TPU v5 tensor core reverse engineering
This project reverse engineers the tensor core on Google's TPU v5 and documents its instruction set. The work stems from the belief that Nvidia GPUs' architecture is a bit subpar for machine learning and that TPUs could be a viable alternative but remain rather undocumented.
What it addresses
Nvidia GPUs' architecture is considered a bit subpar for machine learning, and TPUs could be a viable alternative but are still rather undocumented.
Shared over time
Each entry reflects what was shared on that date. Details may have changed since then.
TPU v5 tensor core reverse engineering
This project reverse engineers the tensor core on Google's TPU v5 and documents its instruction set. The work stems from the belief that Nvidia GPUs' architecture is a bit subpar for machine learning and that TPUs could be a viable alternative but remain rather undocumented.
Problem: Nvidia GPUs' architecture is considered a bit subpar for machine learning, and TPUs could be a viable alternative but are still rather undocumented.
Posted by @gherkinlabs · Original comment ↗
“We've been reverse engineering the tensor core on TPU v5: https://gherkinlabs.com/v5e-tensorcore-instructions”