2 Mar
2021
2 Mar
'21
4:06 a.m.
On Mon, Mar 01, 2021 at 10:40:41PM -0800, Boris Gimbarzevsky via cctalk wrote:
[...]
Out of curiousity, decided to benchmark one of my old,
really cheap PC
laptops that got in 2010 and it managed 30 Mflops using double precision
arithmetic. 10 Mflop performance no longer as impressive as it used to be.
I'll say. Modern kit gets 1 FLOPS per MHz per core, give or take an order of
magnitude depending on the specific architecture. That machine must have
been appallingly bad to only manage 30 MFLOPS. Although perhaps you meant
GFLOPS, in which case it sounds about right.
The Haswell CPU in my 2014-vintage laptop manages "up to" 147 GFLOPS. Which
is an order of magnitude slower than its GPU. Useful FLOPS for scientific
computing rather than contrived numbers for benchmarking may well lose an
order of magnitude or two in overhead, but it's still not hanging about.