Is there something that CELL can still do better than modern CPU/GPU
- Thread starter gongo
- Start date
I am not sure what you mean by efficiency versus flexibility, all GPU designs seem to be heading towards more fexibility. Could you elaborate?
Concerning manufacturability, it is even easier to produce two identical chips instead of two separate designs, i.e. Sony could put two of them in the PS4 as I suggested earlier.
I think Xenus means a single core is more flexible if it's a super-versatile chip, but performance will be less than the equivalent of two discrete, more specialised chips. eg. Larrabee was more flexible but slower than a discrete GPU designed just for graphics.
upnorthsox
Veteran
IMO it's efficiency and flexibiity vs initial design costs and yields with yields being the greast negator. Poor yields are just hard to get around no matter what benefits you may derive from other factors.
The hardest part is you need to commit to the process before you even know what ballpark your yields will be in. Because of this, whatever knowledge you have in manufacturing these components is vital. You certainly wouldn't want to be wedding a cpu with a gpu on a process node where everything is new to you.
I agree to a certain degree, but these kind of comparisons are a bit slippery because we don´t know how well Larrabee could have been with some tailor made graphics routines optimised to its strength. Running some bog standard DX ?.? performance tests and compare the results to some GPU hardware that has been tailor made to support that DX version may not do Larrabee complete justice. Larrabee is just an example, I don´t know the exact reasons why it was postponed.
Then we also have the efficiency consideration that it is easier to have a more continous utlisation and a more evenly distributed load through general purpose hardware than some specialised hardware that can be used only for some special type of calculations.
Deciding what is most efficient isn´t always easy, that general purpose hardware have less ALUs idling may compensate that is not as fast as dedicated hardware for certain tasks.
Wedding a CPU with GPU can be done in different ways, We haven´t seen everything yet. Just adding a tiny ARM core to Fermi would make Fermi capable of booting Linux, you could add some texture units to Cell and it would become a more versatile graphics processor. They would not be perfect but they, but neither is paring a bog standard CPU with a bog standard GPU on the same die.
Just wanted to say that there are many ways to skin a cat.
If you want competitive performance, the GPU will have to be reasonably big. I could expect adding a few tiny CPU cores, but a beefed up Cell wouldnt be tiny. In terms of cost you`d rather have 2 chips than 1 obscenely big one.Wedding a CPU with GPU can be done in different ways, We haven´t seen everything yet. Just adding a tiny ARM core to Fermi would make Fermi capable of booting Linux, you could add some texture units to Cell and it would become a more versatile graphics processor. They would not be perfect but they, but neither is paring a bog standard CPU with a bog standard GPU on the same die.
Just wanted to say that there are many ways to skin a cat.
The only question is if you want 1 CPU and 1 GPU or 2 hybrids, Im not sure if you meant that - it sounds like you would want a single chip. Separate GPU/CPU would have the advantage that you can have large cache and low-latency memory for the CPU and high ban width memory for the GPU, having memory thats both low-latency and high-bandwith aint feasible.
If a PS4 would come out the next year it would have separate CPU/GPU.... but there are a few things Im not certain of in the longer term.
There could be some killer-apps that would require a Hybrid and wont run comparable nice on a Cell(2), Im sure stuff like physics will benefit from hybrids but not to what extend. Which could mean 2 hybrids would be the better solution
Quite possible in a few years you will be more limited by powerdraw than die-size (in a small form-factor like console), which means you fill your powerbudget with rather small chips - at which point you just could use one hybrid. (But if power and thus efficiency is an issue you might aswell be better of with diverse specialized units instead of similar complex ones. rather have half of your chip idle all the time and the other half doing the best with the power)
IT's already the case, some members here gave various numbers on the matters, it's "gross" approximation but that gives an idea. One figure I remember was a shrink will allow for 50% more transistors per mm² but only a decrease 20% in power usage per transistors, so as you cram more and more transistors the power goes up. The trend is pretty clear and not new in the GPUs realm.
Thanks, I have not been paying close attention to those discussions.
Let´s hope Graphene will change those trends.
No arguments from me - that's my view!
ihamoitc2005
Veteran
God of War MLAA
Removing MSAA from GPU (RSX) saves 5ms.
Cost: 5 SPU @ 4ms + scheduling overhead
Benefit: 17.6% extra performance available by RSX for other rendering task at 30hz.
http://www.realtimerendering.com/blog/morphological-antialiasing-in-god-of-war-iii/
Removing MSAA from GPU (RSX) saves 5ms.
Cost: 5 SPU @ 4ms + scheduling overhead
Benefit: 17.6% extra performance available by RSX for other rendering task at 30hz.
http://www.realtimerendering.com/blog/morphological-antialiasing-in-god-of-war-iii/
Sony buys back Toshiba's Cell plant for 50 billion yen, makes a killing and plans a CMOS fab -- from Engadget
Sony's bought back the Cell fab they sold to Toshiba, for less than Toshiba originally paid for it. Sony apparently told Nikkei Business Daily that they were looking to make imaging semiconductors with the plant.
Sony's bought back the Cell fab they sold to Toshiba, for less than Toshiba originally paid for it. Sony apparently told Nikkei Business Daily that they were looking to make imaging semiconductors with the plant.
corduroygt
Banned
A SPE is 7 million transistors without the LS. 8 of them makes 56M transistors, or about the same number of transistors intel uses for x86 decoding on a core processor. With all the sunk costs for SPE programming, maybe 8 SPU's could be Sony's version of x86, they'll put it in the chip to use their existing programming work and libraries and to still have all PS3 software compatible with PS4.
That's assuming they can make part of the cache (2M) reconfigurable as 8 local stores when the spu's are used, or have the option to use it as a part of a larger cache pool when the next gen enchanced PPU's are being predominantly used.
That's assuming they can make part of the cache (2M) reconfigurable as 8 local stores when the spu's are used, or have the option to use it as a part of a larger cache pool when the next gen enchanced PPU's are being predominantly used.
Silent_Buddha
Legend
They could, but it sounds like Sony plan on converting that Fab into CMOS chip production for HD camera's.
What other plants are there still making Cell chips?
Regards,
SB
It would be interesting if that sort of thing was what IBM were looking at with their 'Cell-related work for next-gen consoles', but I don't know if they've got the ability to make cache behave identically to Cell's local store, given the extremely short cycle time the LS has.
Can't wait until we start hearing real details about the next gen.
corduroygt
Banned
LS latencies are right between L1 and L2 cache latencies of intel x86 processors. So it could be treated as a sort of L1 cache (that's huge) and L2 cache could be off chip in the form of edram.
Similar threads
- Locked