George Hotz | Programming | coding in UOps on AMD MI350X tickets.comma-con.com/events/comma/1859964
Date of the stream 1 Nov 2025. COMMA_CON 2025 - https://tickets.comma-con.com/events/comma/1859964 from $999 buy https://comma.ai/shop/comma-3x & best ADAS system in the world https://openpilot.comma.ai & - https://tinycorp.myshopify.com Source: - https://github.com/tinygrad/tinygrad - https://docs.tinygrad.org Follow for notifications: - https://twitch.tv/georgehotz Support George: - https://twitch.tv/subs/georgehotz Order tinybox: - https://tinycorp.myshopify.com Chapters: TBD Official George Hotz communication channels: - https://geohot.com - https://twitter.com/realGeorgeHotz - https://instagram.com/georgehotz - https://tinygrad.org - https://geohot.github.io/blog - https://github.com/geohot We archive George Hotz and comma.ai videos for fun. Follow for notifications: - https://twitter.com/geohotarchive Thank you for reading and using the SHOW MORE button. We hope you enjoy watching George's videos as much as we do. See you at the next video.
Video Chapters
- 2:20 Pitting Tinygrad against Torch on a new AMD GPU
- 3:50 Diving deep into the GPU manual for tensor cores
- 5:20 The tensor core mystery: a floating point surprise
- 7:10 Unlocking a free performance boost from AMD
- 8:40 Introducing Tinygrad's new custom kernel syntax
- 10:15 First test: a barebones matrix multiply from scratch
- 11:20 Live debugging the custom matrix multiply
- 13:20 Visualizing the GPU's execution with a custom debugger
- 15:50 Adding a ReLU activation in a single line
- 17:30 How does this new syntax stack up against TVM?
Original Output
2:20 Pitting Tinygrad against Torch on a new AMD GPU 3:50 Diving deep into the GPU manual for tensor cores 5:20 The tensor core mystery: a floating point surprise 7:10 Unlocking a free performance boost from AMD 8:40 Introducing Tinygrad's new custom kernel syntax 10:15 First test: a barebones matrix multiply from scratch 11:20 Live debugging the custom matrix multiply 13:20 Visualizing the GPU's execution with a custom debugger 15:50 Adding a ReLU activation in a single line 17:30 How does this new syntax stack up against TVM? Timestamps by StampBot 🤖 (302-george-hotz-programming-coding-in-uops-on-amd-mi350x-tickets)
Unprocessed Timestamp Content
0:10 Welcome back- my screen tingles and the mic is very low 1:04 The internet is toxic- let's talk about recent tech layoffs 2:20 Comparing Tinygrad performance to Torch on a brand new AMD GPU 3:50 Let's dive into the GPU manual to understand tensor cores 5:20 Why didn't we get a tensor core- oh it's a float 7:10 AMD gives you double the performance on this GPU for free 8:40 Introducing the brand new custom kernel syntax in the framework 10:15 Writing the stupidest possible matrix multiply from scratch to test it 11:20 Our first custom matmul fails live verification- let's debug it 13:20 Visualizing the GPU program graph and disassembly with our debugger 15:50 Adding a ReLU activation to our matmul with one simple line 17:30 How our new syntax compares to other frameworks like TVM Timestamps by StampBot 🤖 (302-george-hotz-programming-coding-in-uops-on-amd-mi350x-tickets)