specialized HW is fast, but new CPUs are fast as well...

[KimB]

Intel produces massively parallell supercomputers. Look at their site for some info, plus of course the usual suspects for high-performance computing.

There's a difference there, and quite a big one. nVidia and ATI have produced single-chip processors that can execute far more math ops than Intel's processors can. I was talking about manufacturing concerns, not about parallel processing management.

Also, eventually integrated graphics will become the norm as the performance bottlenecks move from the memory interface to the interface between the various chips. In other words, I feel that eventually, everything will be integrated (Well, as much as it can...it may not be feasible to integrate some things, such as quantum and silicon processors...).

 

[Panajev2001a]

There's a difference there, and quite a big one. nVidia and ATI have produced single-chip processors that can execute far more math ops than Intel's processors can. I was talking about manufacturing concerns

Intel produces GENERAL processing unit not specialized 3D accelerator... the number of inherent parallelism you normally extract from the code is not that huge and plus they went for the speed-racer and not braniac ( ~K7 ) approach... and designing and building these GENERAL processor such as the Pentium 4 and Itanium 2 is more complex than a GeForce FX...

still you say "manufacturing concerns"... about INTEL ? do I have to remind you again that Intel is already shipping in HIGH volumes .13u parts and is almost ready to mass produce its 0.09u part ?

 

[KimB]

Intel produces GENERAL processing unit not specialized 3D accelerator... the number of inherent parallelism you normally extract from the code is not that huge and plus they went for the speed-racer and not braniac ( ~K7 ) approach... and designing and building these GENERAL processor such as the Pentium 4 and Itanium 2 is more complex than a GeForce FX...

still you say "manufacturing concerns"... about INTEL ? do I have to remind you again that Intel is already shipping in HIGH volumes .13u parts and is almost ready to mass produce its 0.09u part ?

Yes, which is also part of the point. I don't think Intel could compete with ATI or nVidia if they tried. Not right away, anyway. Obviously if Intel decided to take a couple of years to learn how to properly execute for the high-end 3D graphics market, their superior fab plants and money would allow them to come out ahead, but they couldn't do it the first time out of the gate.

Of course, Intel is doing okay in the extreme low-end 3D market, due to the integrated chipsets, but that doesn't change the fact that their integrated 3D sucks. By "competing" I mean in a performance/image quality standpoint.

 

[g__day]

One a side note - if we look forward 2-3 years I expect chip densities to hit from 200M - 500M transistors. You can do alot with that sort of transistor count. I can understand why analysts say CPUs and GPUs are on a collision course.

In the not too distant future the CGPU.

PS The attribute of a parallel unit defining that most determines its power is not so much the number and speed of its sub units, but their number of interconnects and the ability to arbitrate communications effeciently and effectively and keep them all productive.

 

[KimB]

One a side note - if we look forward 2-3 years I expect chip densities to hit from 200M - 500M transistors. You can do alot with that sort of transistor count. I can understand why analysts say CPUs and GPUs are on a collision course.

Well, I think that what we will see first are integrated GPU's will come first, and then large amounts of integrated RAM. GPU's integrated with CPU's will come later (These integrated processors should actually be entire SOC's...CPU, GPU, RAM, I/O, etc.). The primary problem with integrating CPU's and GPU's is simply that the best developers of both are different companies. It will be interesting to see how companies will manage this. Will there be attempts to make modular chip designs, where, say, one portion of the die could actually run at a different clock speed, and be designed by another hardware vendor? It could be very interesting.