做深度学习离不开算力，就像到达诗和远方的旅途离不开交通工具。本文面向深度学习用户整理市面上常用的各个版本GPU的关键参数，进行汇总整理，方便大家随时查阅。如果数据有误，或有新GPU问世，或者需要补充的参数维度，各位看官可以及时通过联系方式来找我更新数据。
注：仅统计Memory在 16GB以上、发售时间在2017年及以后、算力性能满足深度学习计算要求(Cuda算力性能>=7.0)的GPU。若表中单元格为空，则为暂无数据或数据待补充。博主将不定期移除上市时间最久、已停产过时、市面上已大量退役的GPU数据信息项。
(当前数据整理截止为2025年3月)
Tesla系列
| 版本 | 型号 | Memory | CUDA Core | CUDA性能 | FP8/INT8 T(F)OPs | FP16 半精度TFLOPs | FP32 单精度TFLOPs | FP64 双精度TFLOPs | 总线位宽 | 总线带宽 | 最大功率 | 发布时间 | 原价￥ |
| B系 | B100 | 96 GB *2 HBM3e | 16896*2 | 4096位*2 | 8 TB/s | 700w | |||||||
| B系 | B200 | 180GB HBM3e | 9P | 4.5P | 75 | 37 | 7.7 TB/s | 1000w | 2024 | ||||
| H系 | H100 | 80GB HBM2e/HBM3 | 14592 16896 | 9.0 | 3341 3958 | 756.449 989.43 | 51.2 66.9 | 25.6 33.5 | 5120位 | 2039GB/s | 350w 700w | 2022.03.22 | 26.4w |
| H系 | H200 | 141GB HBM3e | 3341 3958 | 1671 1979 | 60 67 | 30 34 | |||||||
| H系 | H800 | 80GB HBM2e/HBM3 | 18432 | 4P | 2P | 60 | 34 | 2TB/s(HBM2e) 3.9TB/s(HBM3) | 350w 700w | 2023.03 | |||
| H系 | H20 | 96 GB HBM3 | 296 | 148 | 44 | 1 | 4.0 TB/s | ||||||
| L系 | L40 | 48GB GDDR6 | 18176 | 8.9 | 362.066 | 90.516 | 1.414 | 384位 | 864GB/s | 300w | 2022.10.13 | ||
| L系 | L20 | 48GB HBM3 | 10240 | 239 | 119.5 | 59.8 | NA | 384位 | 864 GB/s | ||||
| L系 | L4 | 24GB GDDR6 | 7424 | 8.9 | 121 | 30.3 | 0.49 | 192位 | 300GB/s | 72w | 2023.03.21 | ||
| L系 | L2 | 193 | 75w | ||||||||||
| A系 | A100 | 40GB/80GB HBM2 | 6912 | 8.0 | 312 | 19.5 | 9.7 | 5120位 | 1555GB/s | 250w | 2020.05.14 | ||
| A系 | A800 | 40GB/80GB HBM2 | 6912 | 1248 | 312 | 19.5 | 9.7 | 1.6 TB/s | 400w | 2022.11 | 8.7w | ||
| A系 | A40 | 48GB GDDR6 | 10752 | 8.6 | 149.68 | 37.42 | 1.168 | 384位 | 695.8GB/s | 300w | 2020.10.05 | ||
| A系 | A30 | 24GB HBM2 | 3584 | 8.0 | 165.12 | 10.32 | 5.161 | 3072位 | 933.1GB/s | 165w | 2021.04.12 | ||
| A系 | A10 | 24GB GDDR6 | 9216 | 8.6 | 124.96 | 31.24 | 0.976 | 384位 | 600GB/s | 150w | 2021.04.12 | ||
| A系 | A16 | 4*16GB GDDR6 | 4*1280 | 8.6 | 4*18.432 | 4*4.608 | 1.0848 | 4*128位 | 4*200GB/s | 250w | 2021.04.12 | ||
| A系 | A2 | 16GB GDDR6 | 1280 | 8.6 | 18.124 | 4.531 | 0.14 | 128位 | 200GB/s | 40-60w | 2021.11.10 | ||
| T系 | T4 | 16GB GDDR6 | 2560 | 7.5 | 64.8 | 8.1 | 256位 | 320GB/s | 70w | 2018.09.12 | |||
| V系 | V100 | 16GB/32GB HBM2 | 5120 | 7.0 | 119.192 112.224 105.680 | 14.899 14.028 13.210 | 7.450 7.014 6.605 | 4096位 | 900GB/s 829.44GB/s | 250w | 2017.05.10 |
GeForce系列
| 版本 | 型号 | Memory | CUDA Core | CUDA性能 | FP8/INT8 T(F)OPs | FP16 半精度TFLOPs | FP32 单精度TFLOPs | FP64 双精度TFLOPs | 显存位宽 | 总线带宽 | 最大功率 | 发布时间 | 原价￥ |
| 50系 | RTX 5090 | 32GB GDDR7 | 21760 | 12.8 | 512位 | 575w | $1999 | ||||||
| 50系 | RTX 5090 D | 32 GB GDDR7 | 21760 | 12.8 | 512位 | 1792GB/s | 575w | 16499 | |||||
| 50系 | RTX 5080 | 16GB GDDR7 | 10752 | 12.8 | 256位 | 960GB/s | 360w | 8299 | |||||
| 40系 | RTX 4090 | 24GB GDDR6X | 16384 | 8.9 | 82.58 | 82.58 | 1.290 | 384位 | 1008GB/s | 450w | 2022 10.12 | 12999 | |
| 40系 | RTX 4090D | 24GB | 14592 | 8.9 | 73.54 | 73.54 | 1.149 | 384位 | 1008GB/s | 425w | 2023 12.28 | 11999 | |
| 40系 | RTX 4080 Super | 16GB | 10240 | 8.9 | 51.3 | 51.3 | 0.802 | 256位 | 736GB/s | 320w | 2024 01.31 | 8099 | |
| 40系 | RTX 4080 | 16GB | 9728 | 8.9 | 48.74 | 48.74 | 0.762 | 256位 | 716.8GB/s | 320w | 2022 11.16 | 9499 | |
| 40系 | RTX 4070 Ti Super | 16GB | 8448 | 8.9 | 44.1 | 44.1 | 0.689 | 256位 | 672GB/s | 285w | 2024 01.24 | 6499 | |
| 30系 | RTX 3090 Ti | 24GB | 10752 | 8.6 | 33.54 39.99 | 33.54 39.99 | 0.524 0.625 | 384位 | 1008GB/s | 450w | 2022 03.29 | 14999 | |
| 30系 | RTX 3090 | 24GB | 10496 | 8.6 | 29.38 35.68 | 29.28 35.58 | 0.459 0.558 | 384位 | 935.8GB/s | 350w | 2020 09.02 | 11999 |
Quadro系列
| 型号 | Memory | CUDA Core | CUDA性能 | FP8/INT8 T(F)OPs | FP16 半精度TFLOPs | FP32 单精度TFLOPs | FP64 双精度TFLOPs | 显存位宽 | 总线带宽 | 最大功率 | 发布时间 | 原价￥ |
| RTX 6000 | 48GB GDDR6 | 18176 | 8.9 | 91.1 | 384位 | 960GB/s | 300w | |||||
| RTX 5000 | 32GB GDDR6 | 250w | ||||||||||
| RTX 4500 | 24GB GDDR6 | 210w | ||||||||||
| RTX 4000 | 20GB GDDR6 | 130w | ||||||||||
| RTX 4000 SFF | 20GB GDDR6 | 70w | ||||||||||
| RTX 2000 | 16GB GDDR6 | 70w | ||||||||||
| RTX A6000 | 48GB GDDR6 | 8.6 | 300w | |||||||||
| RTX A5000 | 24GB GDDR6 | 8.6 | 230w | |||||||||
| RTX A4500 | 20GB GDDR6 | 200w | ||||||||||
| RTX A4000 | 16GB GDDR6 | 8.6 | 140w | |||||||||
| Quadro RTX 8000 | 48GB GDDR6 | 7.5 | ||||||||||
| Quadro RTX 6000 | 24GB GDDR6 | 7.5 | ||||||||||
| Quadro RTX 5000 | 16GB GDDR6 | 7.5 | ||||||||||
| Quadro GV100 | 32GB HBM2 | 7.0 |
参考来源
- https://www.nvidia.cn/geforce/graphics-cards/40-series/
- https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
- https://www.nvidia.cn/geforce/graphics-cards/30-series/
- https://www.nvidia.cn/geforce/graphics-cards/compare/
- https://detail.zol.com.cn/1208/1207097/param.shtml
- https://developer.nvidia.com/cuda-gpus
- https://zh.wikipedia.org/wiki/NVIDIA_Tesla
- https://www.bilibili.com/read/cv33922816/
- https://zh.wikipedia.org/wiki/NVIDIA_GeForce_40%E7%B3%BB%E5%88%97
- https://zh.wikipedia.org/wiki/NVIDIA_GeForce_30%E7%B3%BB%E5%88%97
- https://zh.wikipedia.org/wiki/NVIDIA_GeForce_20%E7%B3%BB%E5%88%97
- https://zh.wikipedia.org/wiki/NVIDIA_GeForce_10%E7%B3%BB%E5%88%97
- https://ai.oldpan.me/t/topic/287
- https://en.wikipedia.org/wiki/GeForce_40_series
- https://www.nvidia.com/en-us/design-visualization/rtx-6000/
- https://www.nvidia.com/en-us/design-visualization/desktop-graphics/
- https://resources.nvidia.com/en-us-design-viz-stories-ep/l40-linecard?lx=CCKW39&&search=professional%20graphics
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/NV-DS-Quadro-M6000-24GB-US-NV-fnl-HR.pdf
- https://images.nvidia.com/content/pdf/quadro/data-sheets/192195-DS-NV-Quadro-P5000-US-12Sept-NV-FNL-WEB.pdf
- https://images.nvidia.com/content/pdf/quadro/data-sheets/192152-NV-DS-Quadro-P6000-US-12Sept-NV-FNL-WEB.pdf
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/productspage/quadro/quadro-desktop/quadro-pascal-gp100-data-sheet-us-nv-704562-r1.pdf
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/productspage/quadro/quadro-desktop/quadro-volta-gv100-data-sheet-us-nvidia-704619-r3-web.pdf
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/quadro-rtx-5000-data-sheet-us-nvidia-704120-r4-web.pdf
- https://www.nvidia.cn/design-visualization/rtx-5000/
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/quadro-rtx-6000-us-nvidia-704093-r4-web.pdf
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/quadro-rtx-8000-us-nvidia-946977-r1-web.pdf
- https://viperatech.com/shop/nvidia-hgx-h20/
- https://www.nvidia.com/en-us/data-center/h200/
- https://www.nvidia.com/en-us/data-center/h100/
- https://www.nvidia.cn/geforce/graphics-cards/50-series/rtx-5090-d/
- https://www.nvidia.cn/geforce/graphics-cards/50-series/rtx-5080/
- https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
- https://www.nvidia.com/en-us/data-center/hgx/
- https://blog.csdn.net/Ai17316391579/article/details/132627201
- https://www.jb51.net/hardware/cpu/956950.html
- https://resources.nvidia.com/en-us-blackwell-architecture/datasheet
- https://www.techpowerup.com/gpu-specs/b100.c4275
| 版权声明本博客的文章除特别说明外均为原创，本人版权所有。欢迎转载，转载请注明作者及来源链接，谢谢。本文地址: https://blog.ailemon.net/2024/09/25/nvidia-gpu-params-for-deep-learning/ All articles are under Attribution-NonCommercial-ShareAlike 4.0 |
关注“AI柠檬博客”微信公众号，及时获取你最需要的干货。
WeChat Donate
Alipay Donate
发表回复