Eracer:
Still harping on the same failed points.
Oh, so AMD isn't releasing an 8.8GHz single-core Athlon 64 by choice in a few months ...
Why don't you understand thermal limits? That's what killed the 4GHz Tejas. And the 10GHz P4. That would be just like that overvolted light bulb, burned out in an instant instead of the seven plus years that both AMD and Intel use as the lifetime of a CPU. The more you overvolt it, the quicker it burns up. The clock speed cubed equals power rule is why multiple cores have higher performance.
Good thing yields can go out the window once AMD starts producing integrated CPU-GPUs. That'll make them cheap to produce for those $100-$200 systems.
A 1x1 Fusion die would be good for OLPCs, PICs and palmtops. At 45nm it would have a $20 packaged cost. Getting $50 for it means 60% GMs. $30 packaged gets you a 2x2 Fusion. Selling it for $75 still yields 60% GMs. Higher speed ones would cover the mainstream market. 4x4 Fusions would be used for the high end HPC markets enjoying the higher margins. 2x6 Fusion pairs are likely to be desired by the extreme gaming crowds. Some of the GPU cores would likely be used for physics to allow that blow it away FPS crowd to get its kicks.
The only statement I made was what die sizes and power consumption would be in a straight shrink from 90-nm to 45-nm SOI. I never claimed that a straight shrink is what AMD had planned or that is was even practical. I'm quite sure the clock speeds would increase significantly, but going from 500MHz to 2GHz for what AMD claims will be a low power consumption, low performance system is a very big stretch in my opinion.
You fail to note that less pipelines would be needed at those higher speeds to get the same performance. That is where a lot of the die area reductions would come from. As I stated, it looks that 8 pipelines would fit into a K8L core sans L1 caches normally included. That would mean that the clock could be halved and the same performance still achieved. Given the clock cubed power rule, 8 pipes run at 80% would use the same power as 4 pipe run at 100%. 8 pipes at 50% would have about the same performance, but 1/4 the power used. A 1x1 45nm Fusion would be used in those $100 PCs. 2x2 Fusions would be used for the $200 value PCs at 80% of the clock. The mainstream would be at 100% and up clocking 2x2 Fusions.
Based on AMD's comments so far and the rather short timeline it is a reasonable assumption that AMD/ATI have no plans to completely reinvent the GPU into the likeness of a CPU in the next two years, if ever. If AMD was serious about going that route they wouldn't have needed to buy ATI to do it.
But they can and state that they will use the GPU as an option for replacing a core in a multicore die. Which flavors would sell into which markets is unknown at this point. The point is that AMD has the flexability to find out which percentage mixes yield the best values and profits in each segment. Also as many have proven before, there is value in the "have it your way" approach.
So why would AMD/ATI want an integrated 2GHz 3-pipeline GPU for a low-power consumption, low performance platform when they could cut power consumption by using 6 pipes @ 1GHz instead?
You are missing the point. Its what their customers decide, not what AMD thinks. They will tell them which way to go once both options are tried, benchmarked and tested. They may even be able to decide on the fly which they want as they could change the multipliers for the GPU cores just as they do for the CPU cores. The mix and what their customers would pay for each option is what will determine the number of pipelines in a GPU core. AMd is well known to allow multipliers to be decreased from the limits set during packaging.
Well, isn't it interesting that those same overclockers who try to overclock GPUs also use extreme cooling in order to keep temperatures down when overclocking? How many GPUs do you know of that have been overclocked to 2GHz? How about 1.5GHz?
That is the point! Because AMD will allow a OS, OEM or user to decide what the GPU multiplier should be for their given situation. AMD would limit the multiplier to the bin it was packaged for. It doesn't stop the silent PC crowd from lowering the multiplier to go fanless.
As for the high speed GPU crowd. That is one of the problems. They don't have the option of a high figure of merit process (process that gets high performance per watt expended). So the power requirements require lots of slow pipes. If they could get their hands on a process where they could double speed at the same power, wouldn't they opt for fewer double speed pipelines? Especially, if they would get them for less money?
Pete |
