“You can also check out the TPU paper[2] which places the ratio at >1000 for some unnamed CNNs.”
In-Datacenter Performance Analysis of a Tensor Processing Unit
谷歌首代 TPU 推理比同期 CPU/GPU 快 15–30 倍
论文评估自 2015 年起部署在数据中心、用于神经网络推理的定制 ASIC:TPU。其核心是由 65,536 个 8 位乘加单元组成的矩阵乘单元,峰值 92 TOPS,配 28 MiB 软件管理的片上内存;确定性的执行模型更契合 99 分位响应时间的要求。在占数据中心 95% 神经网络推理需求的 MLP、CNN 和 LSTM 生产负载上,TPU 平均比同期的 Haswell CPU 和 K80 GPU 快约 15–30 倍,每瓦 TOPS 高 30–80 倍。
评论者把它当作 TPU v1 的权威资料,引用其中的数字(92 TOPS、40 W、15–30 倍、15 个月研制)以及与 CPU 和 K80 的对比等设计细节。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Google's TPU paper [0] showed the CPUs were relatively competitive in the machine learning space (within 2x of a K80).”
“One example is: In-Datacenter Performance Analysis of a Tensor Processing Unit”
“You can find the paper with the methodology, theoretical peaks and latency management at”
“From the paper:”
“For the last version of the TPU, Google provided more detail, e.g., in this paper:”
看全部 10 条 →
“One good approach could be to base the architecture on the TPU v1 from [1].”
“Judging by reading the original TPU architecture paper here - In-Datacenter Performance Analysis of a Tensor Processing Unit”
“(92 TOPS @ 700 Mhz - 40W)”
“The first TPU (Seastar) was designed, tested, and deployed in 15 months: In-Datacenter Performance Analysis of a Tensor Processing Unit”
- In-Datacenter Performance Analysis of a Tensor Processing Unit ↗indescions_2017 · 2017-12-06 · 1 分