In a new wave of artificial intelligence represented by "Agentic AI", Unified Memory Architecture (UMA) is rapidly heating up in the PC and computing fields. AMD believes that this is not only an important opportunity, but also a key direction for future product architecture and roadmap.
The so-called unified memory architecture refers to tightly coupling the CPU, GPU and memory on the same system chip (SoC) to form a shared large memory pool, which is dynamically allocated between the CPU and GPU by the system according to load, instead of the traditional "system memory + independent video memory" separation model. In AI workloads, especially large model inference, this design can significantly reduce data copy and bandwidth bottlenecks, so it is gradually becoming one of the mainstream solutions for AI terminals and new forms of PCs.
With the launch of the AMD Ryzen AI MAX series and the addition of NVIDIA RTX Spark and other products, the unified memory architecture has become the common technical cornerstone of AI terminal platforms. AMD said that their first-generation Ryzen AI MAX solution can provide up to 128GB of memory, of which up to 112GB of system memory can be divided for the GPU; NVIDIA RTX Spark also uses a similar idea to dynamically allocate memory between the CPU and GPU based on the workload, allowing the unified memory architecture to cover a wide range of application scenarios from general computing to AI inference.
In an interview with the media, AMD Vice President David McAfee was asked whether more products will use UMA solutions in the future. He responded that the focus on unified memory systems will continue to increase, and the industry will explore "appropriate architectural forms" around such systems and continue to iterate and enhance based on existing platforms. He emphasized that this is a new type of workload and computing space that will open "a whole world of possibilities" for AMD in product selection, road planning, and deployment forms.
AMD has extended its unified memory thinking to its next-generation products, the Ryzen AI MAX 400 series. According to the official introduction, this generation of products can support up to 192GB of unified memory, and can allocate up to 160GB to the GPU, which can be used to locally run large language models (LLM) with a parameter scale of more than 300 billion levels to meet the extreme demands for memory capacity and bandwidth of complex AI workflows and high-end creative workloads.
In the media roundtable, a reporter further asked whether it is possible to see UMA Ryzen processors for games in the future, or a design similar to "Strix Halo + 3D V-Cache / package-level high-bandwidth memory" to further enhance UMA capabilities through tighter integration and lower-latency packaged memory. McAfee said he currently has "no specific answer," but reiterated that platforms such as Strix Halo are entering the same track as Nvidia, which means that system design around UMA will receive more resource investment and architectural exploration in the next few years.
It is worth noting that when McAfee talked about UMA, he not only mentioned mobile and AI terminals, but also high-performance desktop systems. He believes that the continuous improvement of the unified memory architecture's support capabilities and the adoption of this architecture by more ecological participants will promote the overall evolution of high-performance desktops and unified system forms, and reshape the industry's understanding of "high-performance PC + unified memory". In his view, the unified architecture adopted by platforms such as Halo is still the "correct form" of this type of system, and Nvidia's recent related releases can be regarded as an "endorsement" of this architectural path.
McAfee also highlighted that with the rise of Agentic Compute, running "very large models" on endpoints via a unified memory pool becomes one of the unique value propositions of these systems. For AMD, this type of unified system plays a dual role in the overall product portfolio: on the one hand, it supports cutting-edge AI and large model workloads, and on the other hand, it may also become the basic platform form for high-performance desktops and advanced creative workstations.
From an industry perspective, unified memory architecture is no longer a niche experiment, but has rapidly evolved into a basic pillar in new generation computing platforms. As Agentic AI's demand for large-capacity shared memory pools continues to increase, manufacturers such as AMD and NVIDIA have jointly bet on UMA, which also means that this architectural route has received strong industry-level endorsement. AMD's active planning for new platforms such as Ryzen AI MAX 400 and its open attitude towards the future form of high-performance platforms such as Strix Halo show that it is still only the starting point for the development of unified memory architecture.
In a unified system where the boundaries between CPU, GPU and memory are gradually blurred, the new generation platform is expected to achieve simultaneous leaps in performance, energy efficiency and capability boundaries. This is not only applicable to AI and large model workloads, but may also be extended to games and high-end desktop fields. For AMD, the unified memory architecture is becoming one of the core foundations in its next-generation product architecture design and mid- and long-term road planning, and the entire ecosystem has just entered the starting stage of this path.