Why does IndexIVFFlat consume significantly more GPU memory than IndexFlat for the same embeddings?

[jiangtann]

embeddings = np.random.randn(5000000, 2048).astype(np.float32)

index = faiss.IndexFlatIP(2048)
options = faiss.GpuMultipleClonerOptions()
gpu_index = faiss.index_cpu_to_all_gpus(index, co=options, ngpu=1)
gpu_index.add(embeddings)

index = faiss.IndexIVFFlat(index, 2048, 32768)
options = faiss.GpuMultipleClonerOptions()
gpu_index = faiss.index_cpu_to_all_gpus(index, co=options, ngpu=1)
gpu_index.train(embeddings[:1000000])
gpu_index.add(embeddings)

Using the code above, IndexFlatIP occupies 40GB, while IndexIVFFlat occupies about 60GB. I'd like to understand why IndexIVFFlat uses so much more GPU memory. Theoretically, it should only be an additional 32768(nlist) * 2048 * 4 bytes, which is about 0.25GB, right?

[jiangtann]

If use np.random.rand instead of np.random.randn, IndexIVFFlat will occupies 51-52GB GPU memory, but it still uses over 10GB more GPU memory than IndexFlatIP.