ARM SVE 2048 bit vector

[hansl]

I was reading a little bit about new(ish) ARM Scalable Vector Extension and thought it was interesting that the vector length can range from 128 up to 2048 bits, which is up to the chip implementer. Does anyone know if chips exist that use the full 2048bits, or if we are likely to see any in the near future?

How useful would this be for GIMPS style computations?

[retina]

Originally Posted by hansl

I was reading a little bit about new(ish) ARM Scalable Vector Extension and thought it was interesting that the vector length can range from 128 up to 2048 bits, which is up to the chip implementer. Does anyone know if chips exist that use the full 2048bits, or if we are likely to see any in the near future?

How useful would this be for GIMPS style computations?

Maybe if an algorithm can make a few passes over the data while it is still in the registers then it could benefit, but otherwise for large exponents there is still the need to get the data to/from the main memory over the slower external buses.

Also, all that extra computation work done per cycle will raise the temperature, so they might throttle during intensive computations and destroy any benefits.

[rogue]

Originally Posted by retina

Maybe if an algorithm can make a few passes over the data while it is still in the registers then it could benefit, but otherwise for large exponents there is still the need to get the data to/from the main memory over the slower external buses.

Also, all that extra computation work done per cycle will raise the temperature, so they might throttle during intensive computations and destroy any benefits.

Depending upon the underlying instructions to manipulate data in those registers, something like this could greatly speed up sieving for programs built upon the mtsieve framework. Unfortunately I do not know ARM assembly nor do I have an ARM CPU to test any code on. If ARM had intrinsics, that could make a difference as it wouldn't require deep knowledge of assembly.

[ewmayer]

I believe the 2048-bit-ness is by way of future-proofing the ISA more than "we expect someone to implement this soon", but I am reasonably hopeful of seeing some kind of broad-scale 256-bit deployment soon. There've been rumors of Apple switching their PCs to ARM, and a 256-bit implementation would make sense there. Also, smartphones keep getting data-hungrier.

[ldesnogu]

Originally Posted by rogue

If ARM had intrinsics, that could make a difference as it wouldn't require deep knowledge of assembly.

ARM does have intrinsics, but they obviously are different than the ones on Intel CPU, since an intrinsic more or less maps to an instruction of your CPU.


As far as SVE being able to handle up to 2048-bit, that means that registers can be that large depending on what the CPU designer decides. Once vector length is fixed, this doesn't mean your units will be able to process all the vector at once (think of AMD splitting 256-bit vectors into two pieces of 128-bit).


What SVE brings to the table is that the code doesn't have to be different depending on vector length, contrary to x86 where you need different code paths for 128-bit, 256-bit and 512-bit. Note that doesn't mean you'll get the fastest possible code with a single path on ARM, but that's another story.