English
The DGX Spark platform is well-documented, with Dell, ASUS, Acer and Gigabyte all adopting NVIDIA’s GB10 Grace Blackwell reference design. These models share identical core specs: 1,000 TOPS FP4 compute, 128GB unified LPDDR5x memory and dual 200GbE networking within a 150mm chassis. Based on the same foundation, HP’s ZGX Nano G1n AI Station features unique customized tweaks that differentiate it from rival Spark devices.
HP ZGX Nano G1n front bezel
HP distinguishes this unit with sustainable materials and optimized structure. Its chassis contains up to 75% recycled aluminum and 20% recycled steel, while packaging reaches 93% recycled content. The internal upper-and-lower partition simplifies access to serviceable parts such as SSDs and batteries. Thermally, the system runs quietly at 22 dBA idle and 27.6 dBA under load, dissipating a peak 780 BTU/hr.
Security stands out as HP’s biggest upgrade over the reference design. The ZGX Nano integrates FIPS 140-2 certified TPM 2.0, Common Criteria EAL4+ certification, plus BIOS-level secure boot and PXE controls. It deploys factory-fitted OPAL self-encrypting NVMe storage. Targeting regulated industries, this compact desk-side AI node emphasizes supply chain validation, static encryption and tamper resistance for enterprise procurement compliance.
Build and Design
Distinct from other DGX Spark devices from Dell, ASUS and Gigabyte, the HP ZGX Nano G1n adopts a split upper-and-lower chassis for simplified internal access to the SSD and battery. Built with sustainable materials, the unit contains 75% recycled aluminum, 20% recycled steel and recycled plastic, while its packaging includes up to 93% corrugated recycled content. Measuring 5.9 × 5.9 × 2.01 inches, this non-rackmountable ultra-compact device serves purely as a desk-side AI node.
Powered by the NVIDIA GB10 Grace Blackwell Superchip, it features soldered 8533 MHz LPDDR5x memory and a reference NVIDIA board, prioritizing density over user upgradability. The forced-air cooling system delivers a 228W peak power draw and 780 BTU/hr heat dissipation, with noise levels rated at 22 dBA idle and 27.6 dBA under load.
Security and Upgradability
The system features robust enterprise security, including a FIPS 140-2 certified TPM 2.0 module and Common Criteria EAL4+ certification. BIOS protections cover secure boot, removable media boot disablement and PXE controls. Fully non-upgradable, its 128GB soldered memory and factory-locked Gen4 OPAL self-encrypting SSDs (2TB/4TB) form an appliance-style design. HP guarantees spare parts availability for five years post-production.
I/O and Connectivity
With a minimal front panel, the rear I/O includes three 20Gbps USB-C ports with DisplayPort Alt Mode and one HDMI 2.1a video output. Networking comprises a 10GbE port and dual 200GbE QSFP112 ports, supporting enterprise features such as PXE boot, VLAN tagging and time synchronization. It also integrates Wi‑Fi 7 and Bluetooth 5.4 for wireless connectivity. HP recommends its official 240W USB‑C power adapter to avoid instability.
Graphics and Thermal Testing
The embedded Blackwell GPU supports up to 8K@60Hz output via USB‑C and 8K@30Hz over HDMI. Thermal testing shows excellent temperature control: CPU peaked at 77.3°C, GPU hit 69°C, NVMe storage reached 54°C, and the NIC topped at 68°C. All components maintained stable temperatures without excessive thermal throttling. GPU power consumption fluctuated between 2.86W and 69W across different workloads, with idle system power staying low at 36–38W.
AI Inference Performance
Tested via the vLLM serving benchmark alongside rival OEM systems, the HP unit ran six models across Equal ISL/OSL, Prefill Heavy and Decode Heavy workloads. It delivered its strongest throughput in Prefill Heavy scenarios. Key results include 2,773 tok/s for GPT‑OSS‑120B, 4,345 tok/s for GPT‑OSS‑20B and 2,774 tok/s for Llama 3.1 8B FP4, demonstrating reliable batch‑scaling capability for mainstream large language models.
GPU Direct Storage (GDS) Benchmarks
GDS eliminates CPU forwarding to enable direct GPU storage access, lowering latency for data-heavy AI tasks. In 16K tests, read throughput peaked at 4.6GiB/s while write throughput stabilized at 3.3GiB/s. Under 1M block sizes, maximum read throughput hit 5.5GiB/s and write throughput maintained 3.5–3.7GiB/s. Latency naturally increased with thread concurrency, yet remained well-controlled compared to peer Spark systems.
Conclusion
The HP ZGX Nano G1n retains standard DGX Spark performance while adding refined engineering upgrades. It runs cooler and quieter than competing models, with steady vLLM scaling and stable GDS throughput. HP’s differentiated advantages include sustainable recycled materials, user-friendly chassis layout and comprehensive security certifications. Tailored for developers and regulated enterprises, this fixed-configuration AI appliance prioritizes security, stability and sustainability. For users without such requirements, all GB10-based Spark platforms deliver comparable core performance, differing mainly in after-sales support and pricing.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!