FAQ

The H100 is NVIDIA's latest data center GPU based on the Hopper architecture; the A100 is the previous generation based on Ampere. Key differences:

In short: H100 is 5-6x faster than A100 at roughly 2x the price. For LLM inference, H100 offers better value.

These are two different GPU physical form factors:

• SXM: NVIDIA DGX/A100 Server dedicated slot, interconnecting via NVLink with higher bandwidth and lower multi-GPU latency. Ideal for multi-GPU training/inference.
• PCIe: Standard consumer/server PCIe card, connected via PCIe bus at ~60-70% of SXM bandwidth. Ideal for single-GPU or cost-sensitive workloads.

Choose SXM for multi-GPU interconnect needs; choose PCIe for single-GPU inference at better cost efficiency.

NVLink is NVIDIA's high-speed interconnect technology, allowing GPUs to communicate directly, bypassing the PCIe bus:

• PCIe 4.0 x16: Bidirectional bandwidth ~64 GB/s
• NVLink 4.0 (H100): 900 GB/s per link, with higher aggregate bandwidth in 8-GPU clusters

NVLink-connected H100 (8-GPU cluster) is ideal for large model training with near-linear multi-GPU scaling. Regular PCIe multi-GPU is limited by PCIe bandwidth bottlenecks.

Billing starts when you begin using compute and ends when you return it, measured to the minute.

Example: Using H100 4-GPU, billed at $12.50/hr × 4 = $50/hr. If used for 2 hours 30 minutes, the cost is $50 × 2.5 = $125.

Monthly plans cover 730 hours at ~35% discount compared to hourly rates.

We offer minute-level deployment:

• Base image (PyTorch/TensorFlow): ~3-5 minutes
• Custom containers/model loading: ~10-20 minutes
• Large-scale clusters (8+ GPUs): ~20-30 minutes

99.95% of deployments complete within 30 minutes.

We pre-install popular deep learning frameworks:

• PyTorch 2.x (CUDA 12.x)
• TensorFlow 2.x
• JAX, DeepSpeed, Transformers
• vLLM (LLM inference optimization framework)
• TensorRT, TRITON

For custom images, drivers, or libraries, contact support.

Your data security is our top priority:

• Encrypted communication (TLS 1.3) between all compute nodes
• Training/inference data is automatically deleted after compute completes — no retention
• Optional private node isolation (physical exclusive, not shared with other users)
• Data Processing Agreement (DPA) available upon request

We cannot and will not access your model weights, data, or code.

GPU memory (VRAM) is the key bottleneck for LLMs. Common solutions:

• Quantization: FP16 → INT8 → INT4, trading some accuracy for greater capacity
• Tensor Parallel: Splitting the model across multiple GPUs
• Gradient Accumulation: Simulating large batch size with less memory
• Upgrade to higher VRAM GPU: e.g., from A100 40GB to H100 80GB

If you're unsure how much memory you need, contact support for an estimate.

All prices are transparent with no hidden fees:

All prices in USD, supporting WeChat Pay, Alipay, and credit cards.

Unused balance is eligible for a full refund within 7 days — contact support to process. Consumed compute fees are non-refundable.

Monthly/annual plans are non-refundable after service begins, but you may apply for suspension (up to 30 days).

All payments are eligible for official invoices (VAT regular or special invoices).

Invoices are issued and sent to your email within 3 business days after payment. For special VAT invoices, please provide business qualification documents.

A Token is the smallest unit processed by language models. Chinese is roughly 1 token ≈ 0.5-1 characters, English is roughly 1 token ≈ 0.75 words.

• Input Tokens: Text you send to the model (e.g., prompts, conversation history)
• Output Tokens: Text returned by the model (generated content)

GPT-4o pricing: Input $2.5/1M tokens, Output $10/1M tokens. Output is typically 4x the input price due to heavier computation.

Key differences between domestic and international models:

Recommendation: Use domestic models for simple Chinese tasks (cost savings); use international models for complex reasoning/English tasks (quality first).

Follow this selection logic:

• Simple dialogue/customer service/copywriting → Qwen3 72B (¥0.9/1M, very cheap)
• Complex reasoning/code/math → GPT-4o / Claude 3.5 (quality first)
• Deep Chinese tasks → Qwen3 32B / DeepSeek Coder (best value)
• Image understanding → GPT-4o / Claude Sonnet
• Speech synthesis → OpenAI TTS / Fish TTS

Not sure? Start with the cheapest model to test quality, then upgrade if needed.

Yes, limits vary by model:

• Qwen / DeepSeek: Base plan 100 RPM (requests per minute)
• GPT-4o: 500 RPM, 4096 TPM (tokens per minute)
• Claude 3.5 Sonnet: Ranges from 50-500 RPM depending on plan

For higher concurrency, contact support for enterprise quota.

Your API Key is your account credential, equivalent to a password. Please:

• Never hardcode in source code (use environment variables)
• Never commit to GitHub (use .gitignore)
• Rotate your Key regularly (manual reset available in dashboard)

If your Key is compromised, you can reset it anytime in the dashboard — the old Key is immediately invalidated.

Domestic models use direct connections with latency typically 50-200ms, faster than international models (200-500ms).

International models are routed through optimized paths at ~200-400ms average latency, suitable for most online applications.

Get your API Key after registering. Example call (Python):

import openai
openai.api_key = "fw_xxxxxxxxxx"
openai.base_url = "https://api.flowerwolf.net/v1"

response = openai.ChatCompletion.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Write a quicksort in Python"}]
)
print(response.choices[0].message.content)

Compatible with the OpenAI SDK — just change base_url in existing code.

Yes. FlowerWolf Token Market is fully OpenAI API-compatible:

• Endpoint: /v1/chat/completions
• SDK: Existing OpenAI Python/JS SDK works without modification
• Just replace base_url and api_key

We also support unified access to Anthropic, Alibaba DashScope, and other providers.

In your user dashboard you can view:

• Real-time API call volume (grouped by model)
• Token consumption (input/output separated)
• Cost breakdown and monthly bills
• Remaining balance and top-up history

You can also set up usage alerts for automatic notifications when monthly spend exceeds your threshold.