PantheonSim / GPUs / NVIDIA GH200 480GB

Simulated NVIDIA GH200 480GB: a GPU profile for CUDA and ROCm testing

PantheonSim, the GPU simulator for CUDA and ROCm, simulates the NVIDIA GH200 480GB (Hopper, 9.0) as the profile nvidia/gh200-480gb. A Hopper GPU paired with a Grace CPU in the GH200 Superchip. The GPU reaches Grace over NVLink-C2C, which is why its PCIe link in the profile is only x1. Its profile is verified, which means it was read from 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/gh200-480gb`
Vendor and architecture	NVIDIA, Hopper
Compute capability	`9.0`
Device memory	94.5 GiB, HBM
Streaming multiprocessors	132
Warp size	32
Max threads per multiprocessor	2048
Shared memory per block (opt-in maximum)	227 KiB
L2 cache	50 MiB
Memory bus width	6144-bit
Power limit (profile)	900 W
Maximum clocks (profile)	1980 MHz shader, 2619 MHz memory
ECC memory	yes
PCIe link reported	Gen4 x1
Tensor or matrix cores	yes
Verified (measured on a physical card)	yes

What this GPU is for

A Hopper GPU paired with a Grace CPU in the GH200 Superchip. The GPU reaches Grace over NVLink-C2C, which is why its PCIe link in the profile is only x1.

Hopper is compute capability 9.0. The simulator implements Hopper's warpgroup matrix instructions (wgmma), the tensor memory accelerator (TMA), thread-block clusters with their barriers, and distributed shared memory, and checks them against CUTLASS and CuTe's own Hopper unit tests.

Among the 37 NVIDIA profiles it ranks 12th by device memory (94.5 GiB) and 12th by SMs (132), counting the largest as first.

Is this profile verified?

Yes. This profile is marked verified, which means its values were read from a physical NVIDIA GH200 480GB by the simulator's characterization tools, not copied from a data sheet. Verified describes the profile's numbers: it says nothing about speed, because the simulator has no timing model.

Simulate this GPU

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

nvcc -cudart shared -arch=compute_90 my_app.cu -o my_app
vgpu run --gpu nvidia/gh200-480gb ./my_app
vgpu shell -y --gpu nvidia/gh200-480gb --count 4
vgpu info --gpu nvidia/gh200-480gb

This profile is compute capability 9.0. 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. Link the CUDA runtime shared (-cudart shared) so the simulator can stand in for it.

To use it every time, put gpu: nvidia/gh200-480gb in a pantheonsim.yaml settings file, or in a GitHub Actions job give the PantheonSim action gpu: nvidia/gh200-480gb (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 GH200 480GB, as written in the file:

Source note for nvidia/gh200-480gb

Generated by nvidia/tools/characterize.cu from a physical device. Every value below was read from the hardware, not a datasheet.

The GPU index lists every profile, and this one is measured on a physical card.

Other Hopper profiles: