Is it reasonable that the query speed of the index with refine is faster than that without refine?

[sthgreat]

here is my code:
`

// ivfpq + refine
auto start_1 = std::chrono::high_resolution_clock::now();
std::unique_ptr<faiss::IndexIVFPQ> index_ivfpq;
std::unique_ptr<faiss::IndexFlat> qzr = std::make_unique<faiss::IndexFlat>(vector_dim, faiss::METRIC_L2);
index_ivfpq = std::make_unique<faiss::IndexIVFPQ>(qzr.get(), vector_dim, 256, 64, 8, faiss::METRIC_L2);
index_ivfpq->nprobe = 10;
index_ivfpq->own_fields = true;

auto refine_index = std::make_unique<faiss::IndexRefineFlat>(index_ivfpq.get());
refine_index->k_factor = 10;
refine_index->train(num_vectors, vectors.data());

faiss::IndexIVFPQ* const base_ivfpq_index = dynamic_cast<faiss::IndexIVFPQ*>(refine_index->base_index);
base_ivfpq_index->quantizer = qzr.release();

refine_index->add(num_vectors, vectors.data());
std::unique_ptr<faiss::Index> index_ivfpq_refine;
index_ivfpq_refine = std::move(refine_index);

// ivfpq

auto start_1 = std::chrono::high_resolution_clock::now();
std::unique_ptr<faiss::IndexIVFPQ> index_ivfpq;
std::unique_ptr<faiss::IndexFlat> qzr = std::make_unique<faiss::IndexFlat>(vector_dim, faiss::METRIC_L2);
index_ivfpq = std::make_unique<faiss::IndexIVFPQ>(qzr.get(), vector_dim, 256, 64, 8, faiss::METRIC_L2);
index_ivfpq->nprobe = 10;
index_ivfpq->train(num_vectors, vectors.data());
index_ivfpq->own_fields = true;
index_ivfpq->add(num_vectors, vectors.data()); `

Then, under the same CPU occupy, I ran several groups of query data. The result was that the query speed of the index with refine was faster than that of the index without refine.

`
ivfpq ivfpq_refined

QPS 48354 82267
`

And I am confused. As refine index has additional refine calculation consumptions. Shouldn't it be slower?