1

Why some supercomputers use ARM cpus? For example the world's most powerful supercomputer: Fugaku. It uses ARM cpus. x86 cpus can process more complex stuff right?

4
  • 1
    Well, no. But even if that were right, complex is not the same as fast.
    – Darren
    Aug 5 '20 at 18:29
  • Even Apple are moving entirely to ARM from Intel, so there's got to be some sense in it… but you'd have to ask them why.
    – Tetsujin
    Aug 5 '20 at 18:30
  • @Tetsujin Apple wants to move to ARM for these 2 reasons: They want to ditch intel processors and make their own ones so they have COMPLETE control over them, and secondly, they want all of their devices to be powered by the same technology BUT since x86 processors are not ideal for smartphones, tablets and other small devices, in this case its better to use ARM. It allows for MUCH better cross-compatibility ETC Aug 5 '20 at 18:42
  • 1
    Why wouldn’t they? They could also use POWER9, SPARC…. Well, they do, in fact.
    – Daniel B
    Aug 5 '20 at 18:57
1

Because ARM is exponentially more power efficient than x86, as well as producing less heat and requiring less space per core.

There's no real difference in what they can compute (they're both Turing complete instruction sets), the only difference is how many instructions it takes (and in theory how fast those instructions run). However, even though x86 can (sometimes) compute the same things with fewer instructions, that doesn't matter much when dealing with typical workloads for a supercomputer. For such things, parallelization is your key metric (or, alternatively, supercomputer workloads scale out exponentially better than they scale up), not how fast each individual thread of execution can run. For this type of workload, f you can have 4 million threads of execution at 1.8GHz or 2 million at 2.5GHz for the same power consumption, you pick the higher thread count because it will do a better job of running the required workload.

You can actually see this to some extent in a lot of stuff on the Top500 list, they almost always use CPUs that have sub-par clock frequencies compared to what you would find in a desktop system, but have exponentially more threads of execution. A number of the systems also leverage GPUs to do most of the actual processing, which do even better at parallelizing things than CPUs do.

In the particular case of RIKEN's Fugaku, this is taken to an extreme. Because of how lightweight ARM is in terms of power consumption, space, and heat dissipation together that not only does it completely blow the competition on the Top500 list out of the water (it's got literally more than twice the processing power of the number two entrant on the list), it also takes up less space and uses significantly less energy per unit of processing power, although it's also cost far more than most of the others on the list as well.

1

ARM has made enormous strides forward in supercomputer technology.

The world's fastest computer is currently the Japanese Fugaku, currently number one on the TOP500 list. It is also number one on the Green500 list with the lowest power consumption per processor.

This is because thanks to ARM technology, Fugaku has no accelerators, network cards or main memory on the motherboard. Everything has been sucked into the processor package, allowing extremely low power consumption.

In addition, ARM has developed the Scalable Vector Extension (SVE) that enables advanced vectorizing compilers to extract more fine-grain parallelism from existing code and reduce software deployment effort. ARM has made available these compilers for Linux.

How long ARM will stay at the head of the pack is unknown, as the other constructors are making large efforts to outdo it.

For more information see:

0
-1

ARM can have higher performance at lower power than x86. For example, look at Amazon's Graviton, which is quite favourably comparable to x86 on most workloads.

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.