At the Advancing AI 2024 conference in San Francisco, AMD unveiled its Instinct MI325X AI accelerator, aimed directly at competing with NVIDIA's Blackwell chips. The newly introduced MI325X promises advancements in memory and computing power over NVIDIA's H200 platform, boasting 1.8 times more memory, 1.3 times higher memory bandwidth, and 1.3 times greater computing power.
The MI325X AI accelerator, which features AMD's CDNA 3 GPU architecture and is equipped with next-generation HBM3E high-bandwidth memory, is set to begin production in late 2024, with customer deliveries starting in the first quarter of 2025. AMD also revealed its AI chip roadmap, citing plans to release the MI350 series next year using the CDNA 4 architecture, with MI355X achieving peak AI performance of 74 PF.
AMD's CEO, Lisa Su, emphasized that the market for data centers, AI, and accelerators is expected to grow to $500 billion by 2028, driven significantly by generative AI. Su highlighted four core elements of AMD's AI platform: powerful computing engines, open software solutions, a collaborative AI ecosystem, and optimized system design.
Despite its advancements, AMD faces challenges in capturing enterprise market share, as noted by industry analyst Patrick Moorhead. NVIDIA's longstanding dominance in data center GPUs, supported by its CUDA software ecosystem, poses a significant hurdle for AMD. However, AMD is working on enhancing its ROCm software to facilitate AI development on its platforms.
Moorhead also noted that AMD's new GPU, particularly the MI350, shows improved efficiency and performance, marking significant progress. As AMD continues to innovate, the competition with NVIDIA remains intense, with AMD positioning itself as a viable alternative in the AI chip market.
