Skip to content

faq.md

Latest commit

 

History

History
117 lines (79 loc) · 5.09 KB

faq.md

File metadata and controls

117 lines (79 loc) · 5.09 KB

Frequently asked questions

This topic provides answers to frequently asked questions for TheRock users.

General questions

What is TheRock?

TheRock (The HIP Environment and ROCm Kit) is a lightweight, open-source build platform for HIP and ROCm, designed to provide a streamlined and up-to-date ROCm environment.

What does TheRock provide compared to more traditional ROCm releases?

TheRock distributes several types of packages, built daily from the latest ROCm code. These user-space packages are designed to be easy to install, update, and even switch between versions.

Key offerings include:

  • Nightly builds with cutting-edge features
  • Multiple package formats (Python wheels and portable tarballs)
  • Flexible version management without system-level dependencies

Traditional ROCm releases prioritize stability and production use, while TheRock emphasizes rapid access to new developments for contributors and early adopters.

Which GPU architectures are supported by TheRock?

For the most complete and up-to-date information on supported GPU architectures and release history, please refer to the the SUPPORTED_GPUs list, and the RELEASES file.

gfx1151 (Strix Halo) specific questions

Why does PyTorch use Graphics Translation Table (GTT) instead of VRAM on gfx1151?

On Strix Halo GPUs (gfx1151) memory access is handled through GPU Virtual Memory (GPUVM), which provides multiple GPU virtual address spaces identified by VMIDs (Virtual Memory IDs).

GPUVM is the GPU's memory management unit that allows the GPU to remap VRAM and system memory into separate virtual address spaces for different applications, providing memory protection between them. Each virtual address space has its own page table and is identified by a VMID. VMIDs are dynamically allocated to processes as they submit work to the GPU.

On APUs like Strix Halo, where memory is physically unified, there is no discrete VRAM. Instead:

  • Some memory may be firmware-reserved and pinned for GPU use, while
  • GTT-backed memory is dynamically allocated from system RAM and mapped into per-process GPU virtual address spaces.

AI workloads typically prefer GTT-backed allocations because they allow large, flexible mappings without permanently reserving memory for GPU-only use.

For practical implementation details on virtual memory management APIs, see the HIP Virtual Memory Management documentation.

What is the difference between Graphics Address Remapping Table (GART) and GTT?

Within GPUVM, two commonly referenced limits exist:

  • GART defines the amount of platform address space (system RAM or Memory-Mapped I/O) that can be mapped into the GPU virtual address space used by the kernel driver. It is typically kept relatively small to limit GPU page-table size and is mainly used for driver-internal operations.

  • GTT defines the amount of platform address space (system RAM) that can be mapped into the GPU virtual address spaces used by user processes. This is the memory pool visible to applications such as PyTorch and other AI workloads.

Why is allocating to GTT beneficial compared to VRAM?

Allocating large amounts of VRAM permanently removes that memory from general system use. Increasing GTT allows memory to remain available to both the operating system and the GPU as needed, providing better flexibility for mixed workloads. This behavior is expected and intentional on unified memory architectures.

Can I prioritize VRAM usage over GTT?

Yes, if your VRAM is larget than GTT, applications will use VRAM instead. You have two options to prioritize VRAM usage:

  • Increase the VRAM in the BIOS settings.
  • Manually reduce the GTT size so it's smaller than the VRAM allocation (by default, GTT is set to 50% of system RAM).

Note that on APUs, the performance difference between VRAM and GTT is generally minimal.

For information on configuring GTT size, see the next question.

How do I configure shared memory allocation on Linux?

For GPUs using unified memory (including gfx1151/Strix Halo APUs), you can adjust the Graphics Translation Table (GTT) size allocation. See the official ROCm documentation on configuring shared memory.

Note: This applies to Linux systems only and is relevant for any GPU using shared memory, not just Strix Halo.

Troubleshooting

How do I verify my GPU is recognized by TheRock?

See the Verifying your installation section in RELEASES.md for platform-specific instructions.

What should I do if I encounter memory allocation errors?

Check your GTT configuration, ensure sufficient system memory is available, and verify that kernel parameters are correctly set. Review system logs using dmesg | grep amdgpu for specific error messages.