Sujet : Re: Is Intel exceptionally unsuccessful as an architecture designer?
De : anton (at) *nospam* mips.complang.tuwien.ac.at (Anton Ertl)
Groupes : comp.archDate : 18. Sep 2024, 06:40:07
Autres entêtes
Organisation : Institut fuer Computersprachen, Technische Universitaet Wien
Message-ID : <2024Sep18.074007@mips.complang.tuwien.ac.at>
References : 1 2 3
User-Agent : xrn 10.11
mitchalsup@aol.com (MitchAlsup1) writes:
Just last night, I was in a conversation with someone trying to start up
a new company that wants to compete in the "server market". Direct
quote;
"the CPUs are simply I/O managers to the Inference Engines and GPUs."
The investor pitch of a startup that cannot compete on the CPU side.
Who here thinks that CPUs have become the CDC 6600 PPs for the GPUs and
Inference Engines ??
The big difference is, that despite the CDC6600 being a supercomputer,
the majority of its software (measured in lines of code) ran on the
CPU, not on the PPs.
Later supercomputers had a general-purpose computer for doing all the
menial stuff, so that the supercomputer could focus on the matrix
multiplies that it was good at. I can believe that on those systems
the software on the GP computer was bigger than the software on the
supercomputer.
For some machines (e.g., high-end game PCs, game consoles and whatever
is done for training deep neural networks), I expect that the relation
between the CPU and its accelerators are similar to those between the
general-purpose and the supercomputer: Most power spent in the
accelerator, most lines of code run on the CPU. Although, thinking
again, are accelerators used for training, or only for running DNNs?
For servers running data bases, web shops, social networks, etc., and
for, e.g., people doing C++ or Rust development, I expect that
accelerators play little role.
Just last night I talked to Jens Palsberg who works on quantum
computing. He told me that those working on quantum computer
simulators find it too hard to program GPUs, so optimizations for CPUs
are in demand. OTOH, earlier that day he gave a presentation on a
different topic where he worked on an optimization that allowed to
eliminate branches; and his collaborators, who work for Meta, were
very intent on eliminating all branches, because that would allow them
to run on specialized hardware, i.e., an accelerator.
- anton
-- 'Anyone trying for "industrial quality" ISA should avoid undefined behavior.' Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>