA) with no automatic coherence. It might live on as a second level vector register file in some form, - but greatly reduced in size. Having 256KB of processor state in a massively threaded world is doomed.
Cheers
In the near future we will see x86 cores with AVX with its 256 bit SIMD extensions, - and Larrabee with its 512 bit SIMD units. The raw flop density of these cores will be close, - or better than CELL's SPEs, - and with a regular programming model.
The four way SIMD seems borne out of being a good fit for pixel and vertex computations with ARGB and XYZW composants. But with the world + dog moving to a SOA programming paradigm, there is no reason for why vectors-units shouldn't be wider.
The concept of a local store might persevere. It essentially has register semantics, requiring explicit loads and stores (through DMA) with no automatic coherence. It might live on as a second level vector register file in some form, - but greatly reduced in size. Having 256KB of processor state in a massively threaded world is doomed.
Cheers
upnorthsox
Veteran
Well I guess we now know why the PS3 Slim doesn't have a linux option.
Yep. I think the basic principles of the design will be reused (or have been successfully reused
) to create a new processor that is more general purpose, but above all, easier to develop for. I think they'll want to create a chip that's more balanced in its performance, and less specialised, so it can be used for a wider range of purposes and maybe more easily work as a standalone processor at the heart of a personal computer or server, but retains the scaleability and oomph.I've said before but I think if Cell had arrived just a year earlier even - or rather launched on time - it would have created an architectural branch in the industry with some legs. But no use lamenting more than need be I suppose; I'll remain a fan of the tenets the architecture stood for. Even outside of the ultimate 'success' v 'failure' verdict, for an architecture that has really been around for some time now, it is still an extremely strong performer given proper coding.
I think it wins special recognition somewhere in the anals of 2000's microarchitecture history, for sure. Heterogeneous, and first to 1 TFlop on the Global 500 let's remember.
But I agree with Gubbi in the regard that unfortunately where Cell development has stood still, other architectures have marched forward to begin encroaching on its strengths while retaining their own. Of course as an architecture it didn't have to end this way, but Cell never broke out of the niche HPC realm, and the economies of scale weren't there for the development, especially given the economy the last two years in those markets where it would have been targeted.
Well, I am curious to see what will go into PS4. Cell has never been a given to begin with, but I don't think this news rules it out either. Sony/Toshiba still have the resources on their own; the actual architecting work should be minimal if they stick to the SPEs in roughly their present incarnation. But a shift to Larrabee as rumored before wouldn't surprise me either, even before this news.
I think it wins special recognition somewhere in the anals of 2000's microarchitecture history, for sure. Heterogeneous, and first to 1 TFlop on the Global 500 let's remember.
But I agree with Gubbi in the regard that unfortunately where Cell development has stood still, other architectures have marched forward to begin encroaching on its strengths while retaining their own. Of course as an architecture it didn't have to end this way, but Cell never broke out of the niche HPC realm, and the economies of scale weren't there for the development, especially given the economy the last two years in those markets where it would have been targeted.
Well, I am curious to see what will go into PS4. Cell has never been a given to begin with, but I don't think this news rules it out either. Sony/Toshiba still have the resources on their own; the actual architecting work should be minimal if they stick to the SPEs in roughly their present incarnation. But a shift to Larrabee as rumored before wouldn't surprise me either, even before this news.
I think Larrabee conceptually works well with what Sony was trying to do, is potentially more accessible, and does travel down the road of unifying the CPU/GPU. It could have the potential perk of running code designed for the SPEs fairly quickly as well. One thing discussed about a new Cell design is that upping the LS would increase latency and could break compatibility anyhow. With all the CPU-side work Sony engineers are already doing for graphics Larrabee or a Larrabee like design would probably be just what the doctor ordered.
I think that you need to look at this in several different ways. Just because IBM has pulled out of the architecture for their own purposes, that should not mean that either Sony or Toshiba will also do the same. Please correct me if I'm wrong on this, but wasn't the STI partnership there to build a basic architecture, rather than a single product? Wasn't there at least 3 stems on the roadmap itself? Sony with the Cell BE, Toshiba with Spurs (or Micro Cell as it was called), and IBM with whatever they wanted to do? In the end, wasn't Cell BE not what the IBM engineers wanted, but what Sony, being the largest financial contributor to the project, forced on them? Wasn't IBM really along for the technological R&D ride so to speak, and went along with all of it because of a) heavy outside finance, and b) the presumption that PS3 would sell like the PS2 (not to mention go into millions of other CE devices) and that would make the chip as cheap as chips compared to the competition?
All I can see here is that IBM have decided that the Cell BE as is is not what they need going forward, but will use some of the technology behind it to develop something new for themselves. Sony and Toshiba can go their own ways also, Toshiba already did with Spurs, and IMHO, Sony have too much invested into this chip to simply throw it away and start from scratch. It's highly scalable, and in the embedded space really has no equal. It might not suit IBM going forward, but an updated/upscaled Cell BE will do PS4 just fine.
All I can see here is that IBM have decided that the Cell BE as is is not what they need going forward, but will use some of the technology behind it to develop something new for themselves. Sony and Toshiba can go their own ways also, Toshiba already did with Spurs, and IMHO, Sony have too much invested into this chip to simply throw it away and start from scratch. It's highly scalable, and in the embedded space really has no equal. It might not suit IBM going forward, but an updated/upscaled Cell BE will do PS4 just fine.
We don't even know how Larrabee actually performs in the real world do we? I can't see it happening.
EDIT: Serious question, how would increasing LS increase latency in terms of compatibility?
Also, IBM not using it, doesn't mean that IBM would not be willing to do work on the chip on behalf of Sony. They never intended on using Xenon, but they did the job.
Last edited by a moderator:
In an optimistic view, this could allow Sony to tailor a "Cell2" to the needs of a Console without having to think about reuse of the same chip in servers.
Just look at how much unused diespace in the 45nm version of Cell is, surely you could place a couple of other chips there that the PS3 needs anyway, and the FlexIO has nearly half its lanes unused...
Just look at how much unused diespace in the 45nm version of Cell is, surely you could place a couple of other chips there that the PS3 needs anyway, and the FlexIO has nearly half its lanes unused...
Well IBM contribute the PPU and Toshiba the SPU. Toshiba wanted all SPUs but IBM didn't agree and wanted their PowerPC in there. Sony is just the largest buyer for Cell with PS3 and pretty much the one that got them together. Toshiba have some CE products that'll feature Cell, but these are CE product, it'll be a long time before they need 32 SPUs Cell in CE products.
Sony didn't invest a whole lot for Cell. NV invested more in their GPU line compare to Sony with Cell. Sony is invested heavily in their Playstation business, Cell is just one of their venture. They certainly are not tied to it, if they want to use something else in the future. It's not like Intel and the x86 or IBM and PowerPC.
But from this announcement it is unlikely that PS4 will use Cell.
I am always curious about this, but when STI was designing Cell, why didn't they design it to be used for GPU as well as CPU ? Afterall it was going to be used for PS3.
@andypski: Uh, yes I do deserve to be called out on that spelling.
Well, I don't think you should boil it down to that - in fact if anything it even underplays Sony's own involvement. Toshiba was for the SPEs in terms of philosophy definitely, with IBM out for Power and its interests, but IBM engineers contributed hugely to Cell's development, down to the supported instructions of the SPEs. It was a team of volunteer engineers within IBM operating under a different aegis than IBM corporate 'central' we should remember, and a lot of them were onboard with the small/efficient/fast idea.
The architecting was around ~$500M... which is no joke it should be mentioned... and the capital investments in fabs and related were in the billions. They recouped some of that with the fab transfers to Toshiba, but... it definitely is a lot of money.
Anyway that's neither here nor there, I just wanted to comment on those.
Well, I don't think you should boil it down to that - in fact if anything it even underplays Sony's own involvement. Toshiba was for the SPEs in terms of philosophy definitely, with IBM out for Power and its interests, but IBM engineers contributed hugely to Cell's development, down to the supported instructions of the SPEs. It was a team of volunteer engineers within IBM operating under a different aegis than IBM corporate 'central' we should remember, and a lot of them were onboard with the small/efficient/fast idea.
The architecting was around ~$500M... which is no joke it should be mentioned... and the capital investments in fabs and related were in the billions. They recouped some of that with the fab transfers to Toshiba, but... it definitely is a lot of money.
Anyway that's neither here nor there, I just wanted to comment on those.
Well I would easily believe that the cell was closer to what Toshiba and Sony wanted. Cell makes sense in PS3 in HPC realm Ibm needed to married it to opterons.
My opinion (for what it's worse) is that Sony should have listen to IBM more wich would most likely have result in a peace off Toshiba.
Watching at the efforts put in SPUs and how great they are/were (at ther time) in regard to power consumption, throughput, ability to reach close peak figures in some workload, I wonder what they could have achieved by focusing on a more conventional cpu core (whether it had beed PPC or MIPS).
Basically a "Xenon done right" IBM might have want to use in HPC market. They may have offer in 2005/2006 a pretty low power quad core, pretty cheap to produce and with terrific throughput in regard to what X86 was offering back in time. Instead of focusing on GHz they may hace anticipated the move to wider SIMD (say 8 wide) do a clean start in this regard and gain a five years lead on X86.
I would love to see how such a chip would have fare (on the HPC market).
