当前位置：网站首页>【星光04】2022 深度学习GPU表格
【星光04】2022 深度学习GPU表格
2022-07-15 11:46:00 【杨丝儿】
| 型号 | 显存 | 单精(FP32) | 半精(FP16) | 说明 |
|---|---|---|---|---|
| Tesla P40 | 24GB | 11.76 T | 11.76 T | 比较老的Pascal架构GPU，对于cuda11.x之前且对大显存有需求的算法是非常不错的选择 |
| TITAN Xp | 12GB | 12.15 T | 12.15 T | 比较老的Pascal架构GPU，用作入门比较合适 |
| 1080 Ti | 11GB | 11.34 T | 11.34 T | 和TITANXp同时代的卡，同样适合入门，但是11GB的显存偶尔会比较尴尬 |
| 2080Ti | 11GB | 13.45 T | 53.8 T | 图灵架构GPU，性能还不错，老一代型号中比较适合做混合精度计算的GPU。性价比高 |
| V100 | 16/32GB | 15.7 T | 125 T | 老一代专业计算卡皇，半精性能高适合做混合精度计算 |
| 3060 | 12GB | 12.74 T | 约24T | 如果1080Ti的显存正好尴尬了，3060是不错的选择，适合新手。需要使用cuda11.x |
| A4000 | 16GB | 19.17 T | 约76T | 显存和算力都比较均衡，适合进阶过程使用。需要使用cuda11.x |
| 3080Ti | 12GB | 34.10 T | 约70T | 性能钢炮，如果对显存要求不高则是非常合适的选择。需要使用cuda11.x |
| A5000 | 24GB | 27.77T | 约117T | 性能钢炮，如果觉得3080Ti的显存不够用A5000是合适的选择，并且半精算力高适合混合精度。需要使用cuda11.x |
| 3090 | 24GB | 35.58 T | 约71T | 可以看做3080Ti的扩显存版。性能和显存大小都非常够用，适用性非常强，性价比首选。需要使用cuda11.x |
| A40 | 48GB | 37.42 T | 149.7 T | 可以看做是3090的扩显存版。算力和3090基本持平，因此根据显存大小进行选择。需要使用cuda11.x |
| A100 SXM4 | 40/80GB | 19.5 T | 312 T | 新一代专业计算卡皇，除了贵没缺点。显存大，非常适合做半精计算，因为有NVLink加持，多卡并行加速比非常高。需要使用cuda11.x |
边栏推荐
- 可落地的DDD(7)-战术设计上的一些误区
- Meta announced the launch of make-a-scene: AI image generation can be controlled based on text and sketches
- Kotlin Sealed 是什么？为什么 Google 都用
- 【综合笔试题】难度 2/5，递归运用及前缀和优化
- The mystery of any type conversion
- 基础概念
- 2022-04-18 unity getting started 3 - script Basics
- Yunna computer room dynamic environment monitoring expansion scheme
- 2022-04-20 Unity入门9——其他
- Isolation level of MySQL
猜你喜欢
Valuing space over safety, Highlander and link 09, who do you choose
阿里云物联网平台搭建
What is kotlin sealed? Why does Google use it all
Use iceberg in CDP to pressurize the data Lake warehouse
The mystery of any type conversion
PowerDesigner安装教程
![[benefit activity] stack a buff for your code! Click](/img/20/ab63e3373367ed3cecebd90d3bbb29.jpg)
[benefit activity] stack a buff for your code! Click "tea" to receive the gift
No.1 port and shipping AI enterprises in the world have integrated fly pupil, and the mature and standardized AI products of port and shipping have been applied on a large scale, which has greatly imp
「开源摘星计划」Harbor高可用集群设计及部署（实操+视频），基于离线安装方式
03 按键控制LED
随机推荐
10分钟带你进入Swagger的世界，快来看一看吧
Oracle 添加索引造成其他用户对此表的查询权限丢失
Fragment (III) using fragment in viewpager
Practice of kotlin stateflow search function DB + network
The No.1 port and shipping AI enterprise in the world integrates Feitong, port and shipping AI products are mature, standardized and applied on a large scale, and advanced core technologies greatly im
在 PostgreSQL 15 上使用 postgres_fdw 进行事务的并行提交
pygame快速入门
[benefit activity] stack a buff for your code! Click "tea" to receive the gift
Send your code into space and develop 'the greatest work' with Huawei cloud
XPath Combat: climbing the learning ape land (Part 2)
What is kotlin sealed? Why does Google use it all
Adding an index by Oracle causes other users to lose their query permissions on this table
Vector
2022-04-18 Unity入门2——Unity工作原理
Google recommends using kotlin flow in MVVM architecture
Cache avalanche, cache breakdown, cache penetration
MQTT---SUBSCRIBE和SUBACK
Wechat applet uni app uses uni Getuserprofile to realize wechat authorized login (with source code)
Nacos as the registration and discovery center and configuration center of microservice architecture
To realize the remote direct electromagnetic communication of consciousness, the Dongda team, together with the National University of Singapore and others, is expected to build an electromagnetic bra