tinygrad

Abstraction: Minimal neural network framework and companion ML hardware boxes

Key points:

• tinygrad decomposes all neural network ops into just three OpTypes: ElementwiseOps, ReduceOps, and MovementOps (CONVs and MATMULs are derived from these)
• Used in production in openpilot to run the driving model on the Snapdragon 845 GPU, replacing Qualcomm SNPE
• Supports full autodiff (forward and backward passes); similar API to PyTorch but simpler; compiles custom kernels per operation for extreme shape specialization with lazy tensor fusion
• tinybox hardware line: red v2 (778 TFLOPS, 4x 9070XT, 64 GB GPU RAM), green v2 Blackwell (3086 TFLOPS, 4x RTX PRO 6000, 384 GB), exabox (~1 EXAFLOP, 720x GPUs, 25,920 GB GPU RAM, preorder 2027)
• tiny corp raised $5M in 2023; goal is to "commoditize the petaflop and enable AI for everyone"

Connections: Tinygrad · Tiny Corp · George Hotz · Deep Learning Frameworks · AI Hardware · Neural Network Optimization

Source: https://tinygrad.org/#tinybox