GB10 is NVIDIA's Grace Blackwell system-on-chip, the single piece of silicon at the heart of a DGX Spark. It combines a Grace Arm-based processor and a Blackwell-class GPU (graphics processing unit) on one package, sharing a single pool of memory between them rather than splitting it into separate processor and graphics memory.
At a glance
What it is
The Grace Blackwell system-on-chip in a DGX Spark
What is on it
A Grace Arm-based processor and a Blackwell-class GPU on one package
Memory model
One unified pool, shared by processor and GPU
Why the name matters
Kernels must be compiled for its exact architecture, not just for Blackwell
Stack
What sits on a GB10 package
Two compute halves, one memory pool. The shared pool is what lets a small box hold a large model, and what makes the processor and GPU draw from the same budget.
3
Unified memory poolone budget both halves draw from
2
Blackwell-class GPUthe parallel compute half that runs the model
1
Grace Arm-based processorthe general-purpose CPU half
What is GB10?
GB10 is the name of the single chip a DGX Spark is built around. NVIDIA calls
its family Grace Blackwell: Grace is the general-purpose processor, an Arm-based
design, and Blackwell is the GPU (graphics processing unit) architecture that
runs the model. On a normal desktop those would be two
separate parts, a CPU (central processing unit) in a socket and a graphics card
in a slot. On GB10 they sit on one package and share one pool of memory.
Why does the name matter to an operator?
Two reasons. First, the shared pool is the whole trick: it is how a box this
small can hold a model this large, because the GPU is not limited to a few
gigabytes of dedicated graphics memory. The processor and the GPU draw from the
same budget instead. Second, GB10 is a specific architecture, not a generic
Blackwell card. GPU code has to be compiled for the exact silicon it runs on,
the way a program built for one operating system will not run on another. A
package that works on an older NVIDIA card can fall back to a slower path, or
fail outright, until the fast kernels are built for this chip. When something
runs slowly on a Spark and flies elsewhere, the chip architecture is the first
suspect worth checking.
Check it yourself
nvidia-smi --query-gpu=name --format=csv,noheader
On a DGX Spark this reports the Grace Blackwell GPU. The single shared pool is why nvidia-smi memory figures move as the rest of the system uses memory.
What GB10 gives you
A processor and a GPU on one package, no separate graphics card
One memory pool, so a large model fits in a small chassis
Blackwell-generation features in a desktop-sized box
What it does not
A drop-in match for off-the-shelf GPU binaries; kernels need its exact architecture
Separate processor and graphics memory, so the two share one budget
Unlimited memory; the shared pool still has a ceiling