Comment author:sketerpot
02 January 2010 08:29:59AM
0 points
[-]
The current lull in the increase of the speed at which CPUs perform sequential operations comes to an end, yielding a consumer CPU that performs sequential integer arithmetic operations 4x as quickly as a modern 3GHz Xeon (80%).
When you say sequential integer operations, do you mean integer operations that really are sequential? In other words, the instructions can't be performed in parallel because of data dependencies? If not, then this is already possible with a sufficiently wide superscalar processor or really big SIMD units.
But let's assume you really mean sequential integer operations. The only pipeline stage in this example that can't work on several instructions at once is the execute stage, so I'm assuming that's where the bottleneck is here. This means that the speed is limited by the clock frequency. So, here are two ways to achieve your prediction:
Crank up the clock! Find a way to get it up to 12 GHz without burning up.
Make the execute stage capable of running much faster than the rest of the processor does. This is natural for asynchronous processors; in normal operation the integer functional units will be sitting idle most of the time waiting for input, and the bulk of the time and complexity will be in fetching the instructions, decoding them, scheduling them, and in memory access and I/O. But in your contrived scenario, the integer math units could just go hog wild and the rest of the processor would keep them fed. This can be done with current semiconductor technology, I'm pretty sure.
So, either way, kind of an ambitious prediction. I like it.
Comment author:Valkyrie_Ice
02 January 2010 11:20:15AM
-1 points
[-]
Have you not heard that they discovered a way to use graphene as a one to one replacement for copper in chip production. That alone will allow speeds of 12-15GHz.
I would put faster chips using multicore running at many times current speeds will be available by 2011-2012 at near 100% certainty.
= 27d567bc2d48d9f40e9ccc20ea370b15)
= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
= f037147d6e6c911a85753b9abdedda8d)
When you say sequential integer operations, do you mean integer operations that really are sequential? In other words, the instructions can't be performed in parallel because of data dependencies? If not, then this is already possible with a sufficiently wide superscalar processor or really big SIMD units.
But let's assume you really mean sequential integer operations. The only pipeline stage in this example that can't work on several instructions at once is the execute stage, so I'm assuming that's where the bottleneck is here. This means that the speed is limited by the clock frequency. So, here are two ways to achieve your prediction:
Crank up the clock! Find a way to get it up to 12 GHz without burning up.
Make the execute stage capable of running much faster than the rest of the processor does. This is natural for asynchronous processors; in normal operation the integer functional units will be sitting idle most of the time waiting for input, and the bulk of the time and complexity will be in fetching the instructions, decoding them, scheduling them, and in memory access and I/O. But in your contrived scenario, the integer math units could just go hog wild and the rest of the processor would keep them fed. This can be done with current semiconductor technology, I'm pretty sure.
So, either way, kind of an ambitious prediction. I like it.
Have you not heard that they discovered a way to use graphene as a one to one replacement for copper in chip production. That alone will allow speeds of 12-15GHz.
I would put faster chips using multicore running at many times current speeds will be available by 2011-2012 at near 100% certainty.