张量处理单元(TPU)中的Sparsecores对于严重依赖嵌入的AI应用特别有益，例如深度学习推荐模型(DLRMS)。这些模型被广泛用于广告，搜索排名和YouTube等平台。 Sparsecores通过将大型分类空间转换为较小的致密空间来加速嵌入的处理，这对于推荐系统至关重要。例如，TPU V4的Sparsecores使其比TPU V3快3倍，用于推荐模型，并且比基于CPU的系统快5-30倍[1] [3]。
此外，Sparsecores在排名和高级建议工作负载方面是有利的，其中超大嵌入是常见的。在较新的TPU模型(例如Trillium)中，这进一步增强了，该模型将第三代Sparsecores整合以优化这些特定任务的性能[7]。总体而言，任何涉及复杂嵌入或稀疏数据结构的AI应用都可以从TPU中Sparsecores的功能中显着受益。
引用：[1] https://www.kdnuggets.com/2023/04/introducing-tpu-v4-googles-cutting-cutting-dedge-supercomputer-large-large-lange-lange-language-models.html
[2] https://www.wevolver.com/article/tpu-vs-gpu-in-ai-a-ai-a-comprehens-guide-to-their-their-their-their-pher---- impact-simpact-on son-mavinalligence
[3] https://www.vibranium.sg/post/introducing-tpu-v4-google-s-supercomputer-for-large-large-lange-lange-models
[4] https://www.datacamp.com/blog/tpu-vs-gpu-ai
[5] https://cloud.google.com/blog/products/ai-machine-learning/introducing-cloud-tpu-v5p-and-ai-hypercomputer
[6] https://eng.snap.com/training-models-with-tpus
[7] https://futurumgroup.com/insights/the-future-of-ai-infrastructure-unpacking-googles-trillium-tpus/
[8] https://www.zdnet.com/article/5-reasons-why-why-why-googles-trillium-could-transform-transform-ai-and-and-computing-and-computing-and-and-computing-and-and-compacles/
[9] https://cloud.google.com/tpu
[10] https://arxiv.org/pdf/2309.08918.pdf