ASUS Ascent GX10 AI Supercomputer (NVIDIA GB10 Superchip)

The ASUS Ascent GX10 packs the NVIDIA GB10 Grace Blackwell Superchip into a stackable, toolless desktop chassis designed for ML engineers and small AI teams who want DGX Spark-class local compute with ASUS’s engineering and finish quality. It’s built to sit on a desk on its own, or be stacked with a second unit when a project outgrows a single machine’s memory ceiling.

Key Features

• NVIDIA GB10 Superchip — an Arm v9.2-A 20-core CPU (10x Cortex-X925 + 10x Cortex-A725) fused to an integrated Blackwell GPU with 5th-gen Tensor Cores and 4th-gen RT Cores over NVLink-C2C, which ASUS states delivers roughly 5x the bandwidth of PCIe Gen 5 for CPU-GPU communication, rated for up to 1 petaFLOP (1000 AI TOPS) of FP4 performance.
• 128GB Unified LPDDR5x Memory at 273GB/s — a single coherent memory pool at 8533MT/s shared between CPU and GPU, letting the GX10 hold and run large language models that would exceed the VRAM of most discrete workstation GPUs.
• Configurable Storage — 1TB, 2TB, or 4TB NVMe — a user-upgradeable M.2 2242 SSD accessible via the base grille, so storage isn’t a one-time decision locked in at purchase.
• Dual-Port NVIDIA ConnectX-7 (200GbE) — data-centre-class networking for clustering multiple GX10 units to pool memory and run inference on larger model classes, plus a dedicated 10GbE RJ-45 port for standard networking.
• Toolless, Stackable Chassis with a Front Power Switch — a genuinely thoughtful usability detail: unlike some GB10-class competitors that bury the power button on the rear panel, the GX10 puts it on the front, which matters when units are stacked or rack-mounted and the back panel isn’t easily reachable.
• Dual-Fan Cooling — active cooling engineered to stay quiet under sustained AI workloads, with ASUS additionally rating the design for improved thermal headroom versus comparably-sized systems.
• NVIDIA DGX OS, Fully Loaded — Ubuntu-based DGX OS with CUDA, PyTorch, and TensorFlow pre-installed, plus support for agentic AI frameworks including OpenClaw and NVIDIA NemoClaw.
• Complete Connectivity — 3x USB-C 3.2 Gen2x2 (20Gbps) carrying DisplayPort 2.1, 1x USB-C with 180W PD 3.1 EPR power delivery, HDMI 2.1, Wi-Fi 7, and Bluetooth 5.4.

Desktop AI Compute Designed to Scale With You

The case for any GB10-based system rests on the same architectural advantage: 128GB of memory addressable coherently by both CPU and GPU at 273GB/s bandwidth, which is what lets a machine this size hold a 70B-parameter model in FP4/INT4 quantization, or a 30B model at higher precision with a long context window, without the offload penalties that plague discrete-GPU workstations. Every implementation of the GB10 platform shares that foundation — what differs between vendors is the industrial design, thermal engineering, and day-to-day usability layered on top, and that’s where ASUS has clearly spent its engineering time.

The toolless chassis and user-accessible M.2 slot mean upgrading storage later doesn’t require a service call, and the dual-fan thermal design is built to keep the GB10’s 140W package running cool and quiet through sustained fine-tuning and inference sessions rather than just short benchmark bursts. For teams planning to scale beyond a single unit, the stackable form factor and dual ConnectX-7 200GbE ports are the whole point — pool two units’ unified memory to move into significantly larger model classes without re-architecting your setup, using the same NVIDIA-certified interconnect approach used across the GB10 ecosystem.

Being straightforward about scope: the 1 petaFLOP / 1000 TOPS figure is a peak FP4-sparse number, and real dense-precision throughput on a given model will be lower — this is a fine-tuning, inference, and prototyping platform, not a from-scratch foundation-model trainer, which is still cluster or cloud territory. Within that scope, though, running the standard NVIDIA DGX OS and CUDA stack means every workload developed here is portable to any other DGX Spark-class system or DGX Cloud without rework, and built-in support for agentic frameworks like OpenClaw and NemoClaw means the GX10 is ready for the autonomous-agent workloads that are increasingly part of applied AI development, not just chat-style inference.

Specifications

• Superchip: NVIDIA GB10 Grace Blackwell
• CPU: Arm v9.2-A, 20-core (10x Cortex-X925 + 10x Cortex-A725)
• GPU: Integrated NVIDIA Blackwell, 5th Gen Tensor Cores, 4th Gen RT Cores
• CPU-GPU Interconnect: NVLink-C2C
• AI Performance: Up to 1 petaFLOP / 1000 AI TOPS (FP4, with sparsity)
• Memory: 128GB LPDDR5x, 8533MT/s, 273GB/s bandwidth
• Storage: 1TB / 2TB (PCIe Gen4) or 4TB (PCIe Gen5) NVMe SSD, user-upgradeable M.2 2242
• Networking: Dual-port NVIDIA ConnectX-7 (200GbE), 10 Gigabit Ethernet (RJ-45)
• Wireless: Wi-Fi 7, Bluetooth 5.4
• Ports: 3x USB-C 3.2 Gen2x2 (20Gbps) with DisplayPort 2.1, 1x USB-C with 180W PD 3.1 EPR, 1x HDMI 2.1
• Operating System: NVIDIA DGX OS (Ubuntu-based), CUDA/PyTorch/TensorFlow pre-installed
• Agentic AI Support: OpenClaw, NVIDIA NemoClaw compatible
• Power Supply: 240W external USB-C adapter
• Dimensions: 150 x 150 x 51mm
• Weight: 1.48kg
• Cooling: Dual-fan active cooling
• Chassis: Toolless, stackable, front-mounted power switch