Conclusion
Hardware can only deliver “efficiency miracles” for operations that are fundamentally cheap to begin with. This is done by lowering dispatching costs and so increasing throughput per unit of energy. The price paid is reduced flexibility.
This is comparing different metrics against each other, in effect not coming to any conclusion. Yes, things will be “cheaper” by one metric and more costly by a different metric. This is fundamental to anything for which you can provide an economic cost model. Is that all this is trying to say?
Many people seem to believe that implementing something in hardware, regardless of complexity, makes it faster, cheaper, and therefore “better”. The conclusion that there is a tradeoff, and that that performance boosts comes from lowering dispatching costs and increasing parallelism, is apparently not an obvious one to many people.
I think the article says quite a bit more about the complex relationship between (economic) prices and (physical) costs. I’ll refrain from summarizing here, but I found it quite thoughtful.
I’ll also highlight hardware development on nodes that CPU’s use is extremely expensive. I’m talking millions for a chip. They’re better off buying a cluster and/or paying geniuses to optimize their software in most cases.
The value of this post is that he dispels the “myth” that everything can/should/will be implemented in hardware (yes, I’ve been in such a discussion). I think his bulletpoints, “specialization” and “parallelization”, are golden.
This is comparing different metrics against each other, in effect not coming to any conclusion. Yes, things will be “cheaper” by one metric and more costly by a different metric. This is fundamental to anything for which you can provide an economic cost model. Is that all this is trying to say?
Many people seem to believe that implementing something in hardware, regardless of complexity, makes it faster, cheaper, and therefore “better”. The conclusion that there is a tradeoff, and that that performance boosts comes from lowering dispatching costs and increasing parallelism, is apparently not an obvious one to many people.
I think the article says quite a bit more about the complex relationship between (economic) prices and (physical) costs. I’ll refrain from summarizing here, but I found it quite thoughtful.
I’ll also highlight hardware development on nodes that CPU’s use is extremely expensive. I’m talking millions for a chip. They’re better off buying a cluster and/or paying geniuses to optimize their software in most cases.
The value of this post is that he dispels the “myth” that everything can/should/will be implemented in hardware (yes, I’ve been in such a discussion). I think his bulletpoints, “specialization” and “parallelization”, are golden.
[Comment removed by author]