PantheonSim / GPUs / AMD GPUs

Simulated AMD GPUs: 10 profiles

PantheonSim, the GPU simulator for CUDA and ROCm, carries 10 AMD GPUs profiles. They run from Radeon RX 6000 cards to Instinct accelerators, 3 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 gfx target, its multiprocessor count as HIP counts them, its limits and memory, and the values rocm-smi and amd-smi show. The gfx target decides which code objects load. A GPU is named amd/<model>, such as amd/mi300x, 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: CDNA 5, CDNA 4, CDNA 3, CDNA 2, RDNA 4, RDNA 3, RDNA 2.

CDNA 5 (1)

CDNA 5 is the architecture of the Instinct MI455X, built for gfx1250. It is wave32 only, and no physical card exists to check the profile against.

CDNA 4 (1)

CDNA 4 is the architecture of the Instinct MI350 series, built for gfx950. Kernels run in 64-lane wavefronts.

CDNA 3 (2)

CDNA 3 is the architecture of the Instinct MI300 series, built for gfx942. Kernels run in 64-lane wavefronts, and the simulator checks the CDNA 3 code it runs instruction by instruction against ROCm's llvm-objdump.

CDNA 2 (1)

CDNA 2 is the architecture of the Instinct MI200 series, built for gfx90a. Kernels run in 64-lane wavefronts.

RDNA 4 (1)

RDNA 4 is the Radeon RX 9000 generation, gfx1201. Kernels run 32 lanes to a wave, as HIP builds them for it.

RDNA 3 (1)

RDNA 3 is the Radeon RX 7000 generation, gfx1100. Kernels run 32 lanes to a wave, as HIP builds them for it.

RDNA 2 (3)

RDNA 2 is the Radeon RX 6000 generation, gfx1030 and gfx1031. Kernels run 32 lanes to a wave, and there are no matrix instructions, so PyTorch's matrix products run on the vector units.

Use one

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

vgpu list-gpus
vgpu info --gpu amd/mi300x
vgpu run --gpu amd/mi300x ./my_app

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

Questions