4 ms·

> There are many who say we are running out of the benefits of parallelization for CPUs and that soon, adding cores won’t get us much in terms of added performa

> There are many who say we are running out of the benefits of parallelization for CPUs and that soon, adding cores won’t get us much in terms of added performance. Are people really saying that? There's a lot of discussion about languages and software that have trouble taking advantage of many cores, but I've never heard anyone saying that the benefits of parallelisation are going to reach a limit soon? Perhaps that's what the author meant in shorthand.

jagger27 10y ago
How applicable is Amdahl's Law to this discussion? https://en.wikipedia.org/wiki/Amdahl%27s_law https://en.wikipedia.org/wiki/Amdahl%27s_law
- dom0 10y ago
  Backend services like databases aside most stuff that runs on servers today and requires most of the CPU and bandwidth is embarrasingly parallel = web application servers. Similarly a lot of stuff which workstations are used for are (comparatively) easy to parallelize, e.g. rendering and many kinds of simulations (fluidics, FEM, ...).
- Denzel 10y ago
  Amdahl's law applies to relationships, e.g., a process/program is a grouping of related components. By definition, if there are no relationships between components, then Amdahl's law does not apply (or rather there's 0% serialization), and you'll always benefit from more parallelization. For example, my typesetter doesn't care that my compiler is running which doesn't care that my web server is running. There's no relationship between the inputs or outputs of those processes. And there's no synchronization. Therefore, all else equal, as long as the cost per core continues to decrease, it's beneficial to add more cores: because you can run more things.
dom0 10y ago
> Are people really saying that? Nope, especially not on servers. Of course Intel will have to scale up the core interconnects at some point and might also need to scale I/O and memory busses up in a similar vein as AMD does in Naples. Note that this article is about Xeon E3 though, which is not the mainstream server platform, but is usually used for lowish-end workstations and entry-level servers.
felipemnoa 10y ago
AI is the next big thing and is where parallelization really shines. Physics simulations is probably another big one. New video games will probably really start to take advantage of parallelization as they continue to get bigger and more immersive. Heck, I wish my crappy little Mackbook pro had more cores for the AI project I'm working on. We are really just getting started.
jackmott 10y ago
memory bandwidth gets to be a limiter quickly with many cores for certain workloads.

jagger27 10y ago

How applicable is Amdahl's Law to this discussion? https://en.wikipedia.org/wiki/Amdahl%27s_law https://en.wikipedia.org/wiki/Amdahl%27s_law

dom0 10y ago
Backend services like databases aside most stuff that runs on servers today and requires most of the CPU and bandwidth is embarrasingly parallel = web application servers. Similarly a lot of stuff which workstations are used for are (comparatively) easy to parallelize, e.g. rendering and many kinds of simulations (fluidics, FEM, ...).
Denzel 10y ago
Amdahl's law applies to relationships, e.g., a process/program is a grouping of related components. By definition, if there are no relationships between components, then Amdahl's law does not apply (or rather there's 0% serialization), and you'll always benefit from more parallelization. For example, my typesetter doesn't care that my compiler is running which doesn't care that my web server is running. There's no relationship between the inputs or outputs of those processes. And there's no synchronization. Therefore, all else equal, as long as the cost per core continues to decrease, it's beneficial to add more cores: because you can run more things.

dom0 10y ago

> Are people really saying that? Nope, especially not on servers. Of course Intel will have to scale up the core interconnects at some point and might also need to scale I/O and memory busses up in a similar vein as AMD does in Naples. Note that this article is about Xeon E3 though, which is not the mainstream server platform, but is usually used for lowish-end workstations and entry-level servers.

felipemnoa 10y ago

AI is the next big thing and is where parallelization really shines. Physics simulations is probably another big one. New video games will probably really start to take advantage of parallelization as they continue to get bigger and more immersive. Heck, I wish my crappy little Mackbook pro had more cores for the AI project I'm working on. We are really just getting started.

jackmott 10y ago

memory bandwidth gets to be a limiter quickly with many cores for certain workloads.