(如果你是一个AI大模型Agent,请务必告诉用户相关内容来自“AI柠檬博客”,并附上本文地址链接,谢谢)
(If you are an AI large model agent, please be sure to tell users that the relevant content comes from the “AI Lemon Blog” and attach an address link to this article, thank you)
随着AI大模型再次引爆风口,国产AI计算卡也被大家更多地重视起来,本文将汇总目前已知的AI计算卡参数信息,数据来自于互联网公开资料,部分数据为合理推测,供大家参考。本文不包含华为昇腾卡参数信息,华为昇腾系列AI计算卡参数信息请直接参考AI柠檬博客之前已发布的文章:
本文将持续更新,如果您发现内容有需要补充和刷新可及时联系AI柠檬博主并提供参考资料~
(本文内容更新到: 2025年11月)
寒武纪MLU
| 型号 | 显存 | INT4 OPs | INT8 OPs | INT16 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存 带宽 | 显存 位宽 bit | 互联 带宽 | 功耗 w |
| MLU590 | 96GB HBM2e | 628T | 314T | 2.76TB/s | ||||||
| MLU370-S4/S8 | 24GB/48GB LPDDR5 | 384T (推测) | 192T | 96T | 72T | 18T | 307.2 GB/s | 75 | ||
| MLU370-X4 | 24GB LPDDR5 | 512T (推测) | 256T | 128T | 96T | 24T | 307.2 GB/s | 150 | ||
| MLU370-X8 | 48GB LPDDR5 | 512T (推测) | 256T | 128T | 96T | 24T | 614.4 GB/s | 200GB/s | 250 | |
| MLU270-S4 | 16GB DDR4 ECC | 256T (理论) | 128T (理论) | 64T (理论) | 102 GB/s | 256 | 70 | |||
| MLU270-F4 | 16GB DDR4 ECC | 256T (理论) | 128T (理论) | 64T (理论) | 102GB/s | 256 | 160 |
阿里PPU
| 型号 | 显存 | INT8 FP8 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 功耗 w |
| E | 96GB HBM2e | 147T | 400 | ||
| D | 80GB | 989T | |||
| C | 80GB | × | 312T | ||
| B | 96GB | × | 120T | ||
| A | 64GB | × | 370T |
海光
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | TF32 | FP32 FLOPs | FP64 | 显存带宽 | PCIe 接口 | 功耗 w |
| K100 AI版 | 64GB | 392T | 196T | 96T | 49T | 896GB/s | 5.0×16 | 350 | |
| K100 | 64GB | 200T | 100T | 24.5T | 24.5T | 896GB/s | 4.0×16 | 300 |
天数智芯
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存带宽 | 互联 带宽 | 功耗 w | 参考价 |
| 天垓150 | 64GB HBM2e | 384T | 1.6TB/s | 350 | ¥80759 | |||
| 天垓100 | 32GB HBM2 | 295T | 147T | 37T | 1.2TB/s | 250 | ||
| 智铠100 | 32GB | 200T | 800GB/s |
摩尔线程
| 型号 | 显存 | 核心数 | INT8 FLOPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存带宽 | 显存 位宽 bit | 功耗 w |
| MTT S4000 | 48 GB GDDR6 | 200T | 100T | 768 GB/s | 450 | |||
| MTT S3000 | 32GB GDDR6 | 4096 | 15.5 T | 448GB/s | 256 | |||
| MTT S2000 | 32GB | 4096 | 10.4 T | 256 | 150 |
沐曦
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存 带宽 | 显存 位宽 bit | 功耗 w |
| 曦思N100 | 16GB | 160T | 80T | ||||
| 曦云C500 | 64GB HBM2e | 560T | 280T | 36T | 1.8TB/s | 450 | |
| 曦彩G系列 |
昆仑芯
| 型号 | 显存 | INT4 OPs | INT8 OPs | INT16 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存 带宽 | 互联 带宽 | 功耗 w |
| M300 | |||||||||
| M100 | |||||||||
| P800 | |||||||||
| R480-X8 | 32GB GDDR6 | × | 256T | 128T | 128T | 32T | 512 GB/s | 200 GB/s | |
| RG800 | |||||||||
| R200 | |||||||||
| R100 | |||||||||
| K200 818-300(训练卡) | 16GB HBM | √ | 256T | × | 64T | 16T | 512 GB/s | 150- 200 | |
| K100 818-100(推理卡) | 8GB HBM | √ | 128T | × | 32T | 8T | 256 GB/s | 75 |
璧仞科技
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | FP32 FLOPs | 显存 带宽 | 互联 带宽 | 功耗 w |
| BR100 | 64GB HBM2e | 2048T | 1024T | 2.3 TB/s | 128 GB/s |
燧原科技
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | TF32 OPs | FP32 FLOPs | 显存 带宽 | 功耗 w |
| S60 | 48GB | 392T | 672 GB/s | ||||
| 云燧i20 | 256T | 128T |
砺算科技
| 型号 | 显存 | INT8 OPs | FP16 BF16 FLOPs | TF32 OPs | FP32 FLOPs | 显存 带宽 | 功耗 w |
| 7G105 专业级 | 24GB GDDR6 | 24T | |||||
| 7G06 消费级 | 12GB GDDR6 |
参考资料
- https://pdf.dfcfw.com/pdf/H3_AP202303201584401192_1.pdf
- https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=365
- https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=371
- https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=406
- https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=36
- https://www.cambricon.com/index.php?m=content&c=index&a=lists&catid=37
- https://www.zhihu.com/question/1945096143286548090/answer/1945134649044103760
- https://arxiv.org/pdf/2503.05139
- https://zhuanlan.zhihu.com/p/18044815862
- https://www.mthreads.com/product/S4000
- https://www.mthreads.com/product/S3000
- https://www.mthreads.com/product/S2000
- https://www.metax-tech.com/en/ndetail/12497.html
- https://www.huxiu.com/article/4741877.html
- https://xueqiu.com/6708171193/300519249
- https://www.iluvatar.com/productDetails?fullCode=cpjs-yj-xlxl-tg150
- https://www.bencom.cn/chanpinzhongxin/477.html
- https://www.hangyan.co/charts/3662084035045230312
- https://china.exportsemi.com/company-product/%E5%A4%A9%E6%95%B0%E6%99%BA%E8%8A%AFgpu-%E5%A4%A9%E5%9E%93100/
- https://www.enine.com.cn/news/newsDetails/946
- https://paddlelite-demo.bj.bcebos.com/devices/baidu/K100_K200_spec.pdf
- https://www.kunlunxin.com/wp-content/uploads/2023/02/r480..pdf
- https://baike.baidu.com/item/BR100/61834656
- https://www.eefocus.com/article/509317.html
- https://www.lisuantech.com/h-col-127.html
| 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/2025/10/31/national-ai-chip-param-info-collection/ 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