From Go to the GPU, Integrating with CUDA
summary
Builds a Wordle-playing neural net using Go for orchestration and CUDA for heavy lifting, generating 52,726 synthetic games from a 239-word solution space and ~4,739-word action space. Training runs use the Go MLX library with XLA/CUDA, a 146,596-parameter FP32 model (~4 MB on GPU) and achieves 98/100 on a sealed 100-word test set. Inference runs via handwritten CUDA kernels called from Go through Sego/cgo, one block per action with 128 threads, profiled with Nsight profiler and integrated into an HTTP service.
AI-written summary · about
chapters (7)
- 0:00Go and CUDA integration
- 1:31Wordle game problem
- 12:15Training data with Go concurrency
- 15:31Neural network with Go MLX
- 31:53Model performance and overfitting
- 40:17Go CUDA inference engine
- 53:16Live demo and results
related talks
30mGo Faster: Integrating CUDA in Go for GPU Acceleration
50mDeep Learning with Go
37mA Funny Thing Happened On The Way To Reimplementing AlphaGo
34mLearning about AI/ML for Text, with Wordle!
30mUnderstanding GPUs
34mBuilding a Go AI with Kubernetes and TensorFlow
43mDemocratizing NLP content modeling with transfer learning using GPUs
18mMachine Learning and Go