Choose a currency
Choose you country (VAT)
All others
0%
NVIDIA H100，采用新的Hopper架构，是一款强大的GPU，提供强大的AI加速、大数据处理和高性能计算（HPC）能力。
Hopper架构引入了第四代张量核心，与前代相比，速度提升了九倍，从而在各种机器学习和深度学习任务中提升了性能。
配备80GB高速HBM2e内存的这款GPU可以轻松处理大型语言模型（LLM）或AI任务。
Nvidia A100 80GB GPU卡是一款数据中心加速器，专为加速AI训练和推理以及高性能计算（HPC）应用而设计。
A100 80GB GPU还具有市场上任何GPU中最大的内存容量和最快的内存带宽。这使其非常适合训练和部署最大的AI模型，以及加速需要大数据集的HPC应用。
H100 GPU是有史以来构建的最强大的加速器，其AI训练的性能比上一代快4倍，HPC应用的性能快7倍。
H100的架构针对最大的工作负载进行了超级加速，从大型语言模型到科学计算应用都能应对。它还具有高度可扩展性，支持多达18个NVLink互连，以实现GPU之间的高带宽通信。
H100 GPU专为企业使用设计，具备支持PCIe Gen5、NDR Quantum-2 InfiniBand网络和NVIDIA Magnum IO软件等特性，以实现高效的可扩展性。
这些核心专门设计用于加速AI工作负载，它们提供的FP8精度性能比上一代快2倍。
T这一特性允许A100 GPU跳过矩阵中未使用的部分，这可以使某些AI工作负载的性能提高高达2倍。
这一功能允许A100 GPU被分割成多达七个较小的GPU实例，这些实例可以用来同时加速多个工作负载。
以下是Nvidia A100 80GB和H100 GPU与其他GPU在AI和HPC方面的一些基准测试比较：
人工智能基准 |
||||
| 基准 | A100 80GB | H100 | A40 | V100 |
| ResNet-50推理（图像/秒） | 13 128 | 24 576 | 6 756 | 3 391 |
| BERT大型训练（步数/秒） | 1 123 | 2 231 | 561 | 279 |
| GPT-3培训（代币/秒） | 175b | 400b | 87.5b | 43.75b |
HPC基准 |
||||
| 基准 | A100 80GB | H100 | A40 | V100 |
| HPL DP (TFLOPS) | 40 | 90 | 20 | 10 |
| HPCG (GFLOPS) | 45 | 100 | 22.5 | 11.25 |
| LAMMPS（原子/天） | 115t | 250t | 57.5t | 28.75t |
Nvidia H100 GPU在AI和HPC基准测试中的性能显著超过A100 GPU，并且也比其他GPU更快，如A40和V100。但是，Nvidia A100 GPU的价格更低，对于重要的AI和HPC任务来说可能更具成本效益。
| 用于基于PCIe服务器的H100 | A100 80GB PCIe | |
| FP64 | 26 teraFLOPS | 9.7 TFLOPS |
| FP64张量核心 | 51 teraFLOPS | 19.5 TFLOPS |
| FP32 | 51 teraFLOPS | 19.5 TFLOPS |
| TF32张量核心 | 756 teraFLOPS | 156 TFLOPS | 312 TFLOPS |
| BFLOAT16张量核心 | 1,513 teraFLOPS | 312 TFLOPS | 624 TFLOPS |
| FP16 张量核心 | 1,513 teraFLOPS | 312 TFLOPS | 624 TFLOPS |
| FP8张量核心 | 3,026 teraFLOPS | |
| INT8张量核心 | 3,026 TOPS | 624 TOPS | 1248 TOPS |
| GPU内存 | 80GB | 80GB HBM2e |
| GPU内存带宽 | 2TB/s | 1,935 GB/s |
| 解码器 |
7 NVDEC 7 JPEG |
|
| 最大热设计功率（TDP） | 300-350W (可配置) | 300W |
| 多实例GPU | 最多 7 个 MIGS，每个 10GB | 最多 7 个 MIG，每个 10GB |
| 尺寸规格 | PCIe 双槽风冷 |
PCIe 双槽风冷或单槽液冷 |
| 互联 | NVLink: 600GB/s PCIe Gen5: 128GB/s |
NVIDIA® NVLink® Bridge for 2 GPUs: 600 GB/s PCIe Gen4: 64 GB/s |
| 服务器配置 | Partner and NVIDIA-Certified Systems with 1–8 GPUs | Partner and NVIDIA-Certified Systems™ with 1-8 GPUs |
| NVIDIA AI Enterprise | 包括 |
在订购配备Nvidia Tesla H100或A100显卡的GPU服务器之前，请联系我们的销售部门，以了解相关条款和条件。
TIER III Data Centers
DDoS protection
High-bandwidth Internet connectivity
Full control
Eco-friendly
A replacement server is always available
Quick replacement of components
Round-the-clock technical support
Traffic
Free DDoS protection
Customer support 24/7
我们的服务
GPU servers for data science
e-Commerce hosting
Finance and FinTech
Private cloud
Rendering, 3D Design and visualization
Managed colocation
GPU servers for Deep Learning
Wide range of pre-configured servers with instant delivery and sale
资源
Knowledge base
FAQ
Technical support
If you don't find the right configuration, you can always contact our Sales Department. Our managers will help you with your requirements. We are very flexible.
If you don't find the right configuration, you can always contact our Sales Department. Our managers will help you with your requirements. We are very flexible.
You can choose a suitable Data Center in the Netherlands, Germany, Finland, Iceland, Turkey and the USA
We use an individual approach with each client, which is reflected not only in our technological solutions but also in the appropriate Data Center. We offer Data Centers TIER III categories, which allows us to offer the most flexible solutions for the needs of every client.
For business-critical applications, availability is paramount. In this case, you need a certified Tier III category data center at a minimum. For minor tasks, TIER II or even TIER I Data Center will suffice.
A complete list of the Data Centers and their characteristics can be found here.
If availability is crucial to you, we recommend certified Data Centers, i.e. EuNetworks.
You can use a trial period to test the server. To do this, you need to pay for the server for 1 month. If the server does not meet your needs, you can cancel the service at any time. In this case, the funds, minus the amount used, will be returned to your balance. These funds can be used to pay for other HOSTKEY services. Please note: if you rent a server with software that requires a license purchase, including Windows, such servers are not provided on an hourly payment basis - the minimum rental period is 1 month.
All our services are paid for in advance. We accept payments via credit card, PayPal, P2P cryptocurrency payments from any wallet, application or exchange through BitPay. We also accept WebMoney, Alipay and wire transfers. Read more about our payment terms and methods. Read more about payment terms and methods.
We are very confident in our products and services. We provide fast, reliable and comprehensive service and believe that you will be completely satisfied.
You can ask for a test server for 3-4 days for free.
Refund is only possible in case of an accident from our side with your server being offline for 24 hours or more due to that.
Read more about refund procedure.
Customers whose servers come with unlimited bandwidth are committed to a fair usage policy.
That means that servers on the 1 Gbps port cannot use more than 70% of the allocated bandwidth for more than 3 hours a day.
+31 20 820 3777
支持 
文档 
登录 
