AMD unveils AI GPU to challenge Nvidia’s Rubin

AMD unveiled the Instinct MI455X accelerator at its Advancing AI 2026 event, a flagship GPU with 320 billion transistors, 432GB of HBM4 memory, and up to 40 petaflops of FP4 AI compute, designed to challenge Nvidia's Rubin GPU. The MI455X, built on AMD's CDNA 5 architecture and manufactured with TSMC's 2nm and 3nm processes, offers 50% more memory than Rubin and slightly higher bandwidth, and will power AMD's Helios rack-scale platform. Shipping begins at the end of Q3 2026, with volume ramping through Q4 2026 and into H1 2027.

AMD https://www.networkworld.com/article/4021291/amd-latest-news-and-insights.html has answered Nvidia’s Rubin GPU https://www.networkworld.com/article/4188058/nvidia-unveils-vera-rubin-platform-targeting-ai-hpc-infrastructure-customers.html v challenge with the introduction of the Instinct MI455X accelerator, its flagship processor designed for large-scale AI training and inference with absolute whopping specs. The MI455X, unveiled during AMD’s https://www.networkworld.com/article/4089519/amd-outlines-ambitious-plan-for-ai-driven-data-centers.html recent Advancing AI 2026 event, is the new flagship of the company’s Instinct MI400 family and will power its Helios rack-scale AI platform https://www.networkworld.com/article/4199402/helios-marks-amds-biggest-ai-infrastructure-push-yet.html . The launch shows that even though it is lagging in terms of market share, AMD is at least able to match Nvidia when it comes to technology. Spec-wise, the MI455X almost matches and in some cases surpasses the Rubin architecture. The Instinct MI455X boasts 320 billion transistors, 432GB of HBM4 memory, and up to 40 petaflops of FP4 AI compute. The memory is significant and an example of why there is such a memory shortage https://www.networkworld.com/article/4119222/whats-causing-the-memory-shortage.html currently . Rubin is expected to deliver approximately 50 PFLOPS of FP4 performance but has 288GB of HBM4 memory, 50% less than Instinct. AMD also claims memory bandwidth of up to 23.3 TB/s, slightly exceeding Rubin’s 22 TB/s. The MI455X is built on AMD’s new CDNA 5 architecture, representing the company’s most significant accelerator redesign in years. Manufactured using a combination of TSMC’s 2nm and 3nm process technologies, the chip incorporates 320 billion transistors through an advanced chiplet design connected with 3D hybrid bonding. The architecture introduces improvements in tensor processing, cache bandwidth, memory movement, and execution efficiency aimed at accelerating large language models and agentic AI workloads. With up to 40 PFLOPS of FP4 performance and 20 PFLOPS of FP8 compute, the Instinct 455X roughly doubles the AI compute capability of the current Instinct MI350 generation. It also massively increases HBM memory bandwidth from roughly 8 TB/s on MI350 to more than 23 TB/s on the new accelerator. Both NVIDIA and AMD have stopped positioning their GPUs as standalone products. Rather, they are positioned as a component in an overall compute system that includes CPU, networking, security, and software. The Helios rack-scale platform — AMD’s answer to Nvidia’s DGX platform — links 72 GPUs into a unified system. A fully configured Helios rack features Instinct accelerators, EPYC processors, Pensando networking, and the ROCm software stack. AMD says it provides up to 31TB of HBM4 memory and as much as 2.9 exaflops of FP4 AI compute. So for now, AMD has a viable competitor to Nvidia https://www.networkworld.com/article/4140107/amd-accelerates-telecom-network-ai.html at the high end. Whether this moves the needle and helps AMD gain ground remains to be seen. But on paper, the MI455X narrows the hardware gap considerably, offering competitive AI compute alongside substantially more memory than Nvidia’s Rubin architecture, which is a combination designed to appeal to customers training ever-larger AI models. The Instinct MI455X will begin shipping at the end of the third quarter of 2026, with volume ramping through the fourth quarter and into the first half of 2027.