NVIDIA A100
芯片
GPU
上一代
The Ampere workhorse behind GPT-3-era AI — still deployed in huge numbers.
您正在查看一条芯片记录：该处理器是什么、其关键规格、谁设计它、谁制造它、谁供应其零部件，以及它被用在哪里。
The A100 carried the industry through the first scaling wave: GPT-3, PaLM-era research and most 2020–2022 frontier training ran on A100 clusters. It introduced TF32 math, multi-instance GPU partitioning and the 80 GB HBM2e variant that defined memory-bound LLM work before the H100 arrived.
规格与关键参数
每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。
Overview
Designer
NVIDIA
Architecture
Ampere (GA100)
Announced
May 2020 (GTC); 80 GB version November 2020
Status
Prior generation — widely deployed
Manufacturing
Foundry / process
TSMC N7Vendor spec
Transistors
54.2 billionVendor spec
Die size
826 mm²Vendor spec
内存
Memory
40 GB HBM2 or 80 GB HBM2eVendor spec
Memory bandwidth
1,555 (40 GB) / 2,039 (80 GB SXM) GB/sVendor spec
算力
TF32 tensor (dense)
156 TFLOPSVendor spec
NVIDIA datasheets often headline throughput with structured sparsity; figures here are dense unless noted.
BF16/FP16 tensor (dense)
312 TFLOPSVendor spec
INT8 tensor (dense)
624 TOPSVendor spec
FP64 tensor
19.5 TFLOPSVendor spec
电力
TDP
400 (SXM) / 250–300 (PCIe) WVendor spec
Physical
Form factors
SXM4, PCIe
Interconnect
NVLink 3 — 600 GB/s; PCIe 4.0Vendor spec
Partitioning
MIG — up to 7 instancesVendor spec
价格
Launch price
Not publicly disclosed
Market price
~$10,000–15,000 per GPU during the 2021–22 cycleReported
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。 +3 更多分组见下方。
设计方
由……制造
工艺节点
封装
内存技术
用于训练
基础来源
后继产品
前代产品
来源
记录审阅截至 2026-01-15 — 查看 方法论.