Elastic GPU compute
billed per card
Six instance types with T4, P40, V100 and V100S, starting at 4C8G; drivers and CUDA come pre-installed so an inference service runs within 10 minutes. Bill daily, hourly or monthly and release when the training window closes.
What a GPU instance is made of
GPU Cloud Host sits on the same foundation as Global Cloud Host: the same VPC, disks, images, snapshots and API — plus one or more GPUs.
A ZGPU instance = GPU card(s) + vCPU and memory + a local SSD or cloud disk + a VPC subnet and security group + an optional elastic IP. Images with drivers, CUDA and common frameworks pre-installed cut environment setup from hours to minutes.
- Images: Ubuntu 22.04 / 20.04 with the NVIDIA driver, CUDA 11.8 / 12.2, cuDNN, Docker and the NVIDIA Container Toolkit pre-installed.
- Data: keep models and datasets on RSSD / SSD cloud disks so several inference instances share one image and scale fast.
- Billing: daily, hourly or monthly; release when the training window or traffic peak ends.
Four numbers you can verify
Every tile can be checked in nvidia-smi or on your invoice.
Five GPU generations, picked per task
From K80 to V100S: 8.1 to 16.35 TFLOPS FP32 and up to 130 TOPS INT8. T4 for inference, V100 / V100S for training.
Drivers and CUDA pre-installed
Images ship with the NVIDIA driver, CUDA, cuDNN, Docker and the Container Toolkit; start vLLM, Triton or your own container and serve.
GPU 0: Tesla T4 (UUID: GPU-…)
$ docker run --gpus all vllm/vllm-openai … model loaded in 41s
Daily, hourly or monthly
Daily for training windows, hourly for traffic peaks, monthly for always-on inference; billing stops on release, so you never pay for idle GPUs.
Room for large-model inference
A single V100S 32 GB runs FP16 inference for 7B-class models; multi-GPU types are available on request and support distributed training.
Hong Kong, Singapore, Ho Chi Minh City, Denver
GPU stock is concentrated in four regions, others on request; connected privately to cloud hosts, disks and MySQL in the same VPC.
Instance types and GPU specs
Six types across five GPU generations; pick the card by task, then the type by memory and budget.
| Type | GPU | GPU memory | vCPU & memory | System disk | Private network | Typical use | Monthly (USD) |
|---|---|---|---|---|---|---|---|
| G4Inference pick | Tesla T4 × 1 | 16 GB | 4C8G | 20 GB SSD | 10 GbE | Video processing, inference services, graphics rendering | 260 |
| G7-1 | Tesla T4S × 1 | 16 GB | 4C8G | 20 GB SSD | 10 GbE | Elastic inference, graphics workloads, short jobs | 266 |
| G1 | Tesla K80 × 1 | 12 GB | 4C8G | 20 GB SSD | 10 GbE | Entry inference, testing, light graphics | 295 |
| G2 | Tesla P40 × 1 | 24 GB | 4C8G | 20 GB SSD | 10 GbE | Mid-scale inference, parallel computing, rendering | 405 |
| G3 | Tesla V100 × 1 | 16 GB | 4C8G | 20 GB SSD | 10 GbE | Model training, HPC, high-throughput inference | 534 |
| G7-2Training pick | Tesla V100S × 1 | 32 GB | 10C32G | 20 GB SSD | 10 GbE | Distributed training, multi-GPU, heavy compute | 799 |
Multi-GPU types (2 and 4 cards) and newer A- / H-series cards are available on request; contact sales for a quote.
| Spec | Tesla K80 (G1) | Tesla P40 (G2) | Tesla V100 (G3) | Tesla T4 / T4S (G4 / G7-1) | Tesla V100S (G7-2) |
|---|---|---|---|---|---|
| Architecture | Kepler | Pascal | Volta | Turing | Volta |
| CUDA cores | 2,496 | 3,840 | 5,120 | 2,560 | 5,120 |
| FP32 | 8.7 TFLOPS | 12 TFLOPS | 14 TFLOPS | 8.1 TFLOPS | 16.35 TFLOPS |
| INT8 | — | 47 TOPS | — | 130 TOPS | — |
| Tensor performance | — | — | 112 TFLOPS | — | 130 TFLOPS |
| GPU memory | 12 GB | 24 GB | 16 GB | 16 GB | 32 GB |
| Power | 300 W | 250 W | 250 W | 70 W | 300 W |
Driver and CUDA versions in the pre-built images are refreshed quarterly; you can also install from a plain OS image or upload a custom image. The console image list is authoritative for version numbers.
| Image | OS | Driver | CUDA / cuDNN | Pre-installed | Types |
|---|---|---|---|---|---|
| ubuntu-22.04-cuda12.2 | Ubuntu 22.04 LTS | 535.x | 12.2 / 8.9 | Docker, NVIDIA Container Toolkit, Python 3.10 | G3, G4, G7-1, G7-2 |
| ubuntu-22.04-cuda11.8 | Ubuntu 22.04 LTS | 520.x | 11.8 / 8.6 | Docker, NVIDIA Container Toolkit, Python 3.10 | All types |
| ubuntu-20.04-cuda11.4 | Ubuntu 20.04 LTS | 470.x | 11.4 / 8.2 | Docker, Python 3.8 | G1, G2 (older architectures) |
| pytorch-2.x | Ubuntu 22.04 LTS | 535.x | 12.2 / 8.9 | PyTorch 2.x, TorchVision, JupyterLab | G3, G4, G7-1, G7-2 |
| Plain OS image | Ubuntu / CentOS / Debian / Windows | Self-installed | Self-installed | — | All types |
Features
Shares the console, VPC, disks, images, snapshots and API with Global Cloud Host.
Pre-built driver images All types
NVIDIA driver, CUDA, cuDNN, Docker and the Container Toolkit, ready out of the box.
Daily / hourly / monthly All types
Switch between three billing terms; billing stops on release.
GPU monitoring All types
GPU utilization, memory, temperature and power metrics with threshold alerts.
Multi-GPU and distributed On request
2- and 4-card types support NCCL multi-GPU and cross-instance distributed training.
Cloud disks for datasets All types
Attach RSSD / SSD disks for models and datasets; share them across instances via snapshots.
Custom images All types
Freeze a training environment into an image, copy it across regions and create identical nodes in bulk.
VPC and security groups All types
Private connectivity with cloud hosts and MySQL in the same VPC; expose inference through security groups.
OpenAPI and CLI All types
Create and reclaim GPU instances per training job from your own scheduler.
7×24 engineering support All types
Driver, dependency and performance issues answered within 90 seconds, with inference deployment best practices.
Where GPUs are sold
Stock is concentrated in four regions; other regions are provisioned on request, usually within 1 to 3 business days.
| Region | Status | Types in stock | Private connectivity | Recommended for |
|---|---|---|---|---|
| Hong Kong | In stock | G1, G2, G3, G4, G7-1, G7-2 | Cloud hosts, MySQL and disks in the region | Inference for Greater China and Southeast Asia |
| Singapore | In stock | G3, G4, G7-1, G7-2 | Cloud hosts, MySQL and disks in the region | Southeast Asian inference and training |
| Ho Chi Minh City | In stock | G4, G7-1 | Cloud hosts in the region | Local inference and video processing in Vietnam |
| Denver | In stock | G3, G4, G7-2 | Cloud hosts and disks in the region | North American training and rendering |
| Tokyo, Seoul, Frankfurt, Los Angeles and more | On request | Contact sales | Cloud hosts in the region | Regional deployments |
Use cases
Four kinds of GPU workload, each with an architecture, a recommended type and three key points.
Online inference for text, image and support models
- A T4 16 GB runs INT8 / FP16 inference for models under 7B at the lowest cost per card.
- The gateway handles auth, rate limiting and routing; inference instances scale horizontally with traffic.
- Model files live on a cloud disk; new instances start from a snapshot, so scaling never re-downloads the model.
Training, fine-tuning and batch experiments
- V100S 32 GB memory reduces gradient accumulation and gives the highest single-card throughput.
- Datasets sit on three-replica RSSD disks; nodes talk NCCL over the VPC private network.
- Snapshot checkpoints on a schedule and freeze the training environment as an image to rebuild the next experiment in minutes.
Video transcoding, image enhancement and offline rendering
- T4 NVENC hardware encoding for transcoding; P40 24 GB for offline rendering and batch image processing.
- The job queue adds and removes GPU instances with the backlog, billed hourly.
- Tier finished videos and assets to standard disks by access frequency to control long-term storage cost.
Extra compute for training windows and traffic peaks
- When the on-premises cluster is full, rent V100S nodes by the day for the training window instead of buying hardware that idles all year.
- Add T4 inference instances hourly for traffic peaks and release them via the API when the event ends.
- Connect to your own data center over a dedicated line or VPN; data and models move inside the VPC.
Pricing and billing
Six types with monthly list prices, all billable daily and hourly; multi-GPU types and new cards are quoted by sales.
G4 · Tesla T4
Inference services, video processing, graphics rendering
- GPU
- T4 × 1 · 16 GB
- vCPU & memory
- 4C · 8 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 4
G7-1 · Tesla T4S
Elastic inference, graphics workloads
- GPU
- T4S × 1 · 16 GB
- vCPU & memory
- 4C · 8 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 3
G1 · Tesla K80
Entry inference, testing
- GPU
- K80 × 1 · 12 GB
- vCPU & memory
- 4C · 8 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 1
G2 · Tesla P40
Mid-scale inference, parallel computing, rendering
- GPU
- P40 × 1 · 24 GB
- vCPU & memory
- 4C · 8 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 1
G3 · Tesla V100
Model training, high-throughput inference
- GPU
- V100 × 1 · 16 GB
- vCPU & memory
- 4C · 8 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 3
G7-2 · Tesla V100S
Distributed training, multi-GPU
- GPU
- V100S × 1 · 32 GB
- vCPU & memory
- 10C · 32 GB
- System disk
- 20 GB SSD
- Billing
- Monthly / daily / hourly
- In-stock regions
- 3
Billing rules
| Item | Billing | Rules |
|---|---|---|
| GPU instance | Hourly, daily or monthly prepaid | Stopped instances still bill until released; daily ≈ monthly ÷ 25 and hourly ≈ monthly ÷ 720, see the console |
| Multi-GPU types | Linear per card | 2- and 4-card types on request; contact sales for a quote |
| Cloud disk | Per GB · month | Dataset and model disks are bought and released independently and attach to any instance |
| Public bandwidth | Fixed bandwidth per Mbps · month | Fixed bandwidth for inference entries; shared bandwidth packages for fleets |
| Snapshots and images | By storage used | Checkpoint snapshots and environment images bill by size; copies in other regions bill separately |
Unit prices vary by region; the pricing page and create page show live quotes.
Rent GPUs or buy them
Only the dimensions that can be measured.
| Dimension | ZOSOC GPU Cloud Host | Own GPU servers | Major cloud vendors |
|---|---|---|---|
| Upfront cost | From $260 / month, no deposit | Tens of thousands of dollars per server, weeks of lead time | Monthly minimum, some require reserved instances |
| Time to provision | Under 1 minute with drivers pre-installed | Weeks to buy, rack and install drivers | Minutes, drivers usually self-installed |
| Billing granularity | Hourly / daily / monthly | Depreciated yearly | Hourly / monthly |
| Idle cost | Billing stops on release | Depreciation and colocation continue while idle | Billing stops on release |
| Upgrading cards | Rebuild on another type within minutes | Buy again | Rebuild on another type |
| Connectivity to your systems | Same VPC, zero-config private links to cloud hosts and MySQL | Build your own network and lines | Same VPC |
| Support | 7×24, 90-second response, deployment best practices included | Vendor hardware warranty | Tickets, billed by support tier |
| Payment methods | Cards, Alipay, WeChat Pay, Southeast Asian wallets | Purchase contracts | Cards |
Three steps, inference in 10 minutes
Pick a type with a pre-built image, log in and docker run.
Pick a type and a pre-built image
G4 / G7-1 for inference, G3 / G7-2 for training; choose the Ubuntu 22.04 image with CUDA 12.2.
Attach a model disk and set up the network
Add an RSSD / SSD disk for models and datasets; pick a VPC subnet and open the inference port in the security group.
SSH in and verify with nvidia-smi
Driver and CUDA are ready; docker run --gpus all starts vLLM or Triton and serves traffic.
Frequently asked questions
The most common questions on drivers, GPU memory, multi-GPU and billing.
Yes. The pre-built image only saves installation time; uninstall and install any version, or start from a plain OS image, install your own stack and freeze it as a custom image for reuse.
For inference, switch to INT8 / FP16 quantization or move to a P40 24 GB or V100S 32 GB; for training, use gradient accumulation, ZeRO or request a multi-GPU type for model parallelism. Changing type is just a rebuild from the image, and the model disk attaches to the new instance.
2- and 4-card types are available on request with NCCL multi-GPU inside one instance; several instances train distributed over the VPC private network at 10 GbE.
Yes. Use daily or hourly for short jobs and switch to monthly in the console once usage is long-term — it is cheaper; the current period is prorated by actual use.
All current types give you a whole dedicated card with no slicing; for light inference, run several models on a T4 or schedule concurrency at the container layer.
On a separate RSSD or SSD cloud disk rather than the 20 GB system disk; disks move between instances, and a snapshot copies them to new instances so scaling never re-downloads the model.
Instances in the same region and VPC connect by default with under 0.5 ms latency; inference services sit behind a gateway instance over the private network and only the gateway's elastic IP is public.
Contact sales to request stock, usually delivered within 1 to 3 business days; or deploy in the nearest in-stock region and connect to your systems over a dedicated line or the public internet.
A stopped instance still holds the GPU and continues to bill; release the instance when the job ends, after snapshotting the model disk or keeping the data disk.
Console monitoring covers GPU utilization, memory, temperature and power with threshold alerts; you can also export nvidia-smi or DCGM metrics to your own Prometheus.
Your first inference service
running in 10 minutes
Sign up and provision right away; daily or hourly billing lets you release anytime. Not sure which card? Send the model size to an architect.
- 7×24Engineering support around the clock
- 90 sTicket and live-chat response
- 1-on-1Card selection and deployment consulting
- SLA credits99.95% availability commitment