tinygrad maps GPU BARs in userspace for NVIDIA and AMD GPUs

George Hotz's AI stack still relies on Linux plumbing, but the code shows a serious attempt to move GPU control into a smaller runtime.

By · Published

Primary source: Aligned News - AI Intelligence

Why it matters

tinygrad's BAR-mapping work is an attempt to move more GPU control into an inspectable AI runtime, which could weaken vendor lock-in if it proves reliable beyond developer machines.

Illustration of tinygrad mapping GPU BARs for NVIDIA and AMD GPUs, showing layered software and memory architecture.

George Hotz (@realGeorgeHotz)'s tinygrad says it can map PCIe Base Address Registers directly for NVIDIA and AMD GPUs, a low-level runtime claim that puts the AI framework closer to the hardware than most developers ever get through PyTorch, JAX, CUDA or ROCm.

The claim appeared in a July 12th post on X:

tinygrad on X…

Reader comments

Conversation for this story loads after sign-in.