Early this morning, AMD held its 2025 Global AI Development Conference, primarily announcing and introducing its latest AI chips and cloud infrastructure hardware.
OpenAI co-founder and CEO Sam Altman attended the conference as a special guest and jointly announced the Instinct MI400 and Instinct MI350 series of powerful AI chips with AMD. Notably, during the development process, OpenAI continuously provided technical feedback to AMD, helping to optimize its GPUs.
At the press conference, Sam Altman was astonished to hear that a single MI400 would be equipped with 432GB of HBM4 memory, exclaiming, "Impossible!"
AMD's newly released AI chips are mainly set to compete with Nvidia's Blackwell chips, as Nvidia is currently AMD's sole competitor in the AI data center GPU market.
The AMD Instinct™ MI350 series GPUs are the latest products based on the AMD CDNA™ 4 architecture, designed for the demands of modern AI infrastructure. This series includes two GPUs: MI350X and MI355X.
Compared to the previous generation, the MI350 is equipped with 288GB of HBM3E memory and up to 8TB/s of memory bandwidth, boosting AI computing capabilities by 4 times and inference performance by 35 times.
AMD stated that because the chip consumes less power than competitors, the MI355X can deliver 40% more tokens per dollar than Nvidia's chips.
The MI355X platform achieves 161 PFLOPS in FP4 performance, while the MI350X platform reaches 36.8 PFLOPS in FP16 performance. These GPUs not only excel in performance but also offer flexible cooling configurations, including air cooling and direct liquid cooling, supporting large-scale deployments, such as up to 64 GPUs in an air-cooled rack, or up to 128 GPUs in a direct liquid-cooled environment.
To further enhance GPU performance, AMD also open-sourced an AI acceleration platform, ROCm7. Over the past year, ROCm has rapidly matured, delivering leading inference performance, expanding training capabilities, and deeply integrating with the open-source community. ROCm now supports some of the world's largest AI platforms, such as LLaMA and DeepSeek, and will provide over 3.5 times the inference performance boost in the upcoming ROCm 7 release.
ROCm Enterprise AI offers a complete MLOps platform for AI deployments, supporting secure, scalable AI development and providing a rich set of tools for fine-tuning, compliance, deployment, and integration.
The Instinct MI400 is AMD's next-generation flagship AI chip and the core component of the "Helios" AI integrated system. In terms of memory configuration, the MI400 series is expected to feature up to 432GB of HBM4 high-speed memory, a significant increase from the MI350 series' 36TB HBM3E memory. This high-bandwidth memory architecture provides ample data throughput for large AI models, meeting the demands of model parameter loading and rapid computation.
Regarding computing performance, the MI400 series can achieve 40 petaflops of compute power at FP4 precision. This metric is specifically optimized for low-precision computations in AI training, effectively accelerating the training efficiency of mainstream models like Transformer. Additionally, its 300GB/s scale-out bandwidth, achieved through the UALink open standard technology, enables seamless interconnection of 72 GPUs, allowing all GPUs within a rack to work together as a unified computing unit, breaking through traditional architectural communication bottlenecks.
The MI400 series forms a technological synergy with the 6th Gen AMD EPYC "Venice" CPU and Pensando "Vulcano" AI NIC. The Venice CPU, based on the Zen 6 architecture, provides up to 256 cores and 1.6TB/s of memory bandwidth, ensuring efficient task scheduling and resource management for GPU clusters.
Meanwhile, the Vulcano AI NIC supports 800G network throughput, and its UALink and PCIe dual-interface design enables low-latency data transfer between GPUs and CPUs, offering an 8x increase in scale-out bandwidth compared to previous generations, effectively resolving communication congestion issues in high-density clusters.
In terms of architectural design, the MI400 series adopts the open-standard UALink technology, distinguishing it from Nvidia's proprietary interconnect solutions. This technology enables high-speed connections between GPUs via Ethernet tunneling, supporting rack-level unified compute resource pooling. Coupled with the open architectures of OCP and the Ultra Ethernet Consortium, it ensures compatibility with existing data center infrastructures. The MI400 is expected to be available in 2026.
In addition to OpenAI, seven major AI development platforms, including Microsoft, Oracle, Meta, and xAI, are collaborating with AMD to utilize its AI chips.
Oracle will be one of the first industry leaders to adopt Instinct MI355X-powered rack-scale solutions, highlighting Oracle's commitment to providing the broadest AI infrastructure. Oracle Cloud Infrastructure (OCI) supports various critical enterprise workloads with stringent requirements for scalability, reliability, security, and performance.
Mahesh Thiagarajan, Executive Vice President of Oracle Cloud Infrastructure, stated that Oracle Cloud Infrastructure continues to benefit from its strategic collaboration with AMD. "We will be one of the first companies to offer MI355X rack-scale infrastructure using the combined power of EPYC, Instinct, and Pensando."
"We've seen impressive customer adoption of AMD-powered bare metal instances, underscoring how easily customers can adopt and scale their AI workloads. Additionally, Oracle widely relies on AMD technology internally for its own workloads and externally for customer-facing applications. We plan to continue deep cooperation across multiple AMD product generations and are confident in AMD's roadmap and its continued ability to meet expectations."
This article's material is sourced from AMD; please contact us to remove any infringements.
END
Click the image to register now 👇️