PantheonSim / GPUs / NVIDIA GPUs

Simulated NVIDIA GPUs: 38 profiles

PantheonSim, the GPU simulator for CUDA and ROCm, carries 38 NVIDIA GPUs profiles. They run from the Turing T4 to Blackwell and Rubin, 14 of them measured on a physical card. Each card has its own page with the numbers a program can ask it, the command to simulate it, and what its verified status means.

How to read the lists

A profile holds what a program can ask a GPU: its compute capability, its multiprocessor count, its per-multiprocessor limits and memory, and the values nvidia-smi shows. The compute capability decides which kernels load. A GPU is named nvidia/<model>, such as nvidia/h100 or nvidia/t4, and vgpu list-gpus prints every id.

The Measured column says whether the profile was read from a physical card by the simulator's characterization tools. A profile that was not is built from public documents only: some figures are derived or copied from a neighbouring part, and memory sizes are upper bounds. Memory is shown in GiB, as the profile reports it, which can differ from the marketed figure. None of this is about speed, because the simulator has no timing model.

Browse by architecture: Rubin, Blackwell, Hopper, Ada Lovelace, Ampere, Turing.

Rubin (1)

Rubin is the generation after Blackwell. NVIDIA has published no PCI device id for it, so the profile carries a placeholder, and the profile is built from public documents only.

Blackwell (10)

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 4500 and 6000 cards, and 12.1 for the GB10. Only the RTX PRO 4500 Server Edition and the RTX PRO 6000 Server Edition were read from physical cards.

Hopper (6)

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.

Ada Lovelace (6)

Ada Lovelace is compute capability 8.9. The L4 and L40S profiles were read from physical cards on AWS g6 and g6e instances, and the other Ada profiles are built from public documents.

Ampere (13)

Ampere covers two compute capabilities in these profiles: 8.0 for the GA100 data-center parts and 8.6 for the GA10x parts. The simulator runs the SASS a binary carries for the architecture and falls back to PTX when it carries none.

Turing (2)

Turing is compute capability 7.5, the oldest architecture among the NVIDIA profiles. A kernel built for compute_75 runs on every NVIDIA profile here, which is why it is the usual target for a test that has to run on all of them.

Use one

Pick a card by id and run a program you built with the real compiler:

vgpu list-gpus
vgpu info --gpu nvidia/h100
vgpu run --gpu nvidia/h100 ./my_app

Also: AMD GPUs, Tenstorrent machines, AWS Neuron machines, or all of them.

Questions