Which more advanced, GPU or CPU

[Rhonda85]

Now GPUs have over 384million transistors, wow! I have been wondering which chip is more advanced and costs more to manufacture, a CPU lets say an AMD FX60 verses an ATI X1900 R580 GPU?

 

[pm]

Searching on Google appears to reveal that the FX60 has a die size of 199sq.mm. and the R580 has a die size of 315sq.mm. So, in terms of size of the die, the clear winner (or loser ) is the R580. This translates to higher manufacturing cost for the R580 - fewer die per wafers, lower yield due to larger die size.

"Advanced", however, is a non-scientific term. It requires a judgement on what constitutes "advanced" and then applying this to the designs and deciding. A couple of quantitative ideas for "advanced" could be:
1. comparing the size of the design team and the length of the design time
2. comparing the quantity and irregularity of the non-memory structures
3. comparing the complexity of the logic - how regular is it.
4. comparing the length of RTL code

Without access to the numbers for either design, but based on my knowledge of the industry, I would say that the FX60 wins (or loses if you have to fund the development) hands-down on all 4 of these. But really these are measures of complexity - not "advancedness". I'm not sure what constitutes "advanced". But based on my biased opinion (I design CPUs), CPUs are more complex, and are more advanced.

 

[dmens]

Also, consider the difference in design methods. From what I hear, large portions of the GPU are synthesized from rtl with static CMOS with latch/flop sequentials, as opposed to CPU's which at the least use domino (or worse yet, self-reset domino) to achieve speed, resulting in a massive increase in design effort. Even hand-designed static CMOS is about a 2-4X increase in design effort compared to direct synthesis.

If you define "advanced" as "difficult", I'd agree with pm that CPU's are more difficult in the following areas:

- Definition: Protocols are more complicated with non-systolic pipelines (CPU's) as opposed to, as far as I know, more regular pipes such as GPU's and DSP's
- Logic design: Speed targets demand aggressive design methods (hand placement with strange topologies), and so forth
- Validation: Given the breadth of the x86 ISA, it is very difficult to validate all functionality on every protocol corner case (see definition)
- Debug: Not too sure on this one, maybe pm can answer, but silicon debug for CPU's is considered black magic...

 

[Bobthelost]

The technology being used for CPUs is more advanced than theat used for graphics cards. Look at the relative sizes of CPU and GPU manufacturing methods, Nvida just moved down to 90nm i believe, since the CPU industry is leaving 90nm like rats from a sinking ship...

Graphics processing units are also a lot simpler than CPUs, or at least they are much easier to scale up. GPU work being incredibly easy to multithread scaling the core up so there are more pipelines is cheaper than being forced to ramp up the clock speed and shrink the die.

 

[byosys]

To "dumb down" what has been said above: GPUs are very parallel chips, so adding "copies" of a given set of hardware greatly increases performance. For example, adding more pixel pipelines is essentially just copying the design of one pixel pipe and adding more of them. Not very complex in terms of design, but very efficient in terms of performance. While this is true for some specific parts of a CPU, for the majority of a CPU this is not the case. To increase performance in a CPU, designers need to employ all sorts of tricks and methods (and the occasional black vodoo magic as well) to make the CPU run faster as a whole.

Of course, this depends on how you define advanced. The above is based on the advanced as in design.

 

[Never Knows Best]

Well Sony is making the cell chip which is more similar to a gpu then a traditional cpu. It has a main processor and then synergistic processing cores to handle the simpler but more often run algorithms. The main one handles the more complex routines and divides the work among the synergistic cores. Although I would say a cpu cisc chip is more complex to make then a gpu because to increase speed in a cpu you need to redesign a lot like how they have cpus translate cisc commands to risc ones to increase speed. Gpus are designed with a specific function with the exception of programmable shaders so the design is predetermined by the math that goes into processing triangles in three dimensions. Beyond that I really don?t know which is cheaper to make though considering development and production yields.