Introduction
The developer environments for AMD and NVIDIA GPUs share many similarities. To ensure a seamless experience, Crusoe has engineered its platform to provide a consistent and familiar environment for cloud compute workflows. The table below offers a high-level comparison between the AMD and Nvidia developer toolsets, highlighting equivalent components where applicable.
AMD vs Nvidia Toolset
| Nvidia Tool | AMD Tool | Description |
|---|---|---|
| CUDA | ROCm | Developer library enabling interaction with the vendor's GPU stack. |
| nvidia-smi | amd-smi | System Management Interface (CLI + library) to query telemetry, monitor performance, and retrieve diagnostics. |
| DCGM | RDCd | Data center GPU monitoring daemon for in-band telemetry. |
| DCGMI | RDCi | CLI tool for initiating diagnostics and interacting with RDCd. |
| NCCL | RCCL | Collective communications libraries that leverage MPI for all-reduce, all-gather, etc., across GPU clusters. |
| NVSwitch / NVLink | Infinity Fabric | Hardware interconnect enabling intra-node GPU communication. Nvidia uses a hub/spoke model; AMD uses a mesh topology. |
| Infiniband | Infiniband / RoCE | High-speed network fabric supported by both vendors; AMD also highlights RoCE (RDMA over Converged Ethernet). |
| GPUDirect | Peer2Peer (P2P) | Direct GPU-to-GPU communication. "GPUDirect" is Nvidia-specific; AMD refers to it generically as Peer2Peer. |
| Nvidia Docker Container Registry | AMD Docker Container Registry | Prebuilt containers for GPU tools in a platform-agnostic environment. |
Relevant AMD Libraries and Tools
AMD, like Nvidia, offers a robust ecosystem of open source tools and libraries for GPU interaction. Below are the key components relevant to Crusoe's developer and operational workflows:
| AMD Tool/Library | Description |
|---|---|
| ROCm | AMD’s full software stack for GPU computing, including compilers, libraries, and drivers. |
| HIP | Heterogeneous-computing Interface for Portability — a C++ runtime API and kernel language designed to ease migration between CUDA and ROCm environments. Included in ROCm. |
| AMD SMI | System Management Interface for querying and controlling AMD GPU hardware (CLI + API). |
| RDC / RDCd / RDCi | ROCm Data Center tools for real-time GPU telemetry, monitoring, and diagnostics. |
| RVS | ROCm Validation Suite — includes built-in tests for hardware validation and benchmarking. |
| UCX | Unified Communication X — low-level library used by MPI for high-performance socket communication. |
| OpenMPI | Standard Message Passing Interface implementation that works with UCX for inter/intra-node communication. |
| RCCL | ROCm Collective Communication Library — optimized for AMD GPUs and leverages OpenMPI for RDMA-based operations. |
Additional Information
- ROCm Documentation: https://rocm.docs.amd.com/en/latest/
- ROCm user Guide: https://rocm.docs.amd.com/projects/ROCmValidationSuite/en/docs-5.0.2/ug1main.html