How powerful is Cell at graphic rendering.
- Thread starter scatteh316
- Start date
Lysander said:
Programmable FLOPs are things like vertex and pixel shaders. The programmer has control over how the information is processed. Non-programmable FLOPs are hardwired--the chip does the same thing to the data and the programmer has no control over it. Texture filtering would be a non-programmable process on a chip.
Jawed
Legend
No, I'm not in the industry. Hope I haven't given the impression I am 
Things like texture filtering and floating point blending have to be added in.
Do a search on these forums for "1.8tflops" or "2tflops".
Is that number believable? I'm not convinced, to be honest - but there's no doubt that RSX is more than the few hundred GFLOPs you get simply by adding-up the vertex and fragment pipelines' capabilities.
Jawed
Things like texture filtering and floating point blending have to be added in.
Do a search on these forums for "1.8tflops" or "2tflops".
Is that number believable? I'm not convinced, to be honest - but there's no doubt that RSX is more than the few hundred GFLOPs you get simply by adding-up the vertex and fragment pipelines' capabilities.
Jawed
And even if they were "even" on the peak FLOPs that would be completely irrelevant. To adapt a quote from Steve Balmer, "Architecture! Architecture! Architecture!"Jawed said:
GPUs are specialized for processing graphics data. The evolution of the architecture has been with the single minded goal of tackling the goal, graphics rendering, the best way possible with the given silicone budget.
Processors, even CELL, are more general. They have to tackle a broader set of data types and problems so they cannot be quite as specialized. So even if you tossed CELL into a competition with a GPU with a similar FLOPs rating, CELL is going to get spanked. Flip flopping the competition, a GPU is going to be very... very... very slow at most general processing tasks. Your talking about a highly parallel design with little to no silicon spent to effeciently tackle the problem.
So it is not really fair to compare a GPU that has evolved over years and years to solve specialized problems with CELL or Xenon (a first attempt at solving a different problem). Like the first generation GPUs they will keep what is good, toss what is bad, and move forward with those designs. But there is no way they are going to complete against a GPU in graphics processing.
Megadrive, Atari ST and Amiga all had 68000. I cut my teeth on ST and Amiga and my heart will always belong to the 68000, best CPU ever!Lysander said:
The point I was making it that any CPU can do the same work as any other, simple at differenent speeds. So both Cell and 68000 are more capable than any GPU BUT they might be considerible slower, hence the tea comment...
Answering the question more directly, a GPU does vast amounts of computation locked up in fixed function operations. Just the texture samplers alone would take a CPU ALOT of operations to do, things like texture address ops (i.e. modulus for tiling), bilinear filtering, format conversion is all hard coded on a GPU.
And thats fundementally why its so hard to compare, IF you make a CPU do everything a GPU can do, it will be fairly slow. However if you just make a CPU do a form of rendering suited to its architecture, it can do a fair bit.
Seeing as the original plan for the PS3 GPU was a Cell on the same die as a basic GPU, scatteh316s question doesn't seem completely nugatory to me.
Being a general purpose SIMD/MIMD architecture Cell would probably work pretty well as a dedicated unified shader and PSU. Although of course simple repetitive "slave tasks" are better left to dedicated hardware.
The Cell CPU can still perform some of those tasks interleaved with other stuff.
Edit: Changed CPU to GPU in the first line.
Being a general purpose SIMD/MIMD architecture Cell would probably work pretty well as a dedicated unified shader and PSU. Although of course simple repetitive "slave tasks" are better left to dedicated hardware.
The Cell CPU can still perform some of those tasks interleaved with other stuff.
Edit: Changed CPU to GPU in the first line.
Last edited by a moderator:
Squeak said:
First time i see that.
Squeak said:
We heard rumours and saw some strange schematics posted by Version, who loves posting made up stuff, but personally i really cannot remember this "CPU on the same die as GPU" thing. Not saying it's not true, i just never heard.
Lots of people patent ideas. And even if it was in the cards, how could have Sony have known in 1999 that GPUs would advance the way they have. I think Unreal came out in 1998, and it had software rendering on a P2 (which was not a FLOPs monster by any degree compares to the 1-2TFLOPs Sony had in mind). The Voodoo2 was really limited comparatively. Looking at the under powered fix featured GPUs of the time I would have thought a flexible and powerful GPU based on CELL could be a good idea too. The GS seemed to ok this last gen. Obviously with onboard T&L with the GeForce 256 and later programmable shaders and other features to improve IQ graphics chips have come into their own.Squeak said:
A lot of tech, like GPUs, move at a snails pace. GPUs just happen to be on an aggressive development path that makes them better for graphics tasks than a general purpose CPU. If 3D graphic cards had died an early death a CELL based GPU may have been a good answer to the problem.
I don't know that that was ever on the same Die though. People get confused with those patents for visualiser and the 1 teraflop claim. 1 Teraflop was for a 4 Cell Broadband Engine network. This hasn't appeared it seems mostly because 65nm is available and the cost is too high. The visualiser was a patent, not a blueprint for PS3. Companies can patent tech even with no interest in developing it. I met a chap who worked at IBM and said they have a department where you can take ideas to get patented.
It's possible that the patent was for a tech Sony were intending to use in PS3. It's also possible the patent was for a concept to secure the IP for 20 years when it may become more important, and Sony thought they might try it out along with other ideas to see viability in PS3.
I repeat, patents are NOT blueprints and you cannot take patents as being system designs, nor see a final product different to a patent as being a change from original plans.
It's possible that the patent was for a tech Sony were intending to use in PS3. It's also possible the patent was for a concept to secure the IP for 20 years when it may become more important, and Sony thought they might try it out along with other ideas to see viability in PS3.
I repeat, patents are NOT blueprints and you cannot take patents as being system designs, nor see a final product different to a patent as being a change from original plans.
Similar threads
