According to IT Home on December 21, Google recently launched a "learning cost model" training dataset called TpuGraphs, which is mainly used to "optimize the compiler" and "improve AI deep learning capabilities". Google pointed out that current AI deep learning systems are usually trained using frameworks such as TensorFlow, JAX, and PyTorch, which mainly optimize the model through the heuristic algorithm of the underlying compiler, and use the "learning cost model" in the relevant compiler to improve the performance of the compiler and improve the deep learning ability of the final output model. Compared with industry competitors, Google's TpuGraphs dataset is 770 times larger in "average graph size" and 25 times larger in "number of graphs" than the industry training set. Google claims that applying the TpuGraphs dataset to the compiler effectively addresses issues such as "scalability", "efficiency", and "quality" of the final output model.