AMD unveils Halo Station, a workstation with 576 GB of GPU memory for trillion-parameter AI

AMD took advantage of its IFA 2026 conference to unveil the Threadripper Halo Station, a liquid-cooled workstation designed to run artificial intelligence models of more than a trillion parameters entirely locally, without relying on cloud services. The company has not yet announced when the system will be available, but its presentation alone confirms the growing interest in AI-oriented workstation hardware that gives developers direct control over these workloads, including where data is processed and how models are deployed.

On the technical side, the Halo Station pairs a 96-core, 192-thread Ryzen Threadripper PRO 9995WX processor with up to four Instinct MI350P accelerators. At the event, AMD demonstrated a two-card configuration in operation, adding up to 288 GB of HBM3E memory, while the full four-accelerator version would reach 576 GB of GPU memory — enough, according to the company, to host a trillion-parameter model directly on the machine itself. Each MI350P card contributes 144 GB of HBM3E memory and 4 TB/s of bandwidth, a figure AMD places 14 times above that of any LPDDR5X variant, stressing that this high-speed memory capacity is key to storing model weights and processing data efficiently in large-scale AI workloads.

The system also carries significant power and cooling demands: each MI350P accelerator can draw up to 600 watts of total power, so AMD has opted for independent liquid cooling for each card, plus a dedicated liquid-cooling loop for the Threadripper processor. The latter, code-named Shimada Peak, is based on the Zen 5 architecture, reaches a turbo frequency of up to 5.4 GHz and supports up to 2 TB of DDR5 system memory, reinforcing the machine's profile as a very high-end system aimed at intensive workloads.

AMD has not disclosed pricing, availability or full system specifications, nor has it detailed aspects such as storage, network connectivity or chassis options. It is also unclear whether the company will sell the Halo Station directly or through workstation manufacturers and system integrators, leaving many practical details ahead of launch still to be defined.

Jack Huynh, senior vice president and general manager of AMD's Computing and Graphics division, described the product as a new workstation category intended to bring 'supercomputer-class compute' to individual users and developers. The statement neatly sums up the announcement's positioning: this is not a data center server but a desktop system capable of handling large language models without resorting to cloud infrastructure, a proposition that speaks to the demand for data privacy, cost control and local deployment driving this emerging segment of AI hardware.