ROCm 7.0 RC1 release notes

ROCm 7.0 RC1 release notes#

2025-08-07

7 min read time

Applies to Linux

The ROCm 7.0 RC1 is a release candidate for the upcoming ROCm 7.0 major release, which introduces functional support for AMD Instinct™ MI355X and MI350X on single node systems and new features for current-generation accelerators. In this RC1, system support is widened to include more AMD GPUs, Linux distributions, and virtualization options. This preview includes enhancements to the HIP runtime, ROCm libraries, and system management tooling.

This is a first release candidate; expect issues and limitations that will be addressed in upcoming previews.

Important

This preview is not intended for performance evaluation. For the latest stable release for use in production, see the [ROCm documentation](https://rocm.docs.amd.com/en/latest/).

This document highlights the key changes in the RC1 build since the Beta. For a complete history, see the ROCm 7.0 preview release history.

Operating system and hardware support#

This preview supports the following AMD accelerators and Linux distributions in single node setups.

Instinct MI355X, MI350X

Ubuntu	24.04, 22.04
RHEL	9.6
Oracle Linux	9

Instinct MI325X

Ubuntu	22.04

Instinct MI300X

Ubuntu	24.04, 22.04
RHEL	9.6, 8.10
SLES	15 SP7, 15 SP6
Oracle Linux	9, 8
Debian	12

Instinct MI300A

Ubuntu	24.04, 22.04
RHEL	9.6, 8.10
SLES	15 SP7, 15 SP6

Instinct MI250X, MI250, MI210

Ubuntu	24.04, 22.04
RHEL	9.6, 9.4, 8.10
SLES	15 SP7, 15 SP6

Instinct MI100

Ubuntu	24.04, 22.04
RHEL	9.6, 8.10
SLES	15 SP7, 15 SP6

Radeon PRO V710, V620

Ubuntu	24.04, 22.04
RHEL	9.6, 8.10
SLES	15 SP7, 15 SP6

Radeon RX 9000 series

Ubuntu	24.04, 22.04
RHEL	9.6
SLES	15 SP7, 15 SP6

Radeon RX 7000 series

Ubuntu	24.04, 22.04
RHEL	9.6, 8.10
SLES	15 SP7, 15 SP6

See the installation instructions to install ROCm 7.0 RC1 and the Instinct Driver for your hardware and distribution.

Virtualization support#

The RC1 includes support for GPU virtualization on KVM-based SR-IOV and VMware ESXi 8. The following tables detail supported OS configurations per AMD accelerator.

KVM SR-IOV

All supported configurations require the GIM SR-IOV driver version 8.3.0K.

AMD accelerator	Host OS	Guest OS
Instinct MI350X	Ubuntu 24.04	Ubuntu 24.04
Instinct MI325X	Ubuntu 22.04	Ubuntu 22.04
Instinct MI300X	Ubuntu 22.04	Ubuntu 22.04
Instinct MI210	RHEL 9.4	Ubuntu 22.04 or RHEL 9.4
Radeon PRO V710	Ubuntu 22.04	Ubuntu 24.04

ESXi 8

The following configurations are supported on hosts running VMware ESXi 8.

AMD accelerator	Guest OS
Instinct MI325X	Ubuntu 24.04
Instinct MI300X	Ubuntu 24.04
Instinct MI210	Ubuntu 24.04

RC1 release highlights#

This section highlights key features enabled in the ROCm 7.0 RC1.

AI frameworks#

The ROCm 7.0 RC1 supports PyTorch 2.7, TensorFlow 2.19, and Triton 3.3.0.

Libraries#

Composable Kernel#

The RC1 adds functional support for microscaling (MX) data type FP6 in Composable Kernel. This builds upon MX data type support (ROCm 7.0 Alpha).

hipBLASLt#

GEMM performance has been improved for FP8, FP16, BF16, and FP32 data types.

RCCL support#

RCCL is supported for single node functional usage only. Multi-node communication capabilities will be supported in a future release.

HIP#

The following changes improve functionality and runtime performance:

Improved launch latency for device-to-device (D2D) copies and memset operations on AMD Instinct MI300 series accelerators.
Added hipMemGetHandleForAddressRange to retrieve a handle for a specified memory address range. This provides functional parity with CUDA cuMemGetHandleForAddressRange.
Resolved an issue causing crashes in TensorFlow applications. The HIP runtime now combines multiple definitions of callbackQueue into a single function; in case of an exception, it passes its handler to the application and provides the proper error code.

Compilers#

llvm-strip now supports AMD GPU device code objects (EM_AMDGPU).

HIPCC#

The legacy Perl-based HIPCC scripts – hipcc.pl and hipconfig.pl – have been removed.

AMD SMI#

Added:
- New default view when using amd-smi without arguments. The improved default view provides a snapshot of commonly requested information such as bdf, current partition mode, version information, and more. You can obtain the same information in other output formats using amd-smi default --json or amd-smi default --csv.
- New APIs:
  - amdsmi_get_gpu_bad_page_threshold() to get bad page threshold counts.
  - amdsmi_get_cpu_model_name() to get CPU model names (not sourced from E-SMI library).
  - amdsmi_get_cpu_affinity_with_scope() to get CPU affinity.
- API enhancements:
  - amdsmi_get_power_info() now populates socket_power.
  - amdsmi_asic_info_t now also includes subsystem_id.
- CLI enhancements:
  - amd-smi topology is now available in guest environments.
  - amd-smi monitor -p now displays the power cap alongside power.
Optimized:
- Improved overall performance by reducing the number of backend API calls for amd-smi CLI commands.
- Removed partition information from the default amd-smi static CLI command to avoid waking the GPU unnecessarily. This info remains available via amd-smi (default view) and amd-smi static -p.
- Optimized CLI command amd-smi topology in partition mode.
Changed:
- Updated amdsmi_get_clock_info in amdsmi_interface.py. The clk_deep_sleep field now returns the sleep integer value.
- The char arrays in the following structures have been changed.
  - amdsmi_vbios_info_t member build_date changed from AMDSMI_MAX_DATE_LENGTH to AMDSMI_MAX_STRING_LENGTH.
  - amdsmi_dpm_policy_entry_t member policy_description changed from AMDSMI_MAX_NAME to AMDSMI_MAX_STRING_LENGTH.
  - amdsmi_name_value_t member name changed from AMDSMI_MAX_NAME to AMDSMI_MAX_STRING_LENGTH.
- Added new event notification types to amdsmi_evt_notification_type_t: AMDSMI_EVT_NOTIF_EVENT_MIGRATE_START, AMDSMI_EVT_NOTIF_EVENT_MIGRATE_END, AMDSMI_EVT_NOTIF_EVENT_PAGE_FAULT_START, AMDSMI_EVT_NOTIF_EVENT_PAGE_FAULT_END, AMDSMI_EVT_NOTIF_EVENT_QUEUE_EVICTION, AMDSMI_EVT_NOTIF_EVENT_QUEUE_RESTORE, AMDSMI_EVT_NOTIF_EVENT_UNMAP_FROM_GPU, AMDSMI_EVT_NOTIF_PROCESS_START, AMDSMI_EVT_NOTIF_PROCESS_END.
- The amdsmi_bdf_t union was changed to have an identical unnamed struct for backwards compatiblity.
Removed:
- Cleaned up and unified the API by removing unused definitions and redundant components.
  - Removed unneeded API amdsmi_free_name_value_pairs().
  - Removed unused definitions: AMDSMI_MAX_NAME, AMDSMI_256_LENGTH, AMDSMI_MAX_DATE_LENGTH, MAX_AMDSMI_NAME_LENGTH, AMDSMI_LIB_VERSION_YEAR, AMDSMI_DEFAULT_VARIANT, AMDSMI_MAX_NUM_POWER_PROFILES, AMDSMI_MAX_DRIVER_VERSION_LENGTH.
    - Removed unused member year in struct amdsmi_version_t.
  - Replaced amdsmi_io_link_type_t with the unified amdsmi_link_type_t. amdsmi_io_link_type_t is no longer needed. Code using the old enum might need to be updated; this change also affects amdsmi_link_metrics_t, where the link_type field is changed from amdsmi_io_link_type_t to amdsmi_link_type_t.
  - Removed the amdsmi_get_power_info_v2() function as its functionality is now unified in amdsmi_get_power_info().
  - Removed AMDSMI_EVT_NOTIF_RING_HANG event notification type in amdsmi_evt_notification_type_t.
  - Removed enum amdsmi_vram_vendor_type_t. amdsmi_get_gpu_vram_info() now provides vendor names as a string.
- Removed backwards compatibility for the jpeg_activity and vcn_activity fields in amdsmi_get_gpu_metrics_info(). Use xcp_stats.jpeg_busy or xcp_stats.vcn_busy instead. This change removes ambiguity between new and old fields and supports the expanded metrics available in modern ASICs.
Resolved issues:
- Removed duplicated GPU IDs when receiving events using the amd-smi event command.

Instinct Driver/ROCm packaging separation#

The Instinct Driver is now distributed separately from the ROCm software stack and is now stored in its own location in the package repository at repo.radeon.com under /amdgpu/. The first release is designated as Instinct Driver version 30.10. See ROCm Gets Modular: Meet the Instinct Datacenter GPU Driver for more information.

Forward and backward compatibility between the Instinct Driver and ROCm is not supported in the RC1. See the installation instructions.