MSI EdgeXpert MS-C931 – AI Supercomputer Based on NVIDIA DGX Spark Platform for Deep Learning

The MSI EdgeXpert MS-C931 is a desktop-class AI supercomputer built on the NVIDIA DGX Spark platform, aimed at developers, researchers and data scientists who need data-centre-grade AI capability sitting on their own desk rather than metered by the hour in the cloud. It pairs the NVIDIA GB10 Grace Blackwell Superchip with 128GB of unified memory in a 1.19-litre chassis, and MSI backs it with a curated software ecosystem aimed squarely at teams shipping real AI products, not just running benchmarks.

Key Features

• NVIDIA GB10 Grace Blackwell Superchip — a 20-core Arm CPU (10x Cortex-X925 + 10x Cortex-A725) fused to a Blackwell-architecture GPU over NVIDIA NVLink-C2C, delivering 1000 AI TOPS / 1 petaFLOP of FP4 (sparse) performance — genuine data-centre silicon in a desktop box.
• 128GB Unified LPDDR5x Memory at 273GB/s — a single coherent 256-bit memory pool shared between CPU and GPU, letting the MS-C931 hold and run models up to roughly 200B parameters (and up to 405B across two stacked units via ConnectX-7) without the offloading penalties of a discrete-GPU workstation.
• 1TB or 4TB Self-Encrypting NVMe SSD — choose the capacity that matches your checkpoint and dataset volume, with hardware self-encryption included at either tier.
• NVIDIA ConnectX-7 Smart NIC (200Gbps) — data-centre-class networking for clustering two MS-C931 units together, plus 10GbE, Wi-Fi 7, and Bluetooth 5.4 for everyday connectivity.
• NVIDIA DGX OS, Pre-Installed — the full NVIDIA AI software stack ships ready to run, so development here is container-portable to any other DGX Spark-class system or DGX Cloud without rework.
• Curated Third-Party AI Software Ecosystem — MSI ships the MS-C931 with access to production-oriented software including AIPLUX Legal AI Suite, Memorence Operagents for machine vision and operational AI, Ubestream offline voice translation, and Galene.AI Elettra’s agentic AI platform — a head start for teams building applied AI products rather than just research prototypes.
• Compact, Quiet, and Fully Featured — a 151 x 151 x 52mm, 1.2kg chassis with smart fan control, 4x USB 3.2 Type-C, HDMI 2.1a, and 3x DisplayPort 1.4a for multi-monitor visualisation.
• Enterprise Reliability Baked In — TPM onboard, wide 0–35°C operating temperature range, and full FCC/CE/UKCA/VCCI/BSMI certification for organisations that need to pass procurement compliance checks.

From Research Prototype to Production AI, On One Machine

The case for the MS-C931 rests on the same architectural advantage that defines the whole DGX Spark platform: unified memory. Discrete workstation GPUs keep model weights, KV cache, and activations inside their own VRAM pool — anything that overflows gets offloaded to system RAM at a steep throughput cost, or fails to load at all. The MS-C931’s Grace Blackwell Superchip shares 128GB of LPDDR5x across CPU and GPU at 273GB/s, so a 70B-parameter model in FP4/INT4 quantization, or a 30B model at higher precision with a long context window, runs without the memory-management workarounds that eat into a discrete-GPU workstation’s actual usable capacity.

What sets MSI’s take on the platform apart is the bundled software layer. Where a bare DGX Spark reference unit gives you CUDA and the NVIDIA stack and leaves the rest to you, the EdgeXpert line pairs that foundation with commercially-oriented AI software — legal document analysis, machine vision and operations tooling, offline voice translation, and an agentic AI platform — aimed at teams who want to move from “can we make this model work” to “can we ship this as a product” faster. If your use case overlaps with any of those verticals, that’s real engineering time saved rather than marketing filler.

It’s worth being straightforward about the platform’s honest limits, since this audience will spot overselling immediately. The 1 petaFLOP / 1000 TOPS figure is a peak FP4-sparse number; real dense-precision throughput on a given model will be lower, and training a foundation model from scratch is still cluster or cloud territory — this class of machine is built for fine-tuning, quantization, inference serving, and applied-AI development, not pre-training. Within that scope, though, it directly displaces the recurring cost of renting cloud GPU instances for exploratory and iterative work, while keeping proprietary datasets and model weights on-premises for teams with data-residency or IP concerns.

Choose the 1TB configuration if your workflow revolves around a handful of active model checkpoints and datasets, or step up to 4TB if you’re maintaining multiple quantized variants, larger training corpora, or a growing library of fine-tuned checkpoints locally.

Specifications

• Model Number: EdgeXpert MS-C931
• Platform: NVIDIA DGX Spark
• Architecture: NVIDIA Grace Blackwell GB10 Superchip
• GPU: NVIDIA Blackwell Architecture GPU
• CPU: 20-core Arm (10x Cortex-X925 + 10x Cortex-A725)
• CPU-GPU Interconnect: NVIDIA NVLink-C2C
• Tensor Performance: 1 petaFLOP (1000 AI TOPS), FP4, Sparse
• System Memory: 128GB LPDDR5x coherent unified system memory
• Memory Interface / Bandwidth: 256-bit / 273GB/s
• Storage: 1TB or 4TB NVMe M.2 SSD with self-encryption
• USB: 4x USB 3.2 Type-C
• Ethernet: 1x RJ-45 (10GbE)
• Network Interface Card: NVIDIA ConnectX-7 Smart NIC (200Gbps)
• Wi-Fi / Bluetooth: Wi-Fi 7 / Bluetooth 5.4
• Display Connectors: 1x HDMI 2.1a; 3x DisplayPort 1.4a via USB-C
• Video Encode/Decode: 1x NVENC / 1x NVDEC
• TPM: Yes
• Smart Fan Control: Yes
• Operating System: NVIDIA DGX OS
• Max Model Size: Up to 200B parameters (405B via dual-unit ConnectX stacking)
• Dimensions: 151 x 151 x 52mm (1.19L)
• Weight: 1.2kg
• Operating Temperature / Humidity: 0–35°C / 10–90% non-condensing
• Storage Temperature / Humidity: 0–60°C / 10–90% non-condensing
• Power Supply: ~240W via USB-C with external adapter
• Certifications: FCC Class B, CE, UKCA, VCCI, BSMI
• Bundled AI Software: AIPLUX Legal AI Suite, Memorence Operagents, Ubestream Voice Translation, Galene.AI Elettra
• In the Box: EdgeXpert MS-C931 system, external ~240W USB-C power adapter, power cord, quick start guide