Even though this podcast went for three hours, there are many things I wish I'd explained better. So I will attempt a better explanation of my kind-of-throwaway Amdahl's Law comment, as the situation is pretty interesting:https://twitter.com/oxidecomputer/status/1221804818030706688 …
-
-
very much faster than one of the cores on the 32-core CPU. So when Intel wants to deliver more CPU power to customers, the main way they know how to do it is to add cores. Many people can't use most of those cores effectively, but some people can, so there is still a reason
Prikaži ovu nit -
to do this. But, back in the early 2000s, the situation was different. We still knew how to make single processors faster by making their pipelines deeper and making them more out-of-order and speculative. But that was starting to take a lot of transistors, so there was this
Prikaži ovu nit -
idea, hey, instead of using all those transistors to do this out-of-order stuff, we could just have (in the case of PS3) a bunch of blazing-fast SPU coprocessors that are very simple and linear, or (in the case of X360) three in-order main processors instead.
Prikaži ovu nit -
Unlike today, where we don't have a tradeoff to make ... we just know how to add more cores, we can't make one core faster ... back in that time a tradeoff was made: we can have more potential compute power, *as long as we make the slow part of our system slower*.
Prikaži ovu nit -
Because in-order cores are slower than out-of-order cores, potentially *much* slower in some cases. But people decided to do that. And this runs afoul of Amdahl's Law ... with it you should see clearly that making B twice as slow as it could be (or whatever) is going to be
Prikaži ovu nit -
much worse than making A faster, in almost all cases, unless B is a tiny part of your program, which it never is, especially for games. Both consoles made anti-Amdahl tradeoffs; I think both would have been much better with just a single out-of-order CPU at the same price point.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
We can definitely make processors with higher single-threaded performance, they just won't be compatible with existing ISAs aside from maybe some VLIWs. Software pipelining can easily get you a 3-4x improvement, but modern CPUs don't make it practical outside of trivial cases.
-
There's also this: https://youtu.be/MHZDXC4zJ0c Not a way to eliminate Amdahl's law, but a really promising way to speculate around it at low-entropy points in a program.
- Još 10 drugih odgovora
Novi razgovor -
-
-
Not the reason ... The principal reason of the CPU cores size is the fact that the diagonal of your elemental clusters needs to be one cycle « time » max. Then, you build around this ... performance is a compromise between frequency and IPC (<—- parallelism extraction)
-
Parallelism extraction is power consuming for less return on investment than massive threading a la GPU. OoO is expensive in term of power, more hard work than SMT for example.
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
New On the Metal episode w/