4 ms·

This is because CPU’s haven’t gotten any faster in almost 20 years because of the problem of heat dissipation, so all that can be done to make computers feel mo

This is because CPU’s haven’t gotten any faster in almost 20 years because of the problem of heat dissipation, so all that can be done to make computers feel more powerful is to get better at writing parallelizable software… which is just a very hard problem for humans to tackle. Also, the latency to/from main memory is the main bottleneck now for all jobs whether or not they are parallelizable. And it is very hard to get insight into how you can organize things at a high level to optimize for low-level memory usage… humans aren’t good at that either. The problem is that now gains in the experience of using computers have to come from humans getting better at writing software in light of hardware constraints. You can’t get a whole lot these days from throwing new hardware at the problem.

dexwiz 3y ago
I don't get how you claim CPUs are the same as 20 years ago? Maybe the clock speed is the same, but they are demonstrably denser. Lithography continues to improve as extreme costs, so wins are eaked out. The numbers don't seem to change much, but that is because we are approaching a limit of what a transistor can do with current approaches, and alternative approaches are literally millennia of engineering hours behind. Intel and x86(_64) has had a hold for two decades, but GPUs and AArch64 are making inroads addressing memory bandwidth issues.
- apantel 3y ago
  I was basically saying that there are no more order-of-magnitude type gains to be had in CPU performance (edit: with current technology, as far as I’m aware). A top-of-the-line processor today is maybe 3x more capable than one from a decade ago. And that doesn’t help with the real bottleneck of accessing memory, especially main memory, which is bound by the laws of physics. The CPU can’t do anything until it has the data, no matter how fast it is. Also, CPU’s are eking out wins by doing more and more complicated things like speculative execution and out-of-order execution, which makes it harder and harder for the programmer at a high level to reason about performance and actually write software that gets the most out of the hardware. Add in the fact that software is tending to get more and more bloated without equal energy being invested in making it more performant… My argument is that major performance improvements are only hard-won these days by really dedicated developers because we’re getting diminishing returns from hardware and it’s harder than ever to write performant software.
  - johngossman 3y ago
    CPU performance isn't making 2x leaps as quickly as it did in the 80s and 90s, but you're vastly understating the performance and power efficiency improvements that have been made over the last decade.
johngossman 3y ago
This is only partly true. Yes, we hit the heat wall, and what you say was true for a few years, but there have been a lot of other efficiency improvements. A single core of 3-3.5 GHz M chip is vastly faster than a 4GHz Pentium 4, even not accounting for bandwidth (which is also vastly faster).

dexwiz 3y ago

I don't get how you claim CPUs are the same as 20 years ago? Maybe the clock speed is the same, but they are demonstrably denser. Lithography continues to improve as extreme costs, so wins are eaked out. The numbers don't seem to change much, but that is because we are approaching a limit of what a transistor can do with current approaches, and alternative approaches are literally millennia of engineering hours behind. Intel and x86(_64) has had a hold for two decades, but GPUs and AArch64 are making inroads addressing memory bandwidth issues.

apantel 3y ago
I was basically saying that there are no more order-of-magnitude type gains to be had in CPU performance (edit: with current technology, as far as I’m aware). A top-of-the-line processor today is maybe 3x more capable than one from a decade ago. And that doesn’t help with the real bottleneck of accessing memory, especially main memory, which is bound by the laws of physics. The CPU can’t do anything until it has the data, no matter how fast it is. Also, CPU’s are eking out wins by doing more and more complicated things like speculative execution and out-of-order execution, which makes it harder and harder for the programmer at a high level to reason about performance and actually write software that gets the most out of the hardware. Add in the fact that software is tending to get more and more bloated without equal energy being invested in making it more performant… My argument is that major performance improvements are only hard-won these days by really dedicated developers because we’re getting diminishing returns from hardware and it’s harder than ever to write performant software.
- johngossman 3y ago
  CPU performance isn't making 2x leaps as quickly as it did in the 80s and 90s, but you're vastly understating the performance and power efficiency improvements that have been made over the last decade.

johngossman 3y ago

This is only partly true. Yes, we hit the heat wall, and what you say was true for a few years, but there have been a lot of other efficiency improvements. A single core of 3-3.5 GHz M chip is vastly faster than a 4GHz Pentium 4, even not accounting for bandwidth (which is also vastly faster).