GPU 天梯
比较不同 NVIDIA GPU 上 AI 工作负载的性能指标
NVIDIA Tesla V100 PCIe 16 GB
NVIDIA Tesla V100 SXM2 16 GB架构
NVIDIA Volta架构
NVIDIA Volta发布日期
2017.06
发布日期
2019.11
显存
16 GB HBM2
显存
16 GB HBM2
FP8 性能
Not supported
FP8 性能
Not supported
FP16 性能
28.26 TFLOPS
FP16 性能
快 9.8%
31.33 TFLOPS
FP32 性能
14.13 TFLOPS
FP32 性能
快 9.8%
15.67 TFLOPS
FP64 性能
7.07 TFLOPS
FP64 性能
快 9.7%
7.83 TFLOPS
CUDA 核心
5120
CUDA 核心
5120
功耗
低 16.7%
250 W
功耗
300 W
GPU 性能可视化
GPU 基准测试对比
GPU 型号 | 发布日期 | 架构 | CUDA 核心 | 显存 | TGP | FP8 | FP16 | FP32 | FP64 |
|---|---|---|---|---|---|---|---|---|---|
| NVIDIA Tesla V100 PCIe 16 GB | 2017.06 | NVIDIA Volta | 5120 | 16 GB HBM2 | 250 W | Not supported | 28.26 TFLOPS | 14.13 TFLOPS | 7.07 TFLOPS |
| NVIDIA Tesla V100 SXM2 16 GB | 2019.11 | NVIDIA Volta | 5120 | 16 GB HBM2 | 300 W | Not supported | 31.33 TFLOPS | 15.67 TFLOPS | 7.83 TFLOPS |
| NVIDIA Tesla V100 PCIe 32 GB | 2018.03 | NVIDIA Volta | 5120 | 32 GB HBM2 | 250 W | Not supported | 28.26 TFLOPS | 14.13 TFLOPS | 7.07 TFLOPS |
| NVIDIA Tesla V100 SXM2 32 GB | 2018.03 | NVIDIA Volta | 5120 | 32 GB HBM2 | 300 W | Not supported | 31.33 TFLOPS | 15.67 TFLOPS | 7.83 TFLOPS |
| NVIDIA GeForce RTX 3090 | 2020.09 | NVIDIA Ampere | 10496 | 24 GB GDDR6X | 350 W | Not supported | 35.58 TFLOPS | 35.58 TFLOPS | 0.56 TFLOPS |
| NVIDIA GeForce RTX 4090 | 2022.10 | NVIDIA Ada Lovelace | 16384 | 24 GB GDDR6X | 450 W | Not supported | 82.58 TFLOPS | 82.58 TFLOPS | 1.29 TFLOPS |
| NVIDIA GeForce RTX 5090 | 2025.01 | NVIDIA Blackwell | 21760 | 32 GB GDDR7 | 575 W | Not supported | 104.8 TFLOPS | 104.8 TFLOPS | 1.64 TFLOPS |
| NVIDIA A40 PCIe | 2020.10 | NVIDIA Ampere | 10752 | 48 GB GDDR6 | 300 W | Not supported | 37.42 TFLOPS | 37.42 TFLOPS | 0.58 TFLOPS |
| NVIDIA RTX A6000 | 2021.03 | NVIDIA Ampere | 10752 | 48 GB GDDR6 | 300 W | Not supported | 38.71 TFLOPS | 38.71 TFLOPS | 0.60 TFLOPS |
| NVIDIA A100 PCIe 40 GB | 2020.06 | NVIDIA Ampere | 6912 | 40 GB HBM2e | 250 W | Not supported | 77.97 TFLOPS | 19.49 TFLOPS | 9.75 TFLOPS |
| NVIDIA A100 SXM4 40 GB | 2020.05 | NVIDIA Ampere | 6912 | 40 GB HBM2e | 400 W | Not supported | 77.97 TFLOPS | 19.49 TFLOPS | 9.75 TFLOPS |
| NVIDIA A100 PCIe 80 GB | 2021.06 | NVIDIA Ampere | 6912 | 80 GB HBM2e | 300 W | Not supported | 77.97 TFLOPS | 19.49 TFLOPS | 9.75 TFLOPS |
| NVIDIA A100 SXM4 80 GB | 2020.11 | NVIDIA Ampere | 6912 | 80 GB HBM2e | 400 W | Not supported | 77.97 TFLOPS | 19.49 TFLOPS | 9.75 TFLOPS |
| NVIDIA L40 | 2022.10 | NVIDIA Ada Lovelace | 14080 | 48 GB GDDR6 | 300 W | Not supported | 90.52 TFLOPS | 90.52 TFLOPS | 1.41 TFLOPS |
| NVIDIA L40s | 2023.08 | NVIDIA Ada Lovelace | 14080 | 48 GB GDDR6 | 350 W | Not supported | 91.61 TFLOPS | 91.61 TFLOPS | 1.43 TFLOPS |
| NVIDIA H100 PCIe 80 GB | 2023.03 | NVIDIA Hopper | 16896 | 80 GB HBM3 | 350 W | 3026 TFLOPS | 204.9 TFLOPS | 51.22 TFLOPS | 25.61 TFLOPS |
| NVIDIA H100 SXM5 80 GB | 2023.03 | NVIDIA Hopper | 16896 | 80 GB HBM3 | 700 W | 3958 TFLOPS | 267.6 TFLOPS | 66.91 TFLOPS | 33.45 TFLOPS |
| NVIDIA H100 PCIe 96 GB | 2023.03 | NVIDIA Hopper | 16896 | 96 GB HBM3 | 350 W | 3026 TFLOPS | 248.3 TFLOPS | 62.08 TFLOPS | 31.04 TFLOPS |
| NVIDIA H100 SXM5 96 GB | 2023.03 | NVIDIA Hopper | 16896 | 96 GB HBM3 | 700 W | 3026 TFLOPS | 248.3 TFLOPS | 66.91 TFLOPS | 31.04 TFLOPS |
| NVIDIA H200 NVL | 2024.11 | NVIDIA Hopper | 16896 | 141 GB HBM3 | 600 W | 3026 TFLOPS | 241.3 TFLOPS | 60.32 TFLOPS | 30.16 TFLOPS |
| NVIDIA H200 SXM 141 GB | 2024.11 | NVIDIA Hopper | 16896 | 141 GB HBM3 | 700 W | 3958 TFLOPS | 267.6 TFLOPS | 66.91 TFLOPS | 33.45 TFLOPS |
GPU 基准测试
最新的 GPU 硬件和软件性能评测，帮助您做出明智的硬件选择
软件性能
DeepSeek-R1-Distill-Qwen-7B
环境: vllm
DeepSeek-R1-Distill-Llama-8B
环境: vllm
DeepSeek-R1-Distill-Qwen-14B
环境: vllm
DeepSeek-R1-Distill-Qwen-32B
环境: vllm
DeepSeek-R1-Distill-Llama-70B
环境: vllm
DeepSeek-R1-Distill-Qwen-7B
环境: sglang
DeepSeek-R1-Distill-Llama-8B
环境: sglang
DeepSeek-R1-Distill-Qwen-14B
环境: sglang
DeepSeek-R1-Distill-Qwen-32B
环境: sglang
DeepSeek-R1-Distill-Llama-70B
环境: sglang