AMD ROCm 6.2.4 Successfully Ported to RISC-V: A New Era for RISC-V AI

[Wei Wu (吴伟)]

Hi all,

This is big news. All RISC-V processors with a sufficiently powerful
PCI-E interface can be immediately supported. Fully open sourced:

**Kernel Code (Closely Aligned with Upstream):**
https://github.com/revyos/sg2044-vendor-kernel/commits/sg2044-upstream-v6.15.y

**ROCm Code (Multiple Repositories, Consolidated in a Dedicated
Organization for Easy Access):**
https://github.com/orgs/revyos-rocm/repositories
Upstreaming is planned.

(This article is translated from
https://mp.weixin.qq.com/s/ZrCEq-Kp3jzd8LVhISdLjw )

**Title: ROCm 6.2.4 Successfully Ported to SG2044: A New Era for RISC-V AI!**

The RuyiSDK/RevyOS team are thrilled to announce that AMD ROCm 6.2.4
has been successfully ported to the Sophgo SG2044 platform! This major
breakthrough brings powerful GPU computing capabilities to the SG2044,
opening an unprecedented new chapter for high-performance computing
and artificial intelligence applications. This achievement is not
limited to SG2044, but all RISC-V processors with a sufficiently
powerful PCI-E interface can be immediately supported.

**Image: AMD 7900xtx on Sophgo SG2044** (see mail attachment)

The RuyiSDK/RevyOS team has poured tremendous effort into meticulously
adapting the ROCm software stack to the unique architecture of the
SG2044. This includes:

- **Kernel-Level Adaptation**: Ensuring that ROCm drivers and
underlying components seamlessly integrate with the SG2044’s operating
system and hardware, achieving perfect compatibility for low-level
interactions.
- **User-Space Library and Tool Integration**: Fully incorporating
ROCm’s rich ecosystem, including HIP, ROCr, and other key libraries,
enabling developers to immediately leverage these powerful tools.

**Milestone: ROCm Validated on RISC-V Platform for the First Time**
This is not just a simple port; it is a historic moment—it marks the
first time (within the known public scope) that the ROCm platform has
been successfully validated on a RISC-V architecture!

For a long time, AMD’s ROCm platform has demonstrated its outstanding
performance primarily on x86 architecture systems. Its successful
operation on the SG2044, a RISC-V-based platform, clearly proves
ROCm’s strong portability across instruction set architectures.
This opens wide the door for the emerging RISC-V ecosystem to utilize
AMD GPUs for high-performance computing and AI development, greatly
expanding the future application potential of the RISC-V platform.
This success also further highlights ROCm’s flexibility and
adaptability, breaking the perception of its dependency on specific
hardware architectures.

**Looking Ahead: A New Chapter for RISC-V AI**
In summary, the successful porting of ROCm to the SG2044, along with
the smooth operation of applications like the LLaMA model, is not only
a triumph in model deployment but also a milestone in technological
validation. It heralds a broader development landscape for the RISC-V
architecture in the AI domain and expands the range of hardware
supported by the ROCm platform, paving the way for more exciting
innovations.

We firmly believe that the successful porting of ROCm 6.2.4 to the
SG2044 platform will open new pathways for future innovation and
development. We eagerly look forward to seeing the transformative
applications enabled by this enhanced capability. What ideas do you
have for this new capability?

**Kernel Code (Closely Aligned with Upstream):**
https://github.com/revyos/sg2044-vendor-kernel/commits/sg2044-upstream-v6.15.y

**ROCm Code (Multiple Repositories, Consolidated in a Dedicated
Organization for Easy Access):**
https://github.com/orgs/revyos-rocm/repositories

For technical discussions and exchanges, feel free to engage directly
on GitHub Issues or RuyiSDK.cn, in either English or Chinese.

More detailed data was published before July 11:
https://ruyisdk.cn/t/topic/340




--
Best wishes,
Wei Wu (吴伟)