GIM 8.6.0.K Release Note
We are excited to announce a new open-source release of our GIM SR-IOV Driver.
This release has regular bug fixes and improvements.
Hardware / Features Supported for this Release
| Hardware | Host OS | Guest OS | ROCm version | Number of VFs per GPU | Infinity Fabric (XGMI) Configuration |
|---|---|---|---|---|---|
| AMD Instinct MI355X | Ubuntu 24.04 | RHEL 9.6, RHEL 10, Ubuntu 24.04 | ROCm 7.1.1 | 1 | between 8 GPUs |
| AMD Instinct MI350X | Ubuntu 24.04 | RHEL 9.6, RHEL 10, Ubuntu 24.04 | ROCm 7.1.1 | 1 | between 8 GPUs |
| AMD Instinct MI325X | Ubuntu 22.04 | Ubuntu 22.04 | ROCm 7.1.1 | 1 | between 8 GPUs |
| AMD Instinct MI300X | Ubuntu 22.04 | Ubuntu 22.04 | ROCm 7.1.1* | 1 or 8* | between 8 GPUs |
| AMD Instinct MI300X | RHEL 9.4 | RHEL 9.4, Ubuntu 24.04 | ROCm 7.1.1* | 1 or 8* | between 8 GPUs |
| AMD Instinct MI210X | RHEL 9.4 | RHEL 9.4, Ubuntu 22.04 | ROCm 7.1.1 | 1 | between 4/8 GPUs |
| AMD Radeon PRO V710 | Ubuntu 24.04 | Ubuntu 24.04 | ROCm 7.1.1 | 1 or 4 or 12 | N/A |
Partition Mode Support
| Number of VFs per GPU | Dynamic Compute Partitioning | NPS1 | NPS2 | NPS4 |
|---|---|---|---|---|
| 1 | SPX (Default) | MI300X, MI325X, MI35XX | ||
| 1 | DPX | MI35XX | ||
| 1 | CPX | MI35XX (Preview) | ||
| 8* | CPX | MI300X (Preview) |
For the full list of supported partition modes and the relationship between partition mode and number of VFs per GPU, see the GPU Partitioning.
* Multi-VF (8 VF) Support requires a compatible firmware BKC bundle for the GPU.
Please refer to the ROCm Firmware Compatibility Table for the list of required firmware versions and supported configurations.
Host AMD SMI Versions
- Library: 50.2.2
- Tool: 32.4.1
Please checkout out SMI doc for more details.
Getting Started
Please check out our User Guide for instructions on how to set up GIM and example configurations to run SR-IOV enabled VMs.
Known Issues
- When using the configuration of 64 ROCM devices in 1 VM (aka CPX mode with 1 VF) on a MI350 platform, we would see a hang in workloads which use SDMA engines
- MultiVF configuration for MI300X requires specific firmware version. please refer to the ROCm compatibility matrix for supported version
- MI210: Unloading the Guest driver on MI210 may produce kernel call trace. Note that this is just a warning and not a kernel panic. Guest VM reboot is required to restore the operation
- SMI CLI topology --link_type command returns wrong value for the first BDF
- SMI CLI static --vram command returns 0 for the bit width value
- Some BDF values returned from SMI get_link_topology_nearest APIs are wrong