描述
NVIDIA A100 SXM4是一款基于NVIDIA Ampere架构的旗舰级数据中心GPU，以下是其产品详情：
接口与外形：采用SXM4（Server eXpress Module 4）专用接口，需搭配NVIDIA认证的服务器系统，如DGX A100、HGX A100或OEM厂商定制主板。其外形为模块化板卡设计，直接焊接于服务器主板或通过专用插槽连接，无法独立安装至标准PCIe插槽。
功耗：持续负载功耗为400W，依赖液冷或高密度风冷系统，需冗余电源和定制供电模块。
显存配置：提供40GB和80GB HBM2e显存版本，SXM4 80GB的理论峰值显存带宽为2,039GB/s，HBM2e运行频率为3.2Gbps。
NVLink互联能力：支持第三代NVLink，单卡提供12条NVLink通道，多卡间互联带宽高达600GB/s（双向）。在DGX/HGX系统中，可通过NVSwitch实现8卡全互联，甚至扩展至16卡集群。
计算性能：FP32单精度浮点算力为19.5TFLOPS，TF32张量核心算力为156TFLOPS（基础）/312TFLOPS（启用稀疏计算），INT8整型计算算力为624TOPS。
适用场景：适用于超大规模AI训练，如GPT-4、AlphaFold等千亿参数模型；科学计算与仿真，如气候建模、核聚变模拟等需TB级数据实时处理的HPC应用；高密度数据中心，液冷集群可最大限度提升算力密度。
The NVIDIA A100 SXM4 is a flagship data center GPU based on the NVIDIA Ampere architecture.Product details are as follows:
Interface and Form Factor:It uses the dedicated SXM4(Server eXpress Module 4)interface and requires an NVIDIA-certified server system,such as the DGX A100,HGX A100,or a custom OEM motherboard.It features a modular design,either soldered directly to the server motherboard or connected via a dedicated slot;it cannot be installed independently in a standard PCIe slot.
Power Consumption:Continuous load power consumption is 400W,requiring liquid cooling or high-density air cooling,redundant power supplies,and a custom power supply module.
Memory Configurations:Available in 40GB and 80GB HBM2e memory configurations.The 80GB SXM4 has a theoretical peak memory bandwidth of 2,039GB/s,with HBM2e operating at 3.2Gbps.NVLink Interconnect Capabilities:Supports third-generation NVLink,providing 12 NVLink lanes per card and up to 600GB/s(bidirectional)interconnection bandwidth between multiple cards.In DGX/HGX systems,NVSwitch enables full interconnection of 8 cards,and even scales to 16-card clusters.
Computing Performance:FP32 single-precision floating-point computing power of 19.5 TFLOPS,TF32 Tensor Core computing power of 156 TFLOPS(base)/312 TFLOPS(with sparse compute enabled),and INT8 integer computing power of 624 TOPS.
Applicable Scenarios:Suitable for ultra-large-scale AI training,such as GPT-4 and AlphaFold,with hundreds of billions of parameters;scientific computing and simulation,such as climate modeling and nuclear fusion simulation,and other HPC applications requiring real-time processing of terabytes of data;and liquid-cooled clusters for high-density data centers to maximize computing density.
ABB Model: GBU72 3BHE055094R0002 Model: 3BHE055094R0002 GBU72 Model: GBU72 Model: 3BHE055094R0002 Model: 3BHE031197R0001 Model: 3BHB030310R0001 Model: 73BHE055094R0002 GBU72 Model: 73BHE055094R0002 Model: GBU72 Model: ABB PCS6000 PRODUCT FAMLIY ABB 5SHY4045L0006 3BHB030310R0001 3BHE039203R0101 GVC736CE101
