AMD pushes agentic AI to local PCs

LOCAL processing is emerging as a key requirement for a new generation of AI applications as software agents take on more complex tasks, AMD said during an online briefing Wednesday, Sept. 30, outlining its approach to agentic PCs and its Ryzen AI portfolio. Michael Nordquist, AMD corporate vice president for client product marketing, said agentic computing differs from earlier generative AI use because systems are expected to work through tasks and outcomes rather than simply respond to individual prompts.

“As we move into agentic, that means you have an agent working on your behalf on more complex tasks where you really try to assign it a task or an outcome and actually reason through it,” Nordquist said. He said this shift changes computing requirements because agents can run repeatedly and locally, increasing the importance of CPU, GPU, NPU and memory capacity.

Nordquist said AMD has shipped tens of millions of AI PCs across more than 250 designs, while agentic PCs have surpassed 500,000 units across more than 50 designs. Get the latest news delivered to your inbox Sign up for The Manila Times newsletters By signing up with an email address, I acknowledge that I have read and agree to the Terms of Service and Privacy Policy .

The Ryzen AI Max platform combines a Zen 5 CPU with up to 16 cores and 32 threads, up to 192GB of unified memory, RDNA 3.5 graphics and an XDNA 2 NPU. AMD said variable graphics memory can allocate up to 160GB to graphics workloads. “The unified memory, 192 gigs, that's kind of the secret sauce here is the unified memory bringing that all together,” Nordquist said.

“That allows me to take these really large models and bring them into a local device.” AMD said the platform can run large language models locally, including a 320 billion-parameter GLM 5.3 Flash model. Its September 2026 testing showed up to 20 tokens per second for the model and up to 42 tokens per second for Qwen 3.8 Flash Next configurations.

AMD said performance may vary. Nordquist said local execution can also address the cost of cloud-based AI workloads, where usage is commonly metered by tokens. “If you're able to take that and run it all locally, sure you have the device cost or you have the power cost, but it's essentially almost like free tokens that you're able to do and you're still able to get that quality we talked about on the last page,” he said.

A case study involving Emmy Award-winning production company Light Sail VR illustrated one use of local agents. The company developed an AI production coordinator on a Framework desktop powered by an AMD Ryzen AI Max processor and later deployed it on a system with AMD Radeon AI PRO R9700 graphics.

The coordinator is accessed through Slack and handles schedules, deadlines, notes and stakeholder information. Light Sail VR said the system reduced the administrative load involved in managing multiple productions. Founder and CEO Matthew Celia said the company can now oversee six to nine jobs with a comparable level of personal effort, compared with one or two previously.

“For us at Light Sail — we’re a small company, we’re a small team,” Celia said. “In order to operate as a small team, we need to think big. We have to be really efficient.” The case study said Light Sail VR is evaluating additional agents, including one that could review interview material and help assess episode structure while human creators retain editorial control.

AMD noted that the performance and cost-savings claims were provided by Light Sail VR and had not been independently verified by AMD. Nordquist said the shift toward local agentic computing also depends on software and enterprise support. AMD’s portfolio includes Ryzen AI processors, Ryzen AI PRO processors, Ryzen AI Max processors, Radeon AI PRO graphics, the Ryzen AI Halo developer platform and Threadripper Halo, supported by ROCm software.

“Having hardware is great, but if you don't have the ecosystem and the software infrastructure for it, you really can't capitalize on some of the capability,” Nordquist said.