CPU and GPU working together
The Grace Blackwell platform combines general-purpose processing and AI compute for model inference, development, debugging and running applications.
H3C / Local AI Hardware
MegaCube
Desktop AI Supercomputer
Built on NVIDIA GB10 with unified system memory for local model inference, application development and testing.
Product Capabilities
Key specifications and chassis design
Explore compute, memory and connectivity specifications, with five views of the device and its ports.
- Computing Platform
- NVIDIA GB10
- Grace Blackwell Superchip
20-core Arm CPU - AI Computing Performance
- 1 PFLOPS
- Theoretical peak performance with FP4 sparsity
- Unified System Memory
- 128 GB
- LPDDR5x · 273 GB/s bandwidth
Shared by CPU and GPU - Local Storage
- Up to 4 TB
- NVMe M.2 SSD
1 TB / 2 TB / 4 TB options - High-Speed Interconnect
- 200G
- ConnectX-7 · Dual QSFP ports
RoCE support - Compact Chassis
- 150 × 150 mm
- 50.5 mm high · Approx. 1.2 kg
Compute figures are theoretical peaks. See model validation for measured inference performance. Storage options and supplied configurations are subject to your quotation.
Official product illustrations. Appearance and ports are subject to the delivered device.
Hardware Features
Compute, memory and connectivity working together
Understand a single device’s capabilities before planning model deployment or expansion.
Shared unified system memory
The operating system, model weights, inference caches and applications share memory. Whether a model fits depends on its quantization format and context length.
High-speed network connectivity
High-speed networking supports communication between nodes. Validate the network and inference framework separately for model replicas and for model parallelism across devices.
Selection & Deployment
Three things to confirm before purchasing
Consider your target models, existing software and deployment environment together.
Capacity & Performance
Check reports for your model version, quantization, context and concurrency needs. Theoretical compute performance does not directly translate into application speed.
View model validation →System & Compatibility
MegaCube uses Arm architecture. Check application and container compatibility, and confirm how model services and PINVOU will be installed and connected.
View downloads & connections →Deployment & Accessories
Confirm storage, power, ventilation, networking and device management. For multiple devices, also confirm the interconnect cables, switches and deployment services required.
View solutions & configurations →Find the right MegaCube configuration for you
Tell us your target models and workload requirements so we can assess devices, accessories and the scope of delivery together.