AMD outlines comprehensive AI strategy combining Xilinx FPGA acquisition with GPU and software ecosystem development.
AMD is pursuing an ambitious strategy to become a major player in AI compute by developing a broad portfolio of chips covering inference and training across the edge, cloud, and endpoint markets. CEO Lisa Su articulated this vision at AMD's Financial Analyst Day 2022 event, stating: "Our vision here is to provide a broad technology roadmap across training and inference that touches cloud, edge and endpoint, and we can do that because we have exposure to all of those markets and all of those products." Su acknowledged that achieving this goal "will take a lot of work" for AMD to catch up in the AI space, but characterized the market as the company's "single highest growth opportunity."
AMD has already demonstrated early traction in AI compute, with its Epyc server chips being used for inference applications and Instinct datacenter GPUs deployed for AI model training. Multiple cloud service providers are leveraging AMD's ZenDNN library—a Zen Deep Neural Network software optimization integrated with TensorFlow, PyTorch, and ONNXRT that is supported by second and third generation Epyc chips—to provide performance improvements on recommendation engines. Dan McNamara, head of AMD's Epyc business, emphasized that "a large percentage of the inference is happening in CPUs, and we expect that to continue going forward." The upcoming Genoa generation of Epyc chips, arriving later this year, will introduce the AVX-512 VNNI instruction to accelerate neural network processing, with this capability also appearing in the company's Ryzen 7000 desktop chips by year's end.
The $49 billion acquisition of FPGA designer Xilinx, which closed earlier this year, is central to AMD's expanded AI strategy. The AI engine technology from Xilinx—branded under AMD's newly named XDNA banner of "adaptive architecture" building blocks—will be incorporated across multiple products. After its debut in Xilinx's Versal adaptive chip in 2018, the AI engine will be integrated into two future generations of Ryzen laptop chips: Phoenix Point arriving in 2023 and Strix Point in 2024. It will also appear in a future generation of Epyc server chips, though AMD did not specify timing. Additionally, AMD expects to debut its first Zen 5 architecture chips in 2024, which will include new optimizations for AI and machine learning workloads.
On the GPU side, AMD is advancing through the MI200 series of Instinct GPUs and has developed deep partnerships with industry leaders including Microsoft and Meta. David Wang, head of AMD's GPU business, highlighted that AMD has "optimized ROCm for PyTorch to deliver amazing, very, very competitive performance for their internal AI workloads as well as the jointly developed open-source benchmarks." The latest version of ROCm includes optimizations for training and inference workloads on PyTorch and TensorFlow, and support has expanded to consumer Radeon GPUs using the RDNA architecture. AMD is also "developing SDKs with pre-optimized models to ease the development and deployment of AI applications," Wang said.
Looking ahead, AMD is introducing the Instinct MI300, which it calls the "world's first datacenter APU," combining a Zen 4-based Epyc CPU with a GPU using the company's new CDNA 3 architecture. AMD claims the MI300 will deliver "a greater than 8x boost in AI training performance over its Instinct MI250X chip that is currently in the market." Forrest Norrod, head of AMD's Datacenter Solutions Business Group, stated: "The MI300 is a truly amazing part, and we believe it points the direction of the future of acceleration." Through these initiatives spanning CPUs, GPUs, and adaptive architectures, AMD aims to substantially expand its coverage in the AI compute space beyond its traditional datacenter focus.