NVIDIA DGX Spark GB10 — Product Reference: 940-54242-0006-000

What Can Developers Actually Do With 128GB of Unified Memory on NVIDIA DGX Spark?

NVIDIA DGX Spark GB10 — Product Reference: 940-54242-0006-000

When people first see the specifications for NVIDIA DGX Spark, one number tends to dominate the conversation:

Up to 1 PFLOP of FP4 AI performance.

That's impressive, but for developers, I think another specification may be considerably more interesting:

128 GB of coherent unified memory.

Why?

Because when you're experimenting with large language models locally, raw compute isn't always the first wall you hit.

Very often, it's memory.

Let's look at the architecture behind NVIDIA DGX Spark, what its 128 GB unified-memory system changes for local AI development, and where a machine like this actually fits between a conventional workstation and data-center GPU infrastructure.

---

First: What Is NVIDIA DGX Spark?

DGX Spark is NVIDIA's compact desktop AI development system built around the GB10 Grace Blackwell Superchip.

Instead of combining a conventional x86 CPU with a discrete GPU, GB10 integrates NVIDIA's Grace CPU architecture with a Blackwell GPU architecture in a tightly coupled AI computing platform.

The basic hardware is unusual for something this small:

• NVIDIA GB10 Grace Blackwell Superchip
• Blackwell GPU architecture
• 20-core Arm CPU
• 128 GB LPDDR5x coherent unified memory
• Approximately 273 GB/s memory bandwidth
• Up to 1 PFLOP FP4 AI performance with sparsity
• NVIDIA ConnectX-7 high-speed networking
• 10 Gigabit Ethernet
• NVMe storage
• Wi-Fi 7
• NVIDIA DGX OS

And the entire machine is roughly 150 × 150 × 50.5 mm.

That's small enough to sit on a desk.

But physically fitting on your desk isn't what makes it interesting.