We attended the very upbeat AMD Advancing AI 2025 Keynote with its CEO, Lisa Su. Guests included hardware partners such as xAI, Meta, Oracle, HUMAIN, Microsoft, and OpenAI. The company announced that its MI355-series GPU ships commercially in 3Q25 and previewed its MI4xx-series, which it expects to ship in 2026. It also previewed its MI500 GPUs, expected to be available in 2027. The company emphasized its ROCm developer environment, its “open” software environment and support for UA Link and Ultra Ethernet Consortium. AMD compared its MI355-series to its competitor’s existing B200-series and found that it has comparable training and fine-tuning performance. Its guests, Oracle and HUMAIN, shared some interesting plans for the near future: Oracle plans to deploy 27,000 GPUs in a single cluster, and HUMAIN plans to build a data center that consumes 50 Megawatts each quarter. Oracle also emphasized that it is pleased with AMD Pensando systems and has been using EPYC CPUs for over a decade. In summary, AMD is scaling up and scaling out to match the needs of OpenAI and Oracle.
It took two years between the MI350 and its predecessor, the MI300, but AMD now expects to produce new Instinct systems at an “annual cadence.” The company compared the training and fine-tuning performance of its MI355X to that of its main competitor’s B200, and the results are comparable, according to charts it showed at the show.
The company also emphasized its support for UA Link and Ultra Ethernet. Astera Labs and Marvell joined AMD on the stage, offering their UA Link support. Soon, UA Link will support up to 1,000 GPUs, according to Forest Norod, SVP at AMD. Ultra Ethernet Consortium (UEC 1.0 went to full release yesterday), with plans to support up to 1 million GPU nodes. The company also introduced its Pollard 400 Pensando NIC (UEC-ready AI NIC).
AMD expects that the Helios AI Rack will be available in 2026. EPYC, Instinct, Pensando, ROCm, UA Link, Ultra Ethernet. 72 GPUs with 260 TB/sec. AMD claims its Helios-era system will deliver 10 times the performance (presumably compared to its MI300-series).
OpenAI CEO Sam Altman joined Lisa Su onstage. MI450 collaboration. Altman confirmed that OpenAI is already running some on MI300X and is “extremely excited about MI450.”
