AMD Advancing AI 2025: Game-Changing Instinct MI350 Series Unveiled with Up to 35x Inference Performance Boost
On June 13 (Beijing Time), AMD hosted its highly anticipated Advancing AI 2025 event, where Chair and CEO Dr. Lisa Su took the stage alongside senior executives, ecosystem partners, and developers to reveal AMD’s latest breakthroughs in artificial intelligence (AI) and high-performance computing (HPC).
Instinct MI350 Series AI GPUs Debut: A Massive Leap in Performance
The highlight of the event was undoubtedly the unveiling of the next-generation AMD Instinct MI350 series GPUs, including the MI350X and flagship MI355X. Built on the new CDNA 4 architecture and cutting-edge 3nm process, these AI GPUs boast an astonishing 185 billion transistors, pushing the boundaries of AI performance.
Key Specifications:
- Memory: 288GB HBM3E
- Model Capacity: Supports up to 520 billion parameter models (520B) on a single GPU
- AI Precision Support: FP4, FP6, FP8, FP16
- Peak Performance:
- FP64/FP32: 2x faster than competitors
- FP6: Over 2x advantage vs. alternatives
- FP4/FP8/FP16: On par or slightly ahead
- Cost Efficiency: 40% better Tokens/\$ compared to NVIDIA B200
According to AMD’s benchmark tests, the MI355X GPU, when running the Llama 3.1 405B model, delivers:
- 4.2x agent performance vs. MI300X
- 2.9x faster content generation
- 3.8x improved summarization
- 2.6x boost in conversational AI performance
Even more impressively, the MI350 series has achieved a 38x improvement in energy efficiency, surpassing AMD’s original 5-year target of 30x.
Building an Open, Scalable AI Infrastructure
The MI350 platform is designed with openness and scalability in mind, supporting UEC and OCP open standards, and offers three large-scale configuration options:
- 128 GPUs: 36TB HBM3E
- 96 GPUs: 27TB HBM3E
- 64 GPUs: 18TB HBM3E
Combined with AMD’s 5th Gen EPYC x86 CPUs, this platform is purpose-built for large-scale AI model training, inference, and deployment, supporting precision modes from FP4 to FP8. Availability begins in Q3 2025 via AMD’s global partners.
Sneak Peek: Helios Platform Featuring MI400 GPU, Zen 6 CPU, and Vulcano Smart NIC
Looking ahead to 2026, AMD also introduced the Helios AI rack-scale architecture, featuring next-generation components:
- Instinct MI400 GPU: 432GB HBM4, 19.6 TB/s bandwidth, 40 PFLOPS (FP4) / 20 PFLOPS (FP8)
- EPYC “Venice” CPU: Based on Zen 6 architecture
- Pensando Vulcano: Advanced smart networking card
The Helios platform aims to set new standards for high-bandwidth, high-efficiency, and ultra-connected AI infrastructure.
ROCm 7 Launches with Major Upgrades + Developer Cloud Opens Globally
Beyond hardware, AMD also announced the official release of ROCm 7, its upgraded open-source AI software stack, now featuring:
- Enhanced support for major AI frameworks
- Richer toolchains, drivers, APIs, and accelerated libraries
- Greater hardware compatibility across AMD platforms
In parallel, AMD is opening its Developer Cloud to the global AI community. This hosted development platform empowers developers to build, test, and deploy AI models—from prototyping to full-scale production.
Final Thoughts: AMD Emerges as a Dark Horse in AI Infrastructure
The Advancing AI 2025 event showcased not only AMD’s massive hardware strides but also its full-stack approach to AI software, developer support, and data center innovation. With its focus on performance, energy efficiency, and open standards, AMD is well-positioned to compete with giants like NVIDIA and Intel in the generative AI race.
As we move into 2025 and beyond, expect AMD Instinct GPUs, MI355X, and the future-ready Helios architecture to play a pivotal role in shaping next-gen AI infrastructure solutions.