Configuration and Network
- GPU
- NVIDIA A100 NVLink 40GB
- CPU
- 2.6GHz AMD EPYC ROME processor, up to 3.3GHz turbo
- Host Memory
- Paired with eight-channel DDR4 memory (GT4 configuration 96 GiB)
- Storage
- Cloud disk type selectable; new elastic cloud disks can be created for mounting and expansion
- Network
- GT4 规格内网带宽 5 Gbps,网络收发包 120万 PPS(GT4.4XLARGE96)
- Inter-GPU Interconnect
- 600GB/s NVLink
- Delivery method
- Cloud server instance (GPU cloud server managed via CVM)
- 节点地区
- Guangzhou, Shanghai, Nanjing, Beijing
Deployment and Delivery Instructions
The Tencent Cloud GPU cloud server Compute-Optimized GT4 series is built on the NVIDIA A100 NVLink 40GB; official documentation lists the available regions as Guangzhou, Shanghai, Nanjing, and Beijing. This listing corresponds to the smallest configuration among them, GT4.4XLARGE96: A100 per card, 16 vCPU, 96 GiB memory, and 5 Gbps intranet bandwidth. GT4 is positioned for general-purpose computing scenarios such as deep learning training, inference, and scientific computing, delivering 9.7 TFLOPS of double-precision floating-point performance with 600GB/s NVLink interconnect. The page provides no pricing; hourly rates and purchase restrictions must be verified in the Price Center or console before renting; the driver guide notes that A100 requires a Tesla driver of version 450 or later.
Activation Process
- 提交用途、配置、卡数与租期
- 与供方确认资源、网络和正式报价
- 签订服务约定并确认交付
- 验收环境、访问权限与服务指标
Editorial Analysis Inferred from source information
This entry is organized around a single instance: the page lists multiple instance series, and GT4 is the only one that simultaneously has a clear instance identifier, the GPU card model, per-card VRAM, and selling regions, so only GT4.4XLARGE96 is taken. Selection inference (based on documentation): the A100 NVLink 40GB comes with 600GB/s NVLink interconnect and 9.7 TFLOPS double-precision performance, and is officially recommended for deep learning training, inference, and scientific computing (computational fluid dynamics, molecular modeling, etc.), suitable for small- to medium-scale training that requires multi-card NVLink scaling; the 40GB of VRAM per card is tight for training large-parameter language models and is better suited to single-card scenarios after training sharding, or to inference. Deployment checks: confirm instance availability and purchase permissions in the target region; install the Tesla 450+ driver and CUDA; storage uses cloud disks and can be expanded by mounting elastic cloud disks; public network bandwidth must be configured separately as needed. Budget checks: the documentation contains no pricing; actual quotes from the Tencent Cloud Price Center must prevail, and monthly or full-machine prices must not be converted into per-card-hour rates.
- The document is a product specification, not an inventory interface; the actual number of purchasable cards is unknown
- The page provides no pricing or minimum rental period; billing and term information is missing
- The GT4 card model is the A100 NVLink 40GB variant; differences from the 80GB version are per the documentation only
- No measured performance data is included; performance judgments are based solely on official specifications and official recommended scenarios
Compilation Notes 内容由 AI 根据公开资料整理;事实与推断分开呈现。
Content Correction