Threadwell

Recent Notes

DemetriSpanos
Oct 06, 2026
NOTimothyLottes
Oct 06, 2026
bonzajplc
Oct 06, 2026

agilejebrim

65 items with this tag.

Sep 05, 2026
AgileJebrim
Sep 03, 2026
So if a true “No Graphics API” requires 0 lines of user CPU code, does that make it infinitely better than this? Asking for a friend.
Aug 22, 2026
Start with SPIR-V or PTX instead.
May 30, 2026
I don’t think hardware vendors are going to practically be able to ditch the CPU from system architectures any more than they’ll ditch the GPU, but I do think the GPU should become the master device driving decisions rather than just a slave accelerator for heterogeneous apps.
May 24, 2026
I’ve taken to using a simple interview question with folks.
May 12, 2026
We have code reviews at Isochron and I require my devs to produce streamlined, not modular, code.
Feb 27, 2026
A strange notion that assumes that text is the only way to program computers.
Aug 28, 2025
It’s easy to assume that people reading my tweets are already familiar with the context of the problems I aim to solve, but clearly this post took off with a new audience, so let me try to rewind a bit to focus on why I do what I do.
Feb 09, 2025
Check your compiler.
Dec 18, 2024
@munohikari @SebAaltonen You left out the bit where I also said no graphics pipeline.
Aug 22, 2026
@AgileJebrim Why?
Aug 22, 2026
@AgileJebrim ptx, my ass: https://redplait.blogspot.com/2026/08/parser-of-ptx-instructions.html
Aug 22, 2026
@AgileJebrim Are people seriously starting from ptx 😭 I thought you went there after you got a good grip on cuda
May 31, 2026
@AgileJebrim This Isochron https://www.isochron.org/about ?
May 30, 2026
It's ARM based, so it's not like a GPU in the traditional sense, NVIDIA has made CPU/APUs AFAIK.
May 30, 2026
@AgileJebrim Unfortunately few know how to author programs that could fit in the GPUs tiny instruction caches.
May 30, 2026
I don't think thats a good idea, there's some escalating problems where games focus too much on GPU methods and end up suffering as a result.
May 30, 2026
@AgileJebrim Seems unrealistic to have hard real-time in a general purpose OS.
May 30, 2026
@AgileJebrim Why you guys trying to rebuild the world in the image of the raspberry pi?
May 30, 2026
What problem are you solving? What is the CPU lacking, and what does the GPU offer when the OS is ported to it? And I’m curious to know how you deal with stalls.
May 25, 2026
@AgileJebrim As a curious bystander: can you give a rough idea how suboptimal it is to do "the naive thing"?
May 25, 2026
@AgileJebrim Using a hierarchical/structured layers on-device parallel reduction of cascading (multi-pass) kernels.
May 25, 2026
@AgileJebrim Nvidia has good resources on this https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf
May 25, 2026
@AgileJebrim I would allocate a variable in VRAM, then partition or map the array into a texture (or two).
May 25, 2026
Use OpenMP :^)
May 25, 2026
@AgileJebrim when we were first evaluating "compute shaders" at sony it was my job to find the ideal answer to this very question, and i got it right.
May 25, 2026
@AgileJebrim I have no idea about GPU hardware and how it works, can’t answer
May 24, 2026
@AgileJebrim Lots and lots of atomics.
May 24, 2026
@AgileJebrim Shrink the array over and over until it's down to a single value.
May 24, 2026
@AgileJebrim Add up all the columns separately.
May 24, 2026
@AgileJebrim Large array of values is stored in float/double.
May 24, 2026
@AgileJebrim They asked the same question to me at nvidia, it's a good one!
May 24, 2026
@AgileJebrim I think this is a simple example of "look how much I know about a niche topic, let's find something I know but others don't, so I feel superior".
May 24, 2026
@AgileJebrim M x N multipass reduction
May 24, 2026
@AgileJebrim Then make them build it in Minecraft using pre-2016 redstone.
May 24, 2026
@AgileJebrim many people don't write code that touches the gpu
May 24, 2026
@AgileJebrim If the array is in RAM, do it sequentially.
May 24, 2026
@AgileJebrim This is an idiotic interview question
May 24, 2026
@AgileJebrim I have no meaningful GPU coding experience.
May 24, 2026
@AgileJebrim I would embarrass myself by trying to answer because from the application level a GPU is just a way of using money to make pytorch faster.
May 24, 2026
@AgileJebrim Hierarchical map reduce?
May 24, 2026
I'm bored so I'm gonna give this question a good shot (I learned a bunch of GPU architecture stuff back in uni but I had to look a couple things up to make sure I wasn't mixing terms up).
May 24, 2026
@AgileJebrim If you can fit Qwen 3.6 into your gpu, just load that with ollama and it will write the code for you.
May 24, 2026
@AgileJebrim It's for a frontend JavaScript position I see.
May 24, 2026
@AgileJebrim What kind of GPU?
May 24, 2026
@AgileJebrim No idea about GPU programming, cpu hints towards divide by half, but wouldn’t matrix multiplication works better here? Something like ask gpu to multiply source array by column full of 1 of the same length? Would love if somebody can explain if that would work
May 24, 2026
@AgileJebrim divide into small arrays, conquer each array on a core, then sum the sums?
May 24, 2026
@AgileJebrim I am not knowlegable enough with GPUs, but does numerical stability of the result play a role in questions like these?
May 24, 2026
@AgileJebrim Do floor(warpsize/arraysize) sums, keep using the same accumulated warpsize array until the whole array is done.
May 24, 2026
@AgileJebrim you shouldn’t do that on the GPU.
May 24, 2026
@AgileJebrim Easy.
May 24, 2026
@AgileJebrim Divide up the available memory between cores, each with out output slot and as many input slots as possible.
May 24, 2026
@AgileJebrim Does the array initally exist in GPU memory or CPU?
May 24, 2026
@AgileJebrim efficient reductions are a very deep rabbit hole even on single gpu https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf
May 24, 2026
@AgileJebrim i see it as a convolution task.
May 24, 2026
Use a parallel reduction, not atomics from every thread.
May 24, 2026
@AgileJebrim tree reduction?
May 24, 2026
@AgileJebrim Answering blind: hierarchically.
May 24, 2026
@AgileJebrim They need to know how GPU shaders are organized to even answer this.
May 24, 2026
@AgileJebrim mapreduce.
May 24, 2026
@AgileJebrim I had this problem as an example in a uni class covering cuda, curious how their solution and explanation holds up
May 24, 2026
What precision are the numbers? Integer or floating point? Do we care about numerical stability? Can we assume we're on an NVIDIA GPU with threads, warps, CTAs, and grids? How much memory bandwidth do we have? Is that array of numbers big or really big?
Aug 28, 2025
@valigo (This is assuming you’re still keeping everything as chars and not immediately mapping into their own much smaller indexed representation.)
Dec 18, 2024
@AgileJebrim @SebAaltonen @munohikari (3.) If you had multiple line stores collecting on a given line with ECC, you'd have the same bottleneck as fire-and-forget atomics.
Dec 18, 2024
@AgileJebrim @NOTimothyLottes @munohikari Yes, but it's going to be MUCH slower.

Created with Quartz v5.0.0 © 2026

GitHub