AMD Targets High-End Researchers With New Threadripper Halo Workstation

AI-generated image · US National Wire
The company is betting on a high-margin, local AI alternative to cloud subscriptions by bringing Instinct accelerators to the desktop.
AMD is moving to capture the high-net-worth research market with the announcement of the Threadripper Halo, a local AI workstation designed to compete with Nvidia's DGX Station. As The Register first reported, the system marks the first time AMD has offered its Instinct accelerators in a workstation form factor.
Designed for researchers who can afford significant upfront costs to bypass cloud-based dependencies, the Threadripper Halo is powered by a 96-core Threadripper PRO 9995WX CPU. The system supports up to 2 TB of DDR5 memory and can be equipped with up to four PCIe-based MI350P Instinct GPUs. In a quad-GPU configuration, the machine provides up to 576 GB of HBM3e memory and 16 TB/s of memory bandwidth, allowing it to run models exceeding a trillion parameters in GPU memory. The Register notes that by utilizing system memory, the workstation could support massive open weights models, such as the 2.8 trillion-parameter Kimi K3 from Moonshot.AI.
While AMD has not officially disclosed pricing, The Register estimates the system will retail between $100,000 and $150,000. This positions it against Nvidia's DGX Station, which retails for approximately $100,000 and features a 72-core Grace CPU and a 252 GB B300 GPU. AMD claims the Threadripper Halo offers more than double the memory bandwidth and up to 3.4x the total system memory of the DGX Station.
Availability is expected to begin next year, though The Register warns that the power requirements for a quad-GPU setup may necessitate electrical service upgrades for North American users.

