If Moore's law implies that the amount of transistors we can fit into a certain amount of space keeps getting bigger and bigger, is it really necessary to decrease the sophistication of processor logic? Why can't we have the best of both worlds - several processors that are all just as sophisticated as the processors that came before the multi-core era?
Post by @kayvonf:
@tborenest: your thinking is correct. It's really a matter of how sophisticated a single core is today, vs. how sophisticated a single core at today's transistor counts could be.
You could implement very sophisticated branch predictors, prefetchers, huge out-of-order windows, and add in a huge data cache, or you could choose more modest implementations of all these features and place a few cores on the chip. Studies show the former would (for most single-threaded workloads) increase your performance a few percentage points, or maybe at best a factor of two. However, in the the latter case, you now have four or six cores to give a multi-threaded workload a real boost. I should change this slide to say: Use increasing transistor count for more cores, rather than for increasingly complex logic to accelerate performance of a single thread.
Keep in mind that by multi-threaded, I mean multiple single-threaded processes active in the system, running on different cores. This is helpful on your personal machine when multiple programs are running, and it is very helpful on servers, where it's common to run several virtual machines on the same physical box.
A useful discussion from last year.
post by @tborenest:
If Moore's law implies that the amount of transistors we can fit into a certain amount of space keeps getting bigger and bigger, is it really necessary to decrease the sophistication of processor logic? Why can't we have the best of both worlds - several processors that are all just as sophisticated as the processors that came before the multi-core era?
Post by @kayvonf:
@tborenest: your thinking is correct. It's really a matter of how sophisticated a single core is today, vs. how sophisticated a single core at today's transistor counts could be.
You could implement very sophisticated branch predictors, prefetchers, huge out-of-order windows, and add in a huge data cache, or you could choose more modest implementations of all these features and place a few cores on the chip. Studies show the former would (for most single-threaded workloads) increase your performance a few percentage points, or maybe at best a factor of two. However, in the the latter case, you now have four or six cores to give a multi-threaded workload a real boost. I should change this slide to say: Use increasing transistor count for more cores, rather than for increasingly complex logic to accelerate performance of a single thread.
Keep in mind that by multi-threaded, I mean multiple single-threaded processes active in the system, running on different cores. This is helpful on your personal machine when multiple programs are running, and it is very helpful on servers, where it's common to run several virtual machines on the same physical box.
This comment was marked helpful 0 times.