AMD (AMD) Advancing AI 2024 summary

Key announcements and product launches

• Introduced 5th Gen EPYC CPUs (Turin) with up to 192 cores, 5 GHz frequency, and industry-leading performance for AI and enterprise workloads.

• Launched MI325X Instinct accelerator with 256 GB HBM3E memory, 6 TBps bandwidth, and up to 40% better inference performance than competitors.

• Previewed MI350 series with CDNA 4 architecture, offering 7x more AI compute and 35x generational performance leap, launching in H2 2025.

• Announced new networking products: Salina 400 DPU and Pollara 400 AI NIC, supporting 400 Gbps and UEC-ready RDMA, available early next year.

• Unveiled Ryzen AI Pro 300 Series for enterprise PCs, featuring Zen 5 CPU, RDNA 3.5 graphics, and 50+ TOPS NPU for Copilot+ PCs.

Strategic partnerships and ecosystem growth

• Deepened collaborations with Google, Oracle, Microsoft, Meta, Dell, HPE, Supermicro, Lenovo, and Databricks, driving co-innovation across cloud, enterprise, and AI workloads.

• Google Cloud to launch Turin-based VMs in early 2025; Oracle scaling MI300X and planning E6 instances; Meta deployed over 1.5 million EPYC CPUs and ramped MI300X for Llama 405B.

• Dell, HPE, Supermicro, and Lenovo announced new server platforms and edge solutions optimized for AMD CPUs and GPUs, with record-setting performance and energy efficiency.

• Microsoft highlighted joint optimization of hardware and software for Azure and Copilot+ PCs, emphasizing performance per dollar per watt.

• Expanded developer ecosystem with over 150 ISVs for Ryzen AI, and broad support for open-source AI frameworks like PyTorch, JAX, Triton, and vLLM.

Software and AI ecosystem advancements

• ROCm AI software stack now supports over one million Hugging Face models, with day-zero support for major releases and significant performance gains in inference and training.

• Acquired Silo AI to deliver end-to-end AI solutions and accelerate customer deployments.

• ROCm 6.2 delivers 2.4x inference and 1.8x training performance improvements over previous versions.

• Panelists from Reka, Fireworks, Essential, and Luma AI reported rapid integration, high performance, and efficient scaling on MI300X and ROCm.

• Developer sessions featured leaders from OpenAI, Meta, Microsoft, xAI, Cohere, and creators of key AI frameworks, fostering community engagement.