The authors have pointed out an embedded processor spends only little portion of its energy on the actual computation. The figure shows that the processor spends most of its energy on instruction and data supply. 70% of the energy is consumed to supply instruction (42%) and data (28%).
This comment was marked helpful 0 times.
retterermoore
I understand why the code needs to be compute-bound for speedup to occur, but why is floating-point math a problem? Wouldn't that make it more computation intensive than integer math and thus more able to be sped up by parallelism?
This comment was marked helpful 0 times.
spilledmilk
I believe the restriction on floating-point math is due to the fact that processing floating point arithmetic is difficult and therefore, the performance of the ASIC would not reach the ~100x or greater improvement in perf/watt.
Floating point operators generally have dramatically increased logic utilization and power consumption, combined with lower clock speed, longer pipelines, and reduced throughput capabilities when compared to integer or fixed point.
The figure originates from this article.
The authors have pointed out an embedded processor spends only little portion of its energy on the actual computation. The figure shows that the processor spends most of its energy on instruction and data supply. 70% of the energy is consumed to supply instruction (42%) and data (28%).
This comment was marked helpful 0 times.
I understand why the code needs to be compute-bound for speedup to occur, but why is floating-point math a problem? Wouldn't that make it more computation intensive than integer math and thus more able to be sped up by parallelism?
This comment was marked helpful 0 times.
I believe the restriction on floating-point math is due to the fact that processing floating point arithmetic is difficult and therefore, the performance of the ASIC would not reach the ~100x or greater improvement in perf/watt.
Quoting this page,
This comment was marked helpful 0 times.