Ascend910系列:修订间差异

来自About High Performance
跳到导航 跳到搜索
(Sxzhang移动页面Ascend910Ascend910系列,不留重定向)
无编辑摘要
 
第1行: 第1行:
Ascend 910系列是基于华为自研Davinci架构的NPU处理器,主要用于数据中心训练场景。主要的芯片类型包括[[Ascend910A]]和[[Ascend910B]]


=== Architecture of Ascend 910 Chips ===
=== Architecture of Ascend 910 Chips ===


[[Image:Huawei-Ascend-910-AI-Training-Chip-Block-Diagram.jpg|700px]]
[[Image:Huawei-Ascend-910-AI-Training-Chip-Block-Diagram.jpg|700px]]
Main features:
# The Huawei Ascend 910 has 32 DaVinci AI cores.
# The chip combines these DaVinci AI cores with high-speed HBM2 memory.
# The main chip also has additional logic blocks such as 128 channel video decoding engines.
# There is a Nimbus V3 chip that handles most of the I/O and is co-packaged alongside the main chip and the HBM2 stacks.
# The Huawei Ascend 910 is designed to run at a higher power envelope (350W) and higher performance compared to the Ascend 310.
# Huawei claims 256 TFLOPs of FP16 performance for the Ascend 910.
# The Tesla V100 has 112-125 TFLOPs for deep learning thanks to its tensor cores.

2023年11月9日 (四) 10:20的最新版本

Ascend 910系列是基于华为自研Davinci架构的NPU处理器,主要用于数据中心训练场景。主要的芯片类型包括Ascend910AAscend910B

Architecture of Ascend 910 Chips