i doubt that... the actual manufacturing costs for intel and amd cpus are around 30$ as far as i know, for a large chip it might be 60$, maybe 100$, but i dont think its 400$... then again thats how much building a new fab and sustaining the current one actually costs per chip..Quote:
Originally Posted by ahmad
but think about it... a less dense, much bigger, high clocked gpu... that might actually be a great way to scale gpus up...
you dont need a new design, you just bloat it up :D just increase the space between the logic and modify it a bit to reach higher clocks...
im not sure how high the design can be clocked without seriously redesigning it, but if thats not an issue and clocks would scale with die size, then a double sized rv770 performing 50% faster would be worth the higher cost i think...
well it would definately be bandwidth limited, thats for sure :D but it should perform very well i suppose...Quote:
Originally Posted by ahmad
but the way i always thought of it was this, gpus can be massively parallel without losing any efficiency, cpus cant.
perf per watt efficiency is easier to achieve at low clockspeeds than at high clockspeeds
hence gpus are massively parallel and big, dense chips, while cpus are not that parralel and rather need high performance per core, hence few high clockspeed cores, less dense, smaller chip
it seems out of the 3 ways to improve performance, an architectual improvement is the most tricky and risky approach, the easiest is ramping up clockspeeds or doubling the design making it parrallel.
ramping clockspeeds didnt work for cpus anymore, instead of improving the arch significantly they dodge and go the easy way, multiplying the cores :D
thanks for clarifying... i wasnt 100% sure since the author doesnt seem to know much about manufacturing tech and just referred to other sources.Quote:
Originally Posted by Hans de Vries
well its hard to define a node then, isnt it?
its a mix of density performance and power...
ati and nvidia want performance first, power second and density third i assume, other customers like alterra probable want power first, performance second and density third, and others density first, power second and performance third...
hmmm i really cant think of a way to compare nodes except for using reference designs and then comparing the two. something an sram cell but more complex, and the comparing the perf/power/density graphs
what do you mean?Quote:
Originally Posted by vitaminc
