gnufx

84 ms·

931.
by gnufx 6y ago
Intel don't release BLIS, though there is some Intel contribution. Substitute libxsmm, which originally beat MKL.
932.
by gnufx 6y ago
Not on ARM, or POWER, you can't. Why you'd want to run it on AMD, I don't understand. I don't know what fraction of peak BLIS and OpenBLAS get, but it will be high.
933.
Over 25% of the UK likely to have had Covid-19 already(manchester.ac.uk)
3 points by gnufx 6y ago|2 comments
934.
by gnufx 6y ago
Yes, like Dongarra's, for instance. I'll stop there.
935.
by gnufx 6y ago
Non-trivial experience in research computing support says that gfortran is surprisingly more reliable than ifort. "Just use GCC and openmpi" has worked in enough cases for me. A number of HPC benchmark lists I've seen have
936.
by gnufx 6y ago
The fact that universities splurge money on these proprietary tools rather than spending it on long-term improvements isn't a good thing, even without the pain of actually supporting them. The typical differences between stuff using if
937.
by gnufx 6y ago
It's probably worth pointing out that LAPACK/BLAS is a strictly Fortran77 interface, which isn't ideal. Unfortunately there's no de facto standard modern interface available.
938.
by gnufx 6y ago
Aspects of Fortran favour it over C even when restrict is used, e.g. as mentioned in https://news.ycombinator.com/item?id=23124451
939.
by gnufx 6y ago
If you use the reference BLAS routines that ship with R, you're losing an order of magnitude serial performance on relevant operations. (The majority of Intel MKL surely isn't written in Fortran, like OpenBLAS and BLIS.)
940.
by gnufx 6y ago
Ho, ho. I've been told by a NAG representative that we should pay them to talk to our own academics who actually do the serious linear algebra! Also I was told they were contracted to do the old AMD proprietary BLAS, which was inferi
941.
by gnufx 6y ago
The company is distinct from their various products. For instance, their POP services are gratis. The NAG library is proprietary but, when I last knew, probably the majority of it was based on code from Netlib etc.
942.
by gnufx 6y ago
Leadership, we've heard of it (in England). Cue news interviews with huge confusion and worry.
943.
Boris Johnson speech: PM unveils 'conditional plan' to reopen society(bbc.co.uk)
1 points by gnufx 6y ago|2 comments
944.
by gnufx 6y ago
That's history today, that is.
945.
by gnufx 6y ago
Yes, I think it's a local site for local people and others who don't need that explaining. That may be more obvious from the home page. (Who is "he" amongst the contributors?)
946.
by gnufx 6y ago
Just per #54400 above. $ gcc --version | head -1 gcc (Debian 8.3.0-6) 8.3.0 $ cat x.c #include <x86intrin.h> double f(__m128d v){return v[1]+v[0];} $ gcc -Ofast -march=haswell -S x.c $ grep -C1 hadd x.s
947.
by gnufx 6y ago
On x86 you can use something like MAQAO or one of the top-down analysis tools for guidance.
948.
by gnufx 6y ago
GCC 8 -Ofast, targeting Haswell, produces vhaddpd %xmm0, %xmm0, %xmm0
949.
by gnufx 6y ago
Right. The Copenhagen University lab at Ris ø in the '80s used a Nord-something and a Regnecentralen RC4000 for data collection and reduction. [As far as I remember, they only exchanged data on tapes, but it was amusing to think of t
950.
by gnufx 6y ago
Note that simple reductions, like level 1 BLAS, are memory-bound, and optimizing is generally only worthwhile for short arrays. (Contrast level 3, i.e. matrix-matrix operations.) Also it's generally arithmetically incorrect to vector
951.
by gnufx 6y ago
Harding. What a wonderful appointment to such a post.
952.
by gnufx 6y ago
Isn't it reassuring, at least, to see that said? Also presumably an opportunity for the right people to help?
953.
by gnufx 6y ago
But unfortunately these alternatives don't have the same encryption support, do they? (Some seem not to have any.)
954.
by gnufx 6y ago
What fraction of communication on a typical HPC system is overlapped with computation? My guess is rather little, despite the Mellanox sales pitch. I do wish people commonly made cluster-global profiling-type measurements to be able to ans
955.
by gnufx 6y ago
Right, if that's all custom means.
956.
by gnufx 6y ago
Point taken, but for some time Mellanox has touted at least cross-campus-type IB. I never persuaded anyone to evaluate it for use between two machine rooms, but there were obvious non-HPC cases for it within them. I'd be interested to
957.
by gnufx 6y ago
There are different types of molecular dynamics that non-ex people run, and different modes of the same code may be more or less latency sensitive (e.g. two "standard" Gromacs benchmark cases I'd have to look up). While I&#x
958.
by gnufx 6y ago
For example, some of us have speculated that there are periods when around half of the time on the UK Archer system (with the previous generation of Cray interconnect) might be spent in one MPI collective. It's a pity there's no m
959.
by gnufx 6y ago
It's odd to class slingshot/dragonfly as less custom for HPC than even infiniband (and Mellanox is all-out on HPC-specific "offload"). Other interconnects designed for HPC include Bull's BXI, Fujitsu's Tofu, an
960.
by gnufx 6y ago
Gbps isn't a unit of latency (or message rate).

More ›