XENON and NVIDIA: Certified across the entire stack

Powered by the NVIDIA GB10 Grace Blackwell Superchip, DGX Spark puts a petaflop of AI performance on your desk in a compact, power-efficient form factor.

• Up to 1 petaFLOP (1,000 AI TOPS) of FP4 AI performance from the NVIDIA GB10 Grace Blackwell Superchip
• 128GB of coherent unified CPU-GPU memory to run AI models of up to 200 billion parameters locally, and fine-tune models up to 70 billion parameters
• 20-core Arm CPU (10x Cortex-X925 + 10x Cortex-A725) paired with a Blackwell GPU with fifth-generation Tensor Cores
• NVIDIA ConnectX-7 200Gb/s networking: link two units to work with models of up to 405 billion parameters
• NVIDIA DGX OS with the full NVIDIA AI software stack preinstalled (CUDA, NIM microservices, pretrained models) in a 1.2kg chassis

The same NVIDIA GB10 Grace Blackwell platform as DGX Spark, available now in Australia from XENON.

• 1 petaFLOP (1,000 AI TOPS) of FP4 AI performance from the same NVIDIA GB10 Grace Blackwell Superchip as DGX Spark
• 128GB of coherent unified memory: run models of up to 200 billion parameters on a single unit
• Up to 4TB PCIe Gen5 NVMe storage, the fastest storage option of any GB10 system
• ConnectX-7 200Gb/s SmartNIC plus 10GbE and Wi-Fi 7: cluster two units for models up to 405 billion parameters
• Gigabyte AI TOP Utility for turnkey local fine-tuning and RAG workflows, on top of NVIDIA DGX OS and the full NVIDIA AI stack

A desktop AI supercomputer built on the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, delivering data centre-class training and inference at your desk.

• Up to 20 petaFLOPS of FP4 AI performance from the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
• 748GB of coherent memory: 252GB HBM3e GPU memory at 7.1TB/s plus 496GB LPDDR5X CPU memory
• 72-core Grace CPU (Arm Neoverse V2) fused to the Blackwell Ultra GPU over 900GB/s NVLink-C2C
• NVIDIA ConnectX-8 SuperNIC with up to 800Gb/s networking for multi-node scaling
• Supports up to 7 Multi-Instance GPU (MIG) partitions, so one system serves an entire AI team

A deskside AI supercomputer in a standard pedestal tower, bringing the NVIDIA GB300 Grace Blackwell Ultra Superchip to the desktop.

• NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip: 72-core Grace CPU and Blackwell Ultra GPU in one unified architecture
• 748GB of coherent memory: 252GB HBM3E at 7.1TB/s total bandwidth plus 496GB ECC LPDDR5X at up to 396GB/s
• 2x 400Gb/s QSFP ports via the NVIDIA ConnectX-8 SuperNIC for two-node clustering, plus 10Gb/s LAN
• PCIe Gen5 x16 expansion for additional NVIDIA RTX PRO Blackwell GPUs, plus four M.2 slots including PCIe Gen6
• Closed-loop liquid cooling and a single 1600W 80 PLUS Platinum PSU, Ubuntu certified and ready for a standard office

A turnkey deskside AI workstation built on the NVIDIA DGX Station architecture, engineered for next-generation AI, data science and HPC workloads.

• NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip: 72-core Arm Neoverse V2 Grace CPU plus Blackwell Ultra GPU
• Up to 20 petaFLOPS of FP4 AI performance at the desk
• Up to 748GB of coherent unified memory, with 252GB HBM3e GPU memory at 7.1TB/s
• ConnectX-8 SuperNIC with dual 400G ports (800Gb/s aggregate) to cluster two systems
• Ready to work out of the box: liquid cooling, dual 2TB NVMe in RAID 1, 1600W 80 PLUS Platinum PSU and two spare PCIe 6.0 M.2 slots

Part of the NVIDIA DGX platform, DGX H200 is the proven AI powerhouse at the foundation of NVIDIA DGX SuperPOD and DGX BasePOD, accelerated by the NVIDIA H200 Tensor Core GPU.

• 8x NVIDIA H200 Tensor Core GPUs with 1,128GB of total HBM3e GPU memory
• 32 petaFLOPS of FP8 AI performance
• Fourth-generation NVLink with 4x NVSwitch: 900GB/s of bandwidth per GPU, 7.2TB/s aggregate
• 10x NVIDIA ConnectX-7 network interfaces at up to 400Gb/s for compute and storage fabrics
• Dual Intel Xeon Platinum 8480C processors (112 cores total) with 2TB of system memory

The Blackwell Ultra generation of the DGX platform: purpose-built for the age of AI reasoning, and the new building block of NVIDIA DGX SuperPOD.

• 8x NVIDIA Blackwell Ultra GPUs with 2.1TB of HBM3e GPU memory
• 144 petaFLOPS of FP4 inference and 72 petaFLOPS of FP8 training performance: up to 11x the inference performance of the previous generation
• Fifth-generation NVIDIA NVLink with 14.4TB/s of aggregate GPU-to-GPU bandwidth
• 8x NVIDIA ConnectX-8 ports at up to 800Gb/s each, plus 2x dual-port NVIDIA BlueField-3 DPUs
• The first MGX-rack-compatible DGX: air-cooled and deployable in standard enterprise data centres, shipping now