kyle yu

this is how i wish i learned GPU fundamentals not a lengthy textbook. not a static image. every concept is an interactive visualization. covering the SM architecture, memory coalescing, synchronization, and more. what concepts do you want to see next? brrrviz.com

@heave448 modal docs r amazing

@brrrkyle Cool UI but I like modal.com/gpu-glossary much more

this is how i wish i learned GPU fundamentals not a lengthy textbook. not a static image. every concept is an interactive visualization. covering the SM architecture, memory coalescing, synchronization, and more. what concepts do you want to see next? brrrviz.com

great, intuitive resource. worth a few mins playing with as a refresher even if you've been through the fundamentals

@sadjikun 🫡

best thing i've seen so far :o

@XandarXam Thanks for the support man!

Check out this incredible project from @brrrkyle. This is for anyone interested in GPU programming or performance engineering. brrrviz.com

@goyal__pramod glad you like it 🤙

HOLY JESUS THIS IS AMAZING

@goyal__pramod check out brrrviz.com for more gpu visuals 🤙

It's a crime that more people have not read these beautiful blogs! Beautiful visuals, simple explanations, code anyone can understand. I have a new bar for my future blogs now...

Chapter 9 of BrrrViz walks you through both scenarios. brrrviz.com

The cost: serialization. Threads queue at the address one at a time. The more threads contend for the same location, the more your parallelism collapses into a bottleneck. This is why real GPU kernels accumulate locally in registers first, then do a single atomicAdd at the end.

Most GPU bugs don't crash your program. They just give you the wrong answer. Silently. When thousands of threads try to update the same memory address simultaneously, each one does three things: 📖 read the current value ⚡ execute their computation ✍ write back the result

Formez vous à l'inference/kernel engineering. Savoir bien optimiser les GPU kernels dans les workloads d'inference vaut de l'or. Maitriser CUDA ou Triton, vLLM, SGLang, TensorRT-LLM est un vrai plus si vous voulez vous démarquer pour 2026-2027 en que AI/ML Engineer.

@ZakShark shameless plug brrrviz.com

@karan_bangia @ZakShark brrrviz.com is the way

@ZakShark any guides on how to start in this domain,

@Banshee2507 @retr0sushi_ hope brrrviz.com can help ya out 🤝

CUDA DAY 1 It’s a lot to take in, but really interesting so far. Notes: past-porpoise-332.notion.site/CUDA-DAY-1-356… Also sorry if this reads rough I used AI to help since writing isn’t my strong suit. #CUDA #GPU #Learning

@Banshee2507 @retr0sushi_ i got u w brrrviz.com

Started learning CUDA today parallel computing feels like a whole new mindset. Any tips, resources, or beginner pitfalls I should know about? #CUDA #GPU #Learning

@retr0sushi_ heyyy brrrviz.com

always a beginner :) ps : if you have resources or roadmaps don't be shy to share them with me pls!

Chasing utilization without this perspective often means optimizing the wrong thing. Understanding where your kernel sits on this diagram helps you execute better optimizations. Find it at chapter 3 of BrrrViz 👉 brrrviz.com

Memory-bound means your hardware is waiting on data. Fix data movement, locality, and reuse. Compute-bound means the data is there, but the math is slow on the hardware. Fix precision, use tensor cores, or change instruction path.

Stop tuning the wrong bottleneck. GPU optimization isn’t one ceiling, it’s memory bandwidth vs peak compute. The roofline plots both, so you see which one limits your kernel.