PantheonSim / GPUs / NVIDIA GB10

Simulated NVIDIA GB10: a GPU profile for CUDA and ROCm testing

PantheonSim, the GPU simulator for CUDA and ROCm, simulates the NVIDIA GB10 (Blackwell, 12.1) as the profile nvidia/gb10. The Blackwell GPU in the DGX Spark, which shares its 128 GiB of LPDDR5x memory with the CPU. Its profile is built from public documents and not yet measured against a physical card.

The profile at a glance

These are the numbers a program can ask this simulated GPU, and the values nvidia-smi show for it. They come from the profile file in the simulator's repository, pinned to one commit of it.

Property	Value
Profile id	`nvidia/gb10`
Vendor and architecture	NVIDIA, Blackwell
Compute capability	`12.1`
Device memory	128.0 GiB
Streaming multiprocessors	48
Warp size	32
Max threads per multiprocessor	1536
Shared memory per block (opt-in maximum)	99 KiB
Memory bus width	256-bit
Power limit (profile)	140 W
Maximum clocks (profile)	2617 MHz shader, 4266 MHz memory
Tensor or matrix cores	yes
Verified (measured on a physical card)	not yet measured

What this GPU is for

The Blackwell GPU in the DGX Spark, which shares its 128 GiB of LPDDR5x memory with the CPU.

Blackwell is split across several compute capabilities in these profiles: 10.0 for the B200 and GB200, 10.3 for the B300, 11.0 for Jetson Thor, 12.0 for the GeForce RTX 50 series and the RTX PRO 6000 cards, and 12.1 for the GB10. Only the RTX PRO 6000 Server Edition was read from a physical card.

Among the 37 NVIDIA profiles it ranks 7th by device memory (128.0 GiB) and 33rd by SMs (48), counting the largest as first.

Is this profile verified?

No. This profile is marked not verified: it is built from public documents only, and nothing in it was measured on a physical card. Some figures are derived or copied from the nearest profile where the vendor publishes none, so treat them as approximate, and memory sizes as upper bounds. Programs still run on it, and it still answers every call as the architecture's rules require, but a value such as a clock or a limit may differ from a real card's.

Simulate this GPU

Build the simulator (see the docs) or open the playground, then:

nvcc -cudart shared -arch=compute_75 my_app.cu -o my_app
vgpu run --gpu nvidia/gb10 ./my_app
vgpu shell -y --gpu nvidia/gb10 --count 4
vgpu info --gpu nvidia/gb10

This profile is compute capability 12.1. A kernel built for compute_75 runs on it and on every other NVIDIA profile; a kernel built for a newer architecture is refused by a profile that is older than it. Blackwell targets need CUDA 12.8 or newer. Link the CUDA runtime shared (-cudart shared) so the simulator can stand in for it.

To use it every time, put gpu: nvidia/gb10 in a pantheonsim.yaml settings file, or in a GitHub Actions job give the PantheonSim action gpu: nvidia/gb10 (see CUDA and HIP tests in GitHub Actions). The simulator checks behaviour and has no timing model, so a run on a physical GPU before a release still counts. The CUDA guide explains the rest.

Where the numbers come from

Each profile file begins with a note on its sources. This is the note for NVIDIA GB10, as written in the file:

Source note for nvidia/gb10

Public values; verified: false, nothing here has been measured on a card. Where each value comes from: - compute capability 12.1: NVIDIA's CUDA GPUs table ("NVIDIA GB10 (DGX Spark)", developer.nvidia.com/cuda-gpus). - CUDA cores 6,144: NVIDIA's DGX Spark hardware overview. NVIDIA publishes no SM count for the GB10. 48 is DERIVED: 6144 / 128 CUDA cores an SM, the figure NVIDIA gives every GB20x SM (RTX Blackwell whitepaper, Table 1); unverified. - memory: 128 GB LPDDR5x unified system memory, 256-bit, 4266 MHz, 273 GB/s (DGX Spark hardware overview). The GPU shares it with the 20-core Arm CPU, so this is the whole pool, not what a CUDA program can allocate; an upper bound. NVIDIA's "GB" is binary here as elsewhere, 131072 MiB. - power: the hardware overview gives the GB10 SoC's TDP as 140 W. That is the CPU and GPU together, not the GPU alone. - PCI device 0x2E12, "NVIDIA GB10", from the name table in NVIDIA's open GPU kernel modules (g_nv_name_released.h). - Per-SM limits are compute capability 12.x's, the same as nvidia/rtx5090: the CUDA C++ Programming Guide lists 12.0 and 12.1 in one column (64K registers, 1536 threads, 24 blocks, 100 KB shared memory an SM). - SM clock: not published; the RTX PRO 6000's 2617 MHz (another GB20x) until measured. - Not recorded, because the GB10 is a SoC and NVIDIA states neither: the PCIe link, NVLink, ECC, and a memory type (LPDDR5x is neither "hbm" nor "gddr"). - CUDA attributes (the `cuda:` section): the 256-bit bus is NVIDIA's; the L2 is unpublished and left unset.

The GPU index lists every profile, and this one is not yet measured.

Other Blackwell profiles: