Re:The big six matrix factorizations
by the_svd_doctor on Wednesday May 18, 2022 @11:53AM
(#31424286)
It's always been very ambiguous in practice. To avoid ambiguity I usually use `flop/s` but not everyone likes that :)
[ Parent
| on HN ]
Re:The big six matrix factorizations
by MatteoFrigo on Wednesday May 18, 2022 @02:29PM
(#31426062)
Speaking of SVD doctors, I heard many years ago (from Alan Edelman) that Gene Golub's license plate used to be "DR SVD". Later he switched to "PROF SVD".
I couldn't confirm the DR SVD part, but the PROF SVD story appears to be real: https://www.mathworks.com/company/newsletters/articles/profe...
As I recall the Dr SVD was the license on his second car.
He once left the car at SFO airport and when he heard I was flying
in, he asked me to drive it to his house (keys under the seat!)
What he didn't tell me was that it was a "stick shift" and I only once
before had a friend teach me how that works. After driving around
San Francisco for Gene I got better at it, but it was definitely scary at times.
:-)
Re:The big six matrix factorizations
by the_svd_doctor on Wednesday May 18, 2022 @04:59PM
(#31427577)
I heard something very similar from some of his very close (ex-)colleagues, so it appears to be true :)
> To avoid ambiguity I usually use `flop/s` but not everyone likes that :)
Flop/s makes no sense and is outright wrong. The whole point of flops is to express how many floating point operations are required by an algorithm. How many operations are performed per second is a property of the hardware you're using to run an implementation of the algorithm.
You want to express computational complexity in terms of floating point operations.
Re:The big six matrix factorizations
by the_svd_doctor on Wednesday May 18, 2022 @05:01PM
(#31427606)
Well, yeah, I use it in the context of floating-points-operations-per-seconds. That's the most common use in my field. I was replying to the parent comment. No need to use this kind of tone.
Maybe the_svd_doctor uses flop/s to describe hardware, like you said
The number of operations per second varies by at least an order of magnitude on the same hardware depending on the GEMM algorithm you use (reference or Goto), and you quote the performance of the implementation in terms of FLOP/s knowing the number of FLOPs required by the computation. That makes sense to people who implement and measure these things.