Gotta respect the commitment to parallelization
November 12, 2026 at 10:32 PM
Gotta respect the commitment to parallelization
i.e. the tasks already ran in parallel and adding parallelization on the sub-task level did not actually do anything other than make it slower
November 10, 2026 at 3:57 PM
i.e. the tasks already ran in parallel and adding parallelization on the sub-task level did not actually do anything other than make it slower
I see this in regular software projects now: any built time improvements will be seized by the programmers adding more dependencies or making the build graph less amenable to parallelization.
November 11, 2026 at 5:45 AM
I see this in regular software projects now: any built time improvements will be seized by the programmers adding more dependencies or making the build graph less amenable to parallelization.
GPUs are designed for heavy parallelization, and like all too many folks in STEM, Jensen tried to make every problem fit his area of expertise.
November 12, 2026 at 7:48 PM
GPUs are designed for heavy parallelization, and like all too many folks in STEM, Jensen tried to make every problem fit his area of expertise.
I mean parallelization is like his whole thing
EXCLUSIVE: Former Nvidia CEO Huang had a relationship with a second employee, sources say reut.rs/4iEjWXZ
EXCLUSIVE: Former Nvidia CEO Huang had a relationship with a second employee, sources say
Ousted Nvidia CEO Jensen Huang had an intimate relationship with a second employee, according to two people familiar with the matter, Reuters has learned.
reut.rs
November 13, 2026 at 1:01 AM
I mean parallelization is like his whole thing
$550 gpu for 2 minutes of work is over 16 grand an hour. if you assume full market penetration and strategic parallelization we're talking an annualized run rate in the low nine figures. under the standard ai lab accounting.
November 9, 2026 at 3:38 AM
$550 gpu for 2 minutes of work is over 16 grand an hour. if you assume full market penetration and strategic parallelization we're talking an annualized run rate in the low nine figures. under the standard ai lab accounting.
Ok, why is it that on my laptop (Kubuntu Linux, Intel i5, 1.6 GHz, 8 GB RAM), a large #glmmTMB model takes ~50s to fit and ~30s with glmmTMBControl(parallel = ...), but the same model on my desktop (Win 11, Intel i7, 3.2 GHz, 32 GB RAM) takes ~100s and parallelization has no effect? #rstats #linux
November 9, 2026 at 10:09 AM
Learn #parallelization with #MPI through hands-on sessions in C, Fortran or Python. No prior MPI knowledge; programming experience required.
Hybrid: TU Wien or online, 23–26 Nov, 09:00–13:30 CET.
Austrian SMEs may get 80% funding via FFG.
Info and registration: events.asc.ac.at/event/406/
Hybrid: TU Wien or online, 23–26 Nov, 09:00–13:30 CET.
Austrian SMEs may get 80% funding via FFG.
Info and registration: events.asc.ac.at/event/406/
November 11, 2026 at 9:47 AM
Learn #parallelization with #MPI through hands-on sessions in C, Fortran or Python. No prior MPI knowledge; programming experience required.
Hybrid: TU Wien or online, 23–26 Nov, 09:00–13:30 CET.
Austrian SMEs may get 80% funding via FFG.
Info and registration: events.asc.ac.at/event/406/
Hybrid: TU Wien or online, 23–26 Nov, 09:00–13:30 CET.
Austrian SMEs may get 80% funding via FFG.
Info and registration: events.asc.ac.at/event/406/
Gangbang Sankey diagram but it's for industrial espionage agents experiencing what you might call "massive parallelization" with a 60-year-old Asian man who wears a motorcycle jacket all the time
November 12, 2026 at 10:39 AM
Gangbang Sankey diagram but it's for industrial espionage agents experiencing what you might call "massive parallelization" with a 60-year-old Asian man who wears a motorcycle jacket all the time
Our Threaded CSV Loader is split into two parts, one is running in parallel to the main loop and keep on loading, which speeds up the operation a lot.
We'll investigate more general parallelization now.
We'll investigate more general parallelization now.
November 11, 2026 at 3:36 PM
Our Threaded CSV Loader is split into two parts, one is running in parallel to the main loop and keep on loading, which speeds up the operation a lot.
We'll investigate more general parallelization now.
We'll investigate more general parallelization now.
LM-SimBench: A Unified Framework, Dataset, and Benchmark for Large Language Model Training Simulation
Yu Dong et al.
#arXiv #cs.AI
Yu Dong et al.
#arXiv #cs.AI
LM-SimBench: A Unified Framework, Dataset, and Benchmark for Large Language Model Training Simulation
Evaluating candidate parallelization strategies through distributed training runs makes strategy selection prohibitively expensive at scale. Performance simulation offers a lower cost alternative, yet controlled comparisons among existing simulators remain difficult because they differ in the train…
arxiv.org
November 11, 2026 at 8:25 PM
3/7
We completely reimplemented the calculations in PyTorch, taking advantage of GPU parallelization.
Depending on protein size & frustration mode, conventional calculations are now ~70–13,000× faster than with FrustratometeR.
& yes, we checked that we get the same results! 👇
We completely reimplemented the calculations in PyTorch, taking advantage of GPU parallelization.
Depending on protein size & frustration mode, conventional calculations are now ~70–13,000× faster than with FrustratometeR.
& yes, we checked that we get the same results! 👇
November 9, 2026 at 3:50 PM
3/7
We completely reimplemented the calculations in PyTorch, taking advantage of GPU parallelization.
Depending on protein size & frustration mode, conventional calculations are now ~70–13,000× faster than with FrustratometeR.
& yes, we checked that we get the same results! 👇
We completely reimplemented the calculations in PyTorch, taking advantage of GPU parallelization.
Depending on protein size & frustration mode, conventional calculations are now ~70–13,000× faster than with FrustratometeR.
& yes, we checked that we get the same results! 👇
6/7
This is where GPU parallelization really shines. ⚡
In our benchmarks, exhaustive mutational scanning reaches a speed-up of roughly 137,000×.
Calculations that were previously extremely expensive can now be performed at proteome-scale!
This is where GPU parallelization really shines. ⚡
In our benchmarks, exhaustive mutational scanning reaches a speed-up of roughly 137,000×.
Calculations that were previously extremely expensive can now be performed at proteome-scale!
November 9, 2026 at 3:50 PM
6/7
This is where GPU parallelization really shines. ⚡
In our benchmarks, exhaustive mutational scanning reaches a speed-up of roughly 137,000×.
Calculations that were previously extremely expensive can now be performed at proteome-scale!
This is where GPU parallelization really shines. ⚡
In our benchmarks, exhaustive mutational scanning reaches a speed-up of roughly 137,000×.
Calculations that were previously extremely expensive can now be performed at proteome-scale!
In K2Rmini, the integration runs three read-filtering tasks 10.7–19.4× faster than a hash-corrected control, including index construction.
It also makes k-mer membership exact, removing false matches from 32-bit hash collisions. Surrounding parallelization helps here too.
It also makes k-mer membership exact, removing false matches from 32-bit hash collisions. Surrounding parallelization helps here too.
November 10, 2026 at 1:45 PM
In K2Rmini, the integration runs three read-filtering tasks 10.7–19.4× faster than a hash-corrected control, including index construction.
It also makes k-mer membership exact, removing false matches from 32-bit hash collisions. Surrounding parallelization helps here too.
It also makes k-mer membership exact, removing false matches from 32-bit hash collisions. Surrounding parallelization helps here too.
3/
📡 From beamformed IQ data, ConfUSIus offers configurable methods to recover B-mode, power Doppler, and tissue and blood-flow velocities.
Or run your own custom processing! Write a function that applies to a single IQ block, and ConfUSIus handles windowing, memory and parallelization.
📡 From beamformed IQ data, ConfUSIus offers configurable methods to recover B-mode, power Doppler, and tissue and blood-flow velocities.
Or run your own custom processing! Write a function that applies to a single IQ block, and ConfUSIus handles windowing, memory and parallelization.
November 3, 2026 at 8:14 PM
3/
📡 From beamformed IQ data, ConfUSIus offers configurable methods to recover B-mode, power Doppler, and tissue and blood-flow velocities.
Or run your own custom processing! Write a function that applies to a single IQ block, and ConfUSIus handles windowing, memory and parallelization.
📡 From beamformed IQ data, ConfUSIus offers configurable methods to recover B-mode, power Doppler, and tissue and blood-flow velocities.
Or run your own custom processing! Write a function that applies to a single IQ block, and ConfUSIus handles windowing, memory and parallelization.
task parallelization parallelization
October 30, 2026 at 10:23 PM
task parallelization parallelization
From yesterday, a look at the forthcoming Alba 9 release including the cool new "host pooling" feature for more effective parallelization of integration tests:
jasperfx.net/news/alba-9....
jasperfx.net/news/alba-9....
Alba 9 Is Nearly Here
Alba 9.0 RC3 is the last release candidate before the biggest release Alba has had in years: host pooling for parallel integration tests, scenarios addressed by route name, JSON path and problem detai...
jasperfx.net
November 3, 2026 at 11:49 AM
From yesterday, a look at the forthcoming Alba 9 release including the cool new "host pooling" feature for more effective parallelization of integration tests:
jasperfx.net/news/alba-9....
jasperfx.net/news/alba-9....
task parallelization parallelization
October 30, 2026 at 10:23 PM
task parallelization parallelization
But, right now, the project takes 5-10 hours to run on the CPU (and that could increase to days if I expand it). So, a little GPU parallelization would be nice.
October 30, 2026 at 3:47 PM
But, right now, the project takes 5-10 hours to run on the CPU (and that could increase to days if I expand it). So, a little GPU parallelization would be nice.
i've heard rumors that sometimes even gpu pipeline scheduling can affect float operation outcomes for some parts of parallelization in e.g. CUDA and so.
and there's deterministic mode _flags_.
which are not necessarily _on_ by default.
so that's cool. that's fun. i just think that's neat.
and there's deterministic mode _flags_.
which are not necessarily _on_ by default.
so that's cool. that's fun. i just think that's neat.
October 29, 2026 at 9:05 PM
i've heard rumors that sometimes even gpu pipeline scheduling can affect float operation outcomes for some parts of parallelization in e.g. CUDA and so.
and there's deterministic mode _flags_.
which are not necessarily _on_ by default.
so that's cool. that's fun. i just think that's neat.
and there's deterministic mode _flags_.
which are not necessarily _on_ by default.
so that's cool. that's fun. i just think that's neat.
Ig if you use a central control node dedicated to running Ansible playbooks it’d be better, but in that case why wouldn’t you just use a pull based configuration engine like puppet that has better parallelization anyway since it inherently splits work between nodes
October 29, 2026 at 1:18 PM
Ig if you use a central control node dedicated to running Ansible playbooks it’d be better, but in that case why wouldn’t you just use a pull based configuration engine like puppet that has better parallelization anyway since it inherently splits work between nodes
read the thread below👇
A ZK pipeline is:
inputs → circuit → witness → prover → proof → verifier
The bottleneck is rarely “cryptography” in the abstract. It is usually circuit size, witness generation, memory footprint, or poor parallelization.
A ZK pipeline is:
inputs → circuit → witness → prover → proof → verifier
The bottleneck is rarely “cryptography” in the abstract. It is usually circuit size, witness generation, memory footprint, or poor parallelization.
October 28, 2026 at 1:28 PM
read the thread below👇
A ZK pipeline is:
inputs → circuit → witness → prover → proof → verifier
The bottleneck is rarely “cryptography” in the abstract. It is usually circuit size, witness generation, memory footprint, or poor parallelization.
A ZK pipeline is:
inputs → circuit → witness → prover → proof → verifier
The bottleneck is rarely “cryptography” in the abstract. It is usually circuit size, witness generation, memory footprint, or poor parallelization.
So, becuase fluid simulation is ABSURDLY performance expensive, but can be fairly easily optimized for GPU parallelization, this ENTIRE SYSTEM needs to live on the GPU as much as possible.
Which means we're working with Buffers.
Which means we're working with Buffers.
October 28, 2026 at 11:40 PM
So, becuase fluid simulation is ABSURDLY performance expensive, but can be fairly easily optimized for GPU parallelization, this ENTIRE SYSTEM needs to live on the GPU as much as possible.
Which means we're working with Buffers.
Which means we're working with Buffers.
Scale people and compute, parallelization, offshoring RL, disdain for IP and privacy constraints, ease of scraping everything), substantialized through huge $
October 26, 2026 at 9:17 PM
Scale people and compute, parallelization, offshoring RL, disdain for IP and privacy constraints, ease of scraping everything), substantialized through huge $