AMD Threadripper Halo Station: Bringing Trillion-Parameter AI Local
AMD used its IFA 2026 keynote to introduce the Threadripper Halo Station. This new liquid-cooled workstation is designed specifically for running AI models with more than one trillion parameters locally, removing the reliance on a cloud service.
The system points to a growing market for workstation-class AI hardware. This emerging category gives developers more direct control over large-model workloads, allowing them to dictate exactly where data is processed and how models are deployed. Jack Huynh, AMD's senior vice president and general manager of Computing and Graphics, described the product as a new workstation category. He stated it is intended to bring "supercomputer-class compute" to individual users and developers.
Currently, AMD has not announced when the Halo Station will become available or what its pricing will be. The company also did not detail full system specifications, meaning storage, networking, and chassis options remain unknown. Furthermore, it is not yet clear whether AMD plans to sell the system directly or through workstation manufacturers and system integrators. TechSpot, which has served tech enthusiasts for over 25 years, provided a first look at the new hardware during the event.
Core Specifications and Processor Architecture
The Halo Station pairs a 96-core Ryzen Threadripper PRO 9995WX processor with up to four Instinct MI350P accelerators. Code-named Shimada Peak, the Threadripper PRO 9995WX is based on AMD's Zen 5 architecture.
This processor delivers substantial compute density, featuring 96 cores and 192 threads. It is capable of reaching a boost clock of up to 5.4 GHz. For massive data sets, the processor supports up to 2 TB of DDR5 system memory. To manage the thermal output of these powerful components, the Threadripper processor utilizes its own dedicated liquid-cooling setup.
GPU Acceleration and Memory Bandwidth
The workstation's massive AI processing capabilities rely on the Instinct MI350P cards. Each individual Instinct MI350P accelerator includes 144 GB of HBM3E memory and delivers 4 TB/s of memory bandwidth.
This memory capacity and speed are central to the Halo Station's primary purpose. Large AI models require substantial high-bandwidth memory to store weights and process data efficiently. AMD noted that the MI350P's 4 TB/s of bandwidth is 14 times higher than that of any LPDDR5X variant.
Supporting this level of performance requires significant power and thermal management. Each MI350P accelerator can draw up to 600 watts of total board power. To handle this heat generation safely, AMD noted that the accelerator cards are independently liquid-cooled.
Configuration Options and Capacity
During the keynote at IFA, AMD demonstrated a two-card version of the workstation. This specific configuration featured 288 GB of HBM3E memory.
A fully configured model utilizing four accelerators would provide a massive 576 GB of HBM3E memory. This 576 GB capacity is enough, the company stated, to hold a trillion-parameter model entirely on the machine itself.