Scientific software, the joke goes, is your problem somehow turned into matrix algebra. And that algebra runs on highly optimized libraries born in the Fortran era: BLAS and LAPACK.
November 10, 2026 at 8:01 AM
Scientific software, the joke goes, is your problem somehow turned into matrix algebra. And that algebra runs on highly optimized libraries born in the Fortran era: BLAS and LAPACK.
Along the way: how BLAS, LAPACK and NumPy fit together, and ideas for going beyond today’s performance to use the hardware we own but rarely exploit in full. No expertise needed beyond curiosity about the foundations your code runs on. Get your ticket!
📍 Compute! Paris 2026 — Nov 25–26
📍 Compute! Paris 2026 — Nov 25–26
Compute! Paris 2026
Understanding your tools and algorithms is what gives you agency in a digital world. Paris, November 25–26, 2026.
compute.events
November 10, 2026 at 8:01 AM
Along the way: how BLAS, LAPACK and NumPy fit together, and ideas for going beyond today’s performance to use the hardware we own but rarely exploit in full. No expertise needed beyond curiosity about the foundations your code runs on. Get your ticket!
📍 Compute! Paris 2026 — Nov 25–26
📍 Compute! Paris 2026 — Nov 25–26
The DOE data center I visited at 2013 SuperCompute was a Net Zero consumer with solar panels on top of every structure and parking lot, water cooled raised floor (clamping water cooling at the rails! No fluid at the board!). If you caught a chill just run LAPACK :)
October 31, 2026 at 6:09 PM
The DOE data center I visited at 2013 SuperCompute was a Net Zero consumer with solar panels on top of every structure and parking lot, water cooled raised floor (clamping water cooling at the rails! No fluid at the board!). If you caught a chill just run LAPACK :)
We could refactor some BLAS-2 calls (matrix × vector) into BLAS-3 calls (matrix × matrix), be more aggressive in managing memory (swapping buffers, using in-place BLAS/LAPACK, etc.), exploit upper triangle storage, and reuse existing genotype streaming code from GCTA's "GRM V2".
October 29, 2026 at 10:12 AM
We could refactor some BLAS-2 calls (matrix × vector) into BLAS-3 calls (matrix × matrix), be more aggressive in managing memory (swapping buffers, using in-place BLAS/LAPACK, etc.), exploit upper triangle storage, and reuse existing genotype streaming code from GCTA's "GRM V2".
MKL also seems to have non-reference implementations of several eigensolvers, so even though AOCL was better for BLAS (our cluster has AMD CPUs), it was much worse for LAPACK. So using some mash-up of libaries was the key to squeeze the most performance out.
October 29, 2026 at 10:12 AM
MKL also seems to have non-reference implementations of several eigensolvers, so even though AOCL was better for BLAS (our cluster has AMD CPUs), it was much worse for LAPACK. So using some mash-up of libaries was the key to squeeze the most performance out.
biddy blas implies the existence of biddy lapack
October 21, 2026 at 1:31 PM
biddy blas implies the existence of biddy lapack
Se usan las bibliotecas que se escribieron en fortran hace 60 años, LAPACK es una de ellas, si mal no recuerdo. Están tan cagadas a palos y escritas pensando en optimizar recursos que son seguras y eficientes.
September 21, 2026 at 9:00 PM
Se usan las bibliotecas que se escribieron en fortran hace 60 años, LAPACK es una de ellas, si mal no recuerdo. Están tan cagadas a palos y escritas pensando en optimizar recursos que son seguras y eficientes.
A new Linpack/HPL benchmark rewritten in Python. One readable file, numpy as the only dep — and at the end it tells you when your laptop would have been a supercomputer. Mine was world #1 in November 1995 🤯
github.com/dzaragoza/hpl
github.com/dzaragoza/hpl
GitHub - dzaragoza/hpl: The Linpack/HPL benchmark in readable Python — LAPACK through NumPy, with a TOP500 "when was my laptop a supercomputer?" calibration.
The Linpack/HPL benchmark in readable Python — LAPACK through NumPy, with a TOP500 "when was my laptop a supercomputer?" calibration. - dzaragoza/hpl
github.com
September 20, 2026 at 10:58 PM
A new Linpack/HPL benchmark rewritten in Python. One readable file, numpy as the only dep — and at the end it tells you when your laptop would have been a supercomputer. Mine was world #1 in November 1995 🤯
github.com/dzaragoza/hpl
github.com/dzaragoza/hpl
yeah, like, I'm not sure how to read that screenshot other than "gpus are alchemy but blas and lapack get a pass"
September 5, 2026 at 12:35 AM
yeah, like, I'm not sure how to read that screenshot other than "gpus are alchemy but blas and lapack get a pass"
LAPACK/BLAS have a neat vector normalisation trick: find the max axis and divide each component by that before rsqrt and scale. Avoids overflow and gradual underflow due to squaring in the rsqrt. If the max axis is tiny then all axes are tiny so the components can be considered roughly proportional.
October 9, 2025 at 9:58 AM
LAPACK/BLAS have a neat vector normalisation trick: find the max axis and divide each component by that before rsqrt and scale. Avoids overflow and gradual underflow due to squaring in the rsqrt. If the max axis is tiny then all axes are tiny so the components can be considered roughly proportional.
What's up?
gfortran
LAPACK
BLAS
Ongoing Homebrew conflicts
Stores are closed and I am running low on booze.
gfortran
LAPACK
BLAS
Ongoing Homebrew conflicts
Stores are closed and I am running low on booze.
December 25, 2024 at 11:47 PM
What's up?
gfortran
LAPACK
BLAS
Ongoing Homebrew conflicts
Stores are closed and I am running low on booze.
gfortran
LAPACK
BLAS
Ongoing Homebrew conflicts
Stores are closed and I am running low on booze.
Scientific computing in the browser is closer than you think. Gunj Joshi's #HPSFCon2026 talk on stdlib bringing BLAS/LAPACK to JavaScript — with live demos and a perf comparison vs Python/C.
Recap: https://bit.ly/42nOJ24
Recap: https://bit.ly/42nOJ24
April 30, 2026 at 3:12 PM
Scientific computing in the browser is closer than you think. Gunj Joshi's #HPSFCon2026 talk on stdlib bringing BLAS/LAPACK to JavaScript — with live demos and a perf comparison vs Python/C.
Recap: https://bit.ly/42nOJ24
Recap: https://bit.ly/42nOJ24
Cannot locate lapacke.h on windows [duplicate] What is the common way to configure lapack for CMa...
https://stackoverflow.com/questions/79410976/cannot-locate-lapacke-h-on-windows
#c++ #windows #cmake #lapack
Event Attributes
https://stackoverflow.com/questions/79410976/cannot-locate-lapacke-h-on-windows
#c++ #windows #cmake #lapack
Event Attributes
February 4, 2025 at 8:44 AM
Cannot locate lapacke.h on windows [duplicate] What is the common way to configure lapack for CMa...
https://stackoverflow.com/questions/79410976/cannot-locate-lapacke-h-on-windows
#c++ #windows #cmake #lapack
Event Attributes
https://stackoverflow.com/questions/79410976/cannot-locate-lapacke-h-on-windows
#c++ #windows #cmake #lapack
Event Attributes
compared with the LAPACK Complete Orthogonal Decomposition algorithm, among others. The obtained error with both algorithms was similar but the computational cost was at least 20 % smaller with the new algorithm, even though [5/6 of https://arxiv.org/abs/2501.17696v1]
January 30, 2025 at 6:14 AM
compared with the LAPACK Complete Orthogonal Decomposition algorithm, among others. The obtained error with both algorithms was similar but the computational cost was at least 20 % smaller with the new algorithm, even though [5/6 of https://arxiv.org/abs/2501.17696v1]
very cool! as someone who wrote a bunch of FEM in the past (for wave scattering) I appreciate the aside on bandwidth reduction 🥶
great work with the js port of BLAS/LAPACK too!
great work with the js port of BLAS/LAPACK too!
July 22, 2026 at 7:41 PM
very cool! as someone who wrote a bunch of FEM in the past (for wave scattering) I appreciate the aside on bandwidth reduction 🥶
great work with the js port of BLAS/LAPACK too!
great work with the js port of BLAS/LAPACK too!
Why do we need libraries like FlexiBLAS?
Michał Górny's latest blog post explains how multiplexing libraries solve the API/ABI compatibility puzzle in BLAS/LAPACK packaging, making life easier for users, developers, and distributions alike.
Learn how it all fits together: buff.ly/wo861po
Michał Górny's latest blog post explains how multiplexing libraries solve the API/ABI compatibility puzzle in BLAS/LAPACK packaging, making life easier for users, developers, and distributions alike.
Learn how it all fits together: buff.ly/wo861po
BLAS/LAPACK packaging
An overview of differences between BLAS / LAPACK implementations and their packaging.
labs.quansight.org
November 13, 2025 at 4:13 PM
Why do we need libraries like FlexiBLAS?
Michał Górny's latest blog post explains how multiplexing libraries solve the API/ABI compatibility puzzle in BLAS/LAPACK packaging, making life easier for users, developers, and distributions alike.
Learn how it all fits together: buff.ly/wo861po
Michał Górny's latest blog post explains how multiplexing libraries solve the API/ABI compatibility puzzle in BLAS/LAPACK packaging, making life easier for users, developers, and distributions alike.
Learn how it all fits together: buff.ly/wo861po
Exciting work on BLAS/LAPACK packaging!📦
As part of bringing better BLAS/LAPACK support to Meson, Michał Górny learned a lot about how these essential libraries differ across implementations and distributions, and he is sharing it all with the community.
📖 Dive into the details: buff.ly/JAxtRop
As part of bringing better BLAS/LAPACK support to Meson, Michał Górny learned a lot about how these essential libraries differ across implementations and distributions, and he is sharing it all with the community.
📖 Dive into the details: buff.ly/JAxtRop
BLAS/LAPACK packaging
An overview of differences between BLAS / LAPACK implementations and their packaging.
labs.quansight.org
October 23, 2025 at 5:24 PM
Exciting work on BLAS/LAPACK packaging!📦
As part of bringing better BLAS/LAPACK support to Meson, Michał Górny learned a lot about how these essential libraries differ across implementations and distributions, and he is sharing it all with the community.
📖 Dive into the details: buff.ly/JAxtRop
As part of bringing better BLAS/LAPACK support to Meson, Michał Górny learned a lot about how these essential libraries differ across implementations and distributions, and he is sharing it all with the community.
📖 Dive into the details: buff.ly/JAxtRop
Check all the r and r package versions. Also the linear modelling packages like lapack might be different
August 7, 2025 at 7:52 AM
Check all the r and r package versions. Also the linear modelling packages like lapack might be different
(for being accurate, anyway, the other faster ones have tradeoffs)
lapack, on the other hand....I guess libflame exists but ive never seen a reason to use it
lapack, on the other hand....I guess libflame exists but ive never seen a reason to use it
May 27, 2026 at 3:28 PM
(for being accurate, anyway, the other faster ones have tradeoffs)
lapack, on the other hand....I guess libflame exists but ive never seen a reason to use it
lapack, on the other hand....I guess libflame exists but ive never seen a reason to use it
my understanding is that at this point the blas kernels are all best implemented in c/c++ libs right now but netlib lapack is still basically the standard outside of proprietary stuff
January 4, 2026 at 4:44 AM
my understanding is that at this point the blas kernels are all best implemented in c/c++ libs right now but netlib lapack is still basically the standard outside of proprietary stuff
incidentally this is why the common linear algebra ABI (blas/lapack) still uses a bunch of absolutely inscrutable 5 letter function names
October 27, 2025 at 10:22 PM
incidentally this is why the common linear algebra ABI (blas/lapack) still uses a bunch of absolutely inscrutable 5 letter function names