Frequently Asked Questions

Huawei Ascend is a family of AI accelerators (NPU) and servers designed for training and inference of large language models (LLM) and other AI workloads. Ascend is a direct alternative to Nvidia GPU, offering comparable performance at significantly lower cost.

• Independence from US supply chains — no export restrictions
• Lower TCO — 20-40% lower costs over 3 years
• Local support — Czech technical support
• Integration — native support for Chinese LLMs (DeepSeek, Qwen)

Yes. Models trained on Nvidia can be converted to Ascend using ONNX format, CANN toolkit, or MindSpore framework. Most modern LLMs (Llama, DeepSeek, Qwen) are already optimized for Ascend.

Compute Architecture for Neural Networks is Huawei's software stack for Ascend, including:

• ACL — Application Programming Interface
• ATC — Model conversion (Caffe/ONNX/TensorFlow → Ascend)
• Profiling tools — Performance optimization

1. Export model from PyTorch/TensorFlow to ONNX
2. Convert using ATC compiler
3. Optimize for Ascend NPU
4. Test and benchmark

Read our detailed guide: Migrating from CUDA to CANN

• MindSpore — native, best performance
• PyTorch — via CANN backend
• TensorFlow — via CANN backend
• ONNX Runtime — universal inference

Atlas 800 Training Server (8× 910B):

• Price: ~2,500,000–3,000,000 CZK
• Includes 3-year support
• Delivery: 4–6 weeks

Atlas 300I Pro Inference Card:

• Price: ~80,000–120,000 CZK
• PCI-Express card for inference deployment

Yes, we offer:

• Managed hosting — our server in your data center
• Cloud instances — API access
• Lease — 3-year financing options

Yes:

• Remote demo — access to our server
• POC — 30-day testing period
• Benchmark — comparison with your current solution

• Rack space: 2U–4U depending on configuration
• Power: 2× 3000W (for training server)
• Cooling: Front-to-back airflow, 35–45 dB
• Network: 2× 10G/25G/100G Ethernet

Yes, we support:

• Scale-up — up to 8× 910B in single server
• Scale-out — multiple servers via RoCE/InfiniBand
• Kubernetes — container orchestration
• Slurm — HPC workload management

On-premise deployment means:

• Data never leave your infrastructure
• No cloud dependency
• Full control over access
• Audit logs for compliance

Yes:

• Data sovereignty — data stays in EU
• No third-party access
• Right to deletion — full control
• Audit trail — access logging

• ISO 27001 — information security
• ISO 9001 — quality management
• CE — European conformity
• RoHS — environmental standard