NVIDIA DGX Spark GB10 Grace Blackwell 20-Core Arm 128GB LPDDR5x RAM 4TB SSD AI Workstation PC

Description

NVIDIA DGX Spark GB10 Grace Blackwell AI Workstation for Local Agents in Bangladesh

The NVIDIA DGX Spark GB10 Grace Blackwell AI Workstation is designed to run always-on autonomous agents and large models directly on-premises. Instead of relying on cloud-based token generation resources, this workstation focuses on powerful local AI compute paired with large memory capacity—helping teams keep workloads closer to their data for faster iteration, improved privacy control, and more predictable performance. Built around the NVIDIA GB10 Grace Blackwell Superchip architecture, it is a compact, power-efficient platform intended for continuous agent workloads right from your desktop setup.

Processor & AI Compute Architecture

Processor: 20-Core ARM (10 Cortex-X925 + 10 Cortex-A725). This 20-core ARM configuration supports efficient parallel processing for AI workloads and agent pipelines. The system is powered by the NVIDIA Grace Blackwell GB10 architecture, which is specifically aligned with NVIDIA’s modern AI compute approach. For AI acceleration, it includes NVIDIA Blackwell generation CUDA cores, 4th Generation RT Cores, and 5th Generation Tensor Cores—a combination that targets both high-throughput tensor operations and advanced compute features used in contemporary AI applications.

Tensor Performance: 1 PFLOP. With this level of tensor capability, the workstation is well-suited for tasks such as local LLM experimentation, inference acceleration, agent orchestration, and AI workflows that benefit from strong matrix/tensor throughput. If you’re comparing local deployment versus cloud-only execution, DGX Spark helps reduce dependency on external compute for token generation, which can be especially valuable when you need consistent latency or want to minimize data exposure.

Memory & Storage for Large Local Models

RAM: 128GB LPDDR5x (128GB installed). The LPDDR5x memory type supports high-bandwidth data movement for AI workloads, helping reduce bottlenecks when working with large model contexts, multi-step agent reasoning, and data-heavy pipelines. The workstation also specifies a 256-bit memory interface with up to 273 GB/s memory bandwidth, which is important for sustaining performance during intensive training-like operations, fine-tuning workflows, or high-volume inference.

Storage: 4TB SSD (M.2 NVMe PCIe with self-encryption). With 4TB of NVMe SSD storage, you get ample space for model files, datasets, container images, logs, and agent runtime artifacts. The self-encryption capability adds an extra layer of protection for stored data, which is useful for organizations handling sensitive datasets or regulated information. Since the system has no HDD and no optical drive, it’s optimized for fast boot and rapid access to AI assets—ideal for production-like agent environments.

Graphics, Video Engines & Display Output

Graphics chipset: NVIDIA Grace Blackwell GB10 architecture. The graphics memory is shared/integrated (graphics memory accessibility: integrated; graphics memory: shared). For AI workstation buyers, this matters because the system is primarily focused on AI compute acceleration rather than discrete gaming-style graphics. It still includes dedicated video engine features: NVENC 1x and NVDEC 1x, which are commonly used for hardware-accelerated encoding/decoding in streaming, video analytics, and media-related AI pipelines.

Display connectors: HDMI 2.1a with HDMI port: 1. This makes it straightforward to connect a compatible monitor for local control, monitoring, and development sessions. For best results, pair it with a modern HDMI 2.1a-capable display to ensure stable high-quality output during long-running agent tasks.

Connectivity & External I/O

LAN: 1 x RJ-45 connector with 10GbE. The system also includes a ConnectX-7 NIC @ 200 Gbps, supporting high-speed networking for distributed workflows, fast data transfer, and efficient communication with storage or internal services. For teams running multi-node pipelines or transferring large datasets, this high-throughput networking can significantly reduce waiting time.

WiFi: Wi-Fi 7. Bluetooth: Bluetooth 5.4 with LE. These wireless capabilities help with flexible setup in offices and labs where cabling may be limited, while still supporting modern peripheral connectivity.

External I/O ports: 4 x USB Type-C / Thunderbolt port. This is useful for connecting high-speed storage, development peripherals, and docking solutions. The workstation also provides audio output via HDMI multichannel audio output, which can simplify audio setup for presentations, demos, or monitoring sessions.

Power, Size, Color & Warranty

Power consumption: 240W. A 240W power profile is relatively manageable for an AI workstation class system, supporting continuous operation for always-on agent workloads. Dimensions (WxHxD): 150 x 50.5 x 150mm, and weight: 1.2 kg—making it compact enough for desktop or space-conscious deployments. The workstation color is listed as Golden, giving it a distinctive look for professional workspaces.

Operating system: NVIDIA DGX. Warranty: 3 Year (box mandatory while claiming). This warranty coverage is important for enterprise buyers and institutional deployments where service continuity matters.

Who Should Buy It & Buying Advice

Choose the NVIDIA DGX Spark GB10 Grace Blackwell AI Workstation if you want to run AI agents and large models locally with strong tensor performance, large memory capacity, and fast NVMe storage. It’s especially relevant for organizations in Bangladesh that need predictable performance, reduced reliance on cloud token generation, and a compact workstation footprint for dedicated AI workstations. Before purchasing, confirm your required peripherals and display setup (HDMI 2.1a monitor, USB Type-C/Thunderbolt accessories) and plan for continuous power availability, since the system is built for always-on agent workloads. The price starts from 800,000 (Bangladesh), and may vary based on customization and product availability.