跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators

AWS Inferentia2AWS Inferentia2

回答「Inferentia2 单加速器的 FP16 算力、HBM 容量与 Inf2 实例上限」。190 TFLOPS 是 FP16;32GB HBM 是单卡容量;12 卡是 Inf2 实例上限,不是单机默认配置。
不主张:不是训练芯片(训练看 Trainium);不是通用 GPU;实际吞吐取决于 Neuron 编译、批处理与模型结构。
Answers the per-accelerator FP16 compute, HBM capacity and the Inf2 instance ceiling. 190 TFLOPS is FP16; 32GB HBM is per accelerator; 12 accelerators is the Inf2 instance maximum, not a default configuration.
It does not claim: this is not a training chip (see Trainium); it is not a general-purpose GPU; real throughput depends on Neuron compilation, batching and model shape.

关键参数Key specs

FP16 算力
190 TFLOPS / 加速器
HBM
32 GB / 加速器
NeuronCore
2 个 / 加速器
实例上限
12 个加速器 / Inf2
内存类型
HBM

官方原文怎么说What the source says

Each Inferentia2 accelerator supports up to 190 tera floating operations per second (TFLOPS) of FP16 performance.
1 https://aws.amazon.com/machine-learning/inferentia/
Inferentia2 offers 32 GB of HBM per accelerator
2 https://aws.amazon.com/machine-learning/inferentia/
Each Inferentia2 accelerator has two second-generation NeuronCores
3 https://aws.amazon.com/machine-learning/inferentia/
with up to 12 Inferentia2 accelerators per EC2 Inf2 instance
4 https://aws.amazon.com/machine-learning/inferentia/

资料与许可Source and licence

资料时间Source checked
2026-11-19
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://aws.amazon.com/machine-learning/inferentia/
类目Category
加速卡Accelerators

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute