When simulating scaled down result, why would the results not predict reality? If this simulated machine is identical to the real machine and we run the most complex, memory intensive problem that is feasible on the machine then shouldn't it predict reality? I understand that the simulation might be imperfect but if it is near perfect then it should run the same.
jmackama
@acortes,I think problems usually come from the stuff that does not stay the same when you scale. For example, the logic for each ILP datapath might work exactly the same regardless of the number of datapaths, but the logic for managing them and interfacing them with the rest of the chip will need to change slightly to support each number of datapaths. I imagine you could use either a less expensive abstraction for the simulation, or do a unit test of the scaled up version to verify that small part of the logic still works at scale.
When simulating scaled down result, why would the results not predict reality? If this simulated machine is identical to the real machine and we run the most complex, memory intensive problem that is feasible on the machine then shouldn't it predict reality? I understand that the simulation might be imperfect but if it is near perfect then it should run the same.
@acortes,I think problems usually come from the stuff that does not stay the same when you scale. For example, the logic for each ILP datapath might work exactly the same regardless of the number of datapaths, but the logic for managing them and interfacing them with the rest of the chip will need to change slightly to support each number of datapaths. I imagine you could use either a less expensive abstraction for the simulation, or do a unit test of the scaled up version to verify that small part of the logic still works at scale.