7 Mar
2023
7 Mar
'23
1:08 p.m.
Larry McVoy <lm(a)mcvoy.com> wrote:
It's funny because back in my day those GPUs
would have been called vector
processors, at least I think they would. It seems like somewhere along
the way, vector processors became a dirty word but GPUS are fine.
Color me confused.
I think vector processing is used for things like
for (i = 0; i < SOME_CONSTANT; i++)
a[i] = b[i] + c[i]
that is, vectoring general purpose code. GPUS are pretty specialized
SIMD machines which sort of happen to be useful for certain kinds
of parallelizable general computations, like password cracking.
Today there are both standardized and proprietary ways of programming
them.
About the only reason I can see to keep things divided
between the CPU
and the GPU is battery power, or power consumption in general. From
what little I know, it seems like GPUs are pretty power thirsty so
maybe they keep them as optional devices so people who don't need them
don't pay the power budget.
But even that seems suspect, I would think they could put some logic
in there that just doesn't feed power to the GPU if you aren't using
it but maybe that's harder than I think.
You're on target, not just in the GPU world but also in the CPU
world. Modern Intel CPUS have a lot of circuits for turning power
consumption up and down dynamically.
Modern-day CPU development is much harder than we software types generally
realize. I worked at Intel as a software guy (bad juju there, let me
tell you!) and learned a lot about it, from the outside. For a given
x86 microarchitecture, from planning until it's in the box in the store
is like a 5+ year journey. These days maybe even more, as I left
Intel 7.5 years ago.
Arnold