FCLC (London 🦁)

a white woman with blonde wavy hair wearing a puffy dress is fist bumping a tall red ghostly woman creature

me accepting that my demons are a part of me and that they deserve love too

YouTube video by alt-J alt-J - In Cold Blood (Official Audio)

youtu.be/vlMePO0qywA?...

foreground: bench power supply. it's unremarkable. it's a cheap looking one, the kind you find on aliexpress or ebay or whatever. background: with the blissful expression only stock photo actors are able to pull off, here's a technican-looking guy, starring off into the mid distance. he's holding a laptop, which he's not looking at. the laptop has some panels removed, exposing its ram and whatever. our guy is *holding a medical stethoscope* against the body of the laptop. thankfully he is wearing safety glasses

dear lady who's holding a soldering iron by the tip, and soldering a motherboard with no solder. i have guy for you

One of the biggest things missing from Bsky continues to be Polls.

I use socials in no small part as a way to aggregate and share information. Polls are *easily* the easiest, lowest friction way of doing that.

@jay.bsky.team please please please can that functionality be added?

*preparing for RISCV vector, IME and AME meetings*

"Just another Matrix Mondayyyyyyy"🎶

For users of #OpenFOAM, what is your most used solver?

#hpc #CFD

💥Spack v1.0 is out!💥

This is a huge milestone. We reworked the core to add compiler dependencies, and we're introducing a stable package API.

🚀1.0 also adds concurrent builds, better includes, and much more -- read it all in the release notes!

github.com/spack/spack/...

Hey folks! Will be in the Bay Area from ~25 of July to 2nd of August! If we can meet up, we should!!

Remember when defragmenting a hard drive was a thing that was important to do and you paid for software to do it?

Come see us at #PEARC25!

all constant floats aren't floats James!!!!!
given
float foo = (1/3)*3
and
const float bar = (1/3)*3

foo != bar

!!!!!!!!!!!!!!!!!!!!!!

I have been convinced that Go is mostly an unserious language that should never have been allowed to escape the lab.

Solar power is now California's largest source of electricity. ☀️🔌💡

seems intel has since redacted this section...

Help a journalist out? I'm looking to talk to researchers, engineers, students or entrepreneurs who are considering leaving the U.S. because they lost funding or are worried about visas or anything like that. I'm at mimsical.94 on Signal

CppNorth, The Canadian C++ Conference 2025: Why are software engineers so hard to re... View more about this event at CppNorth, The Canadian C++ Conference 2025

CppNorth 2025: Are Devs Obsolete? 🤔

Join Heather Crawford & John Pavan: "Why are software engineers so hard to replace?" Explore why tech from COBOL to AI, meant to replace us, has only made developers more essential.
🔗 sched.co/21xPC
Tickets: CppNorth.ca
🍁 Join us in Toronto, July 20-23! #cpp #dev

Glenn Posted his ISC recap, and as always, it's excellent. The exert on Mixed precision and Ozaki did have me giggling whilst reading.

I think it's relevant to add some high level context on *why* we vendors decrease FP64; it's not just chasing AI, it's the shear *cost* of FP64.

The short ...

We’ve got a request for information out on where we want to take Livermore Computing and other #HPC centers in the next five years.

hpc.llnl.gov/fg-hpcc-rfi

Check it out and send us your thoughts.

GitHub - tenstorrent/tt-npe: A simple network-on-chip performance estimator (NPE) for Tenstorrent Tensix-based devices A simple network-on-chip performance estimator (NPE) for Tenstorrent Tensix-based devices - tenstorrent/tt-npe

github.com/tenstorrent/...

WAIT

picture of me holding a box containing a Tenstorrent Blackhole accelerator card

picture of Tenstorrent Blackhole accelerator card on my desk

i am going to take @tenstorrent.bsky.social’s beautiful matrix-multiplying machine and make it do bizarre and horrible things

The answer is decoupling the operation from where it’s run.

We can continue to run FP64 DGEMM on vector for those that need it; we’ve been doing it for decades.

If you have a need for “more than fp32”, Ozaki to the rescue!

Part of the power of the scheme is selecting how much precision you need

Depends on training requirements; training still needs some FP32, even if only for accumulation!

ISC'25 recap I had the pleasure of attending the 40th annual ISC High Performance conference this month in Hamburg, Germany. It was a delightful way to t...

Here's my notes from ISC'25 in Hamburg: blog.glennklockwood.com/2025/06/isc2...

#HPC

A nice summary, and a call to action, from the organisers of the PASC'25 Mini-Symposium on Application Perspective on SYCL, a Modern Programming Model for Performance and Portability
youtu.be/4K612eNB6cI
#APEROL #SYCL

Satoshi said this during his talk, but his explanation felt overly simplistic, and I didn't fully buy it. Felix's description makes more sense to me. This quadratic relationship between precision and floorspace is brutal.

#HPC

But what you net out is you're still loosing out on performance/area vs other designs that don't.

Balance in all things is relevant, but I can't put on FP64 (or FP128) just for the sake of it.

i suspect the trend we'll see is FP64 (sometimes FP128) on scalars, FP32-64 on Vectors, =<FP32 matrix

You have techniques for getting some of that back (for example, muxing have the multipliers, giving an extra scalar unit in the middle of your multiplier tree etc. ) that would allow you to fit 2 FP32s in the existing FP64 unit while only costing 10% more.

you can play the same tricks with FP16

... Version is that, doubling the precision of a given FPU quadruples the amount of area needed for that FPU.

That's to say, to fit 1 FP64 unit, I need to give up 4 FP32s. I need to give up 16 FP16/BF16/TF19s

and compared to MXFP8? Aprox factor 64.