AMD MI350 and CDNA 4 Architecture Launched with ROCm 7

AMD has launched its Instinct MI350 series, based on the CDNA 4 architecture, targeting AI optimization with a focus on lower precision compute and increased memory capacity and bandwidth. The MI350 series uses the OAM UBB form factor and features new accelerator compute dies (XCD) on the N3P process. AMD is also emphasizing software development with ROCm 7, aiming for easier installation and broader support, including notebooks and Windows, signaling a commitment to AI and developer accessibility.

Analysis

Advanced Micro Devices (AMD) has unveiled its Instinct MI350 series, underpinned by the new CDNA 4 architecture, signaling a strategic pivot towards AI optimization, particularly in lower precision compute. This new generation, utilizing the industry-standard OAM UBB 8-GPU form factor, sees its accelerator compute dies (XCDs) fabricated on an N3P process, an upgrade from the previous N5. Architecturally, the MI350 features eight XCDs, totaling 256 compute units, which, while fewer than the MI300X/MI325X, are described by AMD as 'beefier' due to enhancements within CDNA 4. A significant change includes a reduction in I/O dies (IODs) from four to two, which AMD suggests streamlines data pathways by having each IOD manage a larger compute and memory topology. The MI350 series emphasizes increased memory capacity and bandwidth per compute unit to mitigate data bottlenecks, a common challenge in AI workloads. Notably, CDNA 4 enhances support for FP4 and FP6 precision, with a particular focus on FP6, while TF32 support is relegated to software emulation. Physically, the liquid-cooled MI355X platforms can scale to 128 GPUs within approximately 48U of rack space, a density figure AMD contrasts with Nvidia's GB200 NVL72 offering 72 GPUs per rack. Concurrently, AMD launched ROCm 7, its updated software stack, underscoring a significant investment in improving developer experience with initiatives like 'pip install rocm' for easier installation, the AI Developer Cloud, and Enterprise AI efforts. AMD also plans to extend ROCm support to notebooks and provide full Windows compatibility (without WSL) by late 2025, indicating a comprehensive strategy to build out its AI ecosystem from developer workstations to large-scale data centers.

Key Decisions for Investors

• Investors should recognize AMD's MI350 launch and CDNA 4 architecture as a targeted effort to capture greater share in the AI accelerator market, focusing on enhanced lower-precision performance, memory bandwidth, and a competitive rack-scale density offering.
• The significant emphasis on software development, highlighted by ROCm 7 and its planned expansion to broader platforms including Windows, is a critical factor to monitor for its potential to improve developer adoption and address historical software ecosystem challenges relative to competitors.
• Consider the implications of AMD's architectural choices, such as the shift away from HPC FP64 dominance towards AI-centric low-precision formats like FP6, and assess how these align with evolving AI model requirements and total cost of ownership for data center deployments.