AMD has announced the Threadripper Halo Station, a workstation that combines the Ryzen Threadripper PRO processor with Instinct HBM accelerators. The system is designed to allow users to run heavy AI models and hundreds of AI agents locally.
During the keynote at IFA, AMD showcased the Threadripper Halo Station. The workstation is specifically designed for those who currently rely on cloud infrastructure to train or run large language models. According to AMD, the Halo Station aims to eliminate this dependency by running frontier-class models directly on the device. The system will be released to the market in 2027.
Up to 2.6 TB of onboard memory #
The Threadripper Halo Station combines a Ryzen Threadripper PRO 9995WX processor, built on the Zen 5 architecture, with Instinct-class HBM accelerators. The chip features 96 cores and 192 threads. Regarding memory, up to 576 GB of HBM3e memory is available on the GPU, supplemented by up to 2 TB of RDIMM system memory. Together, this accounts for up to 2.6 TB of total system memory.
In terms of bandwidth, AMD promises up to 16 TB/s of combined GPU memory bandwidth and up to 410 GB/s of RDIMM bandwidth. Combined, this delivers a total system memory bandwidth of up to 16.4 TB/s. AMD states that this is more than sufficient to keep heavy models continuously loaded and responsive, even when multiple AI agents are active simultaneously.
Comparison with Nvidia #
In the announcement, AMD explicitly draws a comparison with Nvidia’s DGX Station. According to the company, the Threadripper Halo Station offers up to 3.4 times more total system memory than the competing device when RDIMM and HBM3e memory are combined. With this, AMD clearly positions its new workstation as an alternative for those seeking computing power for AI workloads without compromising on memory capacity.
The company emphasizes that users can save on costs they would otherwise spend on cloud inference. By running models locally, the need to constantly pay for external computing power is eliminated. This can make a significant difference to the budget, especially for organizations with continuous AI workloads.
Target audience and applications #
AMD positions the Threadripper Halo Station as a workstation for those who want to train or execute complex AI models, and for those who want to run hundreds of agentic workflows simultaneously. As such, it is not an entry-level machine, but a device aimed at researchers, developers, and organizations that want to perform heavy AI tasks locally instead of via the cloud.
Concrete pricing information is currently unavailable. AMD does indicate that it is the successor to previous Threadripper workstations, now equipped for the first time with Instinct memory accelerators. The system will not be released to the market until 2027.