I don't think his post demostrated it's false. If anything it just another opinion based on assumptions. Nvidia's been making all sorts of chips these days: sound chips, Opteron chipsets, GPUs. I don't see cpu's being anymore difficult. Sure maybe if you compared Intel's monopoloy with it's x86 chips then yeah Nvidia probably wouldn't be able to compete...has little to do with whether or not Nvidia can or cannot design a competitive part. Like I said Intel wouldn't be able to design a GPU that's competitive with Nvidia's either so no Pana's post doesn't prove anything.
Official PS3 Thread
- Thread starter Sonic
- Start date
- Status
I don't think his post demostrated it's false. If anything it just another opinion based on assumptions. Nvidia's been making all sorts of chips these days: sound chips, Opteron chipsets, GPUs. I don't see cpu's being anymore difficult. Sure maybe if you compared Intel's monopoloy with it's x86 chips then yeah Nvidia probably wouldn't be able to compete...has little to do with whether or not Nvidia can or cannot design a competitive part. Like I said Intel wouldn't be able to design a GPU that's competitive with Nvidia's either so no Pana's post doesn't prove anything.
If Intel wasn't hosed with the x86 ISA and built an new ISA, they'd whoop people. Oh look, The Itanium has been picking up steam like mad it's Int performance is rather good and it's FP performance is insane -- it'll only be getting better. They achieve very high clock rates and have far more complex designs.
Nvidia couldn't hope to design a competitive CPU. You don't just stick units together and say look I got a wicked fast CPU. RISC chips from IBM, Alpha, HP, Sun, Motorolla and so on are having their asses handed to them by Intel and AMD with the crappy ISA -- see Spec -- most usually in the price performance department. I wonder why so many people stopped buying those higher powered exotic workstations? Do you really think Nvidia with so little experience in solving the problems of a high performance CPU could hope to make something competitive? The skill transfer isn't as high when going from GPU to CPU, but the other way around it sure is. You have far more experience in designing more complex high performance parts in the latter instance. Pay close attention to the type of performance I'm talking about here.
Chipsets, APUs, GPUs are simple, when compared to CPUs. Chipsets have to go through a fair bit of validation but nothing like what CPUs have to go through. APU are dsps, wow, really hard. IIRC Nvidia used someone elses DSPs threw it together and said, look we made an audio processor. GPUs, they are just a lot of execution units in parallel and I doubt there is much custom work in them.
If Intel seriously designed a new GPU, I'm sure their wealth of information with execution units, memory controllers, cache heirarchies, busses and so on would mean one wicked GPU.
You keep on claiming GPUs are so complex when compared to CPUs, yet you show nothing to indicate that they are. Have you even taken a simple digital systems and computer architecture courses to have any idea of the amount of work it takes to design some of this stuff?
There are a fair bit of tools out there that automate a lot of this process, the problem is that in high performance CPU design, you can't really use them since they're far from optimal designs. GPUs use a lot of this, this is one of the reasons you see low clock rates from them, the other is the use of libraries rather than building custom circuitry for everything or nearly everything. Then there is the fact that the problems being solved by GPUs is simple if not closer to trivial than what is being solved with high performance CPUs. GPUs don't have to operate at ludicriously high frequencies and worry about heavily varied work loads. They have an narrow optimal work load definition and a low frequency --relatively speaking-- and their work load is very well suited to their optimal.
This is why CPUs require more design effort while GPUs don't! This is what makes CPUs more complex than GPUs and this isn't a small difference.
This isn't a matter of opinion, this is fact, supported by a whole lot of history.
I don't see Intel's non x86 cpus in the top 3 supercomputers. I also don't think the legacy x86 ISA is holding back performance that much otherwise the Itanium would be the top performing non x86 cpu in the world which it isn't. Heck CELL is gonna crush the Itanium2 like an ant and it's a lot simpler in design something that Nvidia could probably design ie lots of parallelism. I'd bet a CELL chip running at only 1 GHz would beat the crap out of anything Intel could come up with in the next couple of years.
Panajev2001a
Veteran
nVIDIA does not have the logic designers, the compiler designers and the chip architects and the experience in high speed micro-processor design to compete witha giant like Intel...
GPUs are chips born with a silver spoon and can afford to increase the number of parallel units and a deep pipeline as they get fed with highly parallel input data streams...
General Purpose CPUs are designed to run a huge set of different applications and they mostly work and earn their money by EXTRACTING parallelism from the code... compilers and the processor work in sinergy to have the best effect...
With the way nVIDIA design chips they would be stuck with a low clocked CPU that gets its ass kicked all over the place expect a handful of scenarios which all involve easily vectorizable code sequences...
An approach like Cell is designed to do GREAT for multimedia applications, but opther areas it doesn't need to be more than just acceptable performance...
A Processor like Prescott or Opteron are IMHO considerably more complex than a GeForce FX for example...
GPUs are chips born with a silver spoon and can afford to increase the number of parallel units and a deep pipeline as they get fed with highly parallel input data streams...
General Purpose CPUs are designed to run a huge set of different applications and they mostly work and earn their money by EXTRACTING parallelism from the code... compilers and the processor work in sinergy to have the best effect...
With the way nVIDIA design chips they would be stuck with a low clocked CPU that gets its ass kicked all over the place expect a handful of scenarios which all involve easily vectorizable code sequences...
An approach like Cell is designed to do GREAT for multimedia applications, but opther areas it doesn't need to be more than just acceptable performance...
A Processor like Prescott or Opteron are IMHO considerably more complex than a GeForce FX for example...
Panajev2001a
Veteran
As ironic as seeing a specialized part beating a general purpose processor in a benchmark designed for the specialized processor's strenghts...
Panajev2001a
Veteran
PC-Engine said:Panajev2001a said:
A a modified CELL or something similar to it could probably run Linux better than an Itanium2 and Nvidia doesn't need to know how to design the latter only the former
Are you talking about x86 Linux or IA-64 Linux ?
In both cases ( now finally Intel has discovered the joys of x86 dynamic emulation in software ) I would disagree...
Also you would be comparing a 4 GHz processor with a 1.2 GHz one... not that fair
Also Cell is a bit more general purpose than the average nVIDIA chip and has legions of compile writers and experienced architects working on the ISA, compilers, tools, APIs, etc...
Panajev2001a
Veteran
Sorry Paul, I do not see any picture
crystalcube
Newcomer
Paul said:What do you guys think what the actuall case of ps3 should be like? I wouldn't mind something like this.
Although this one is a little big, but you get the idea. The design of this thing is pretty cool.
It could be as sony designs its game console's looks to easily fit in a living room
unlike MS , which comes up with a shrunk desktop and Nintendo , which has a cuter toy like appeal.
PC-Engine said:A a modified CELL or something similar to it could probably run Linux better than an Itanium2 and Nvidia doesn't need to know how to design the latter only the former
Linux is not an application, its an operating system. Performance is going to be entirely dependent on what you chose to run on it.
Problems with lots of exploitable parallellism is going to flie like sh*t off a shovel on CELL.
This however requires that the developers is going to use a programming language that can express this parallism. C and C++ doesn't cut the mustard here.
I think SONY will deploy a new programming language that is designed to take advantage the parallel nature of CELL. But with a syntax close to something people are familiar with. Like OCCAM (C like) or maybe a C/C++ implementation of CSP (like JCSP is a Java class library that provides CSP primitives).
At the very least they'll provide libraries so that people won't have to start from scratch (linear solvers etc.)
But writing a compiler that can take C/C++ spaghetti and divide the code and data into chunks so that both can fit in eac APUs local memory and get good performance is just not going to happen.
Cheers
Gubbi
Panajev2001a
Veteran
Was this quote taken from H&P Gubbi ?
It does tell the truth though
- Status
Similar threads
- Locked
