AMD is launching the Threadripper Halo Station, a massive AI workstation designed to bring datacenter-level processing power directly to a desk. This shift matters because it allows researchers and developers to run trillion-parameter models locally without relying on cloud infrastructure. The system aims to address the memory limitations that typically constrain high-end AI inference on personal hardware.
Workstation features 96-core CPU and up to four MI350P accelerators
The workstation is built around the AMD Ryzen Threadripper PRO 9995WX processor, which packs 96 Zen 5 cores into a single socket. To handle heavy workloads, the chassis accommodates up to four AMD Instinct MI350P GPUs based on the CDNA 4 architecture. This configuration creates a unified system where the CPU and GPU accelerators work together for complex AI tasks.
Specifications
- CPU: AMD Ryzen Threadripper PRO 9995WX (96 Zen 5 cores)
- GPU Accelerators: Up to 4x AMD Instinct MI350P (CDNA 4 architecture)
- GPU Memory: Up to 576GB HBM3e
- System Memory: Up to 2TB RDIMM
- Total Memory: 2.6TB
Memory capacity defines the core capability of this machine, offering a total of 2.6TB across the system. The GPU accelerators provide up to 576GB of HBM3e memory, while the system supports up to 2TB of RDIMM system memory. This combination delivers up to 16.4TB/s of total memory bandwidth, with 16TB/s dedicated to GPU operations and 410GB/s for system memory.
AMD claims the Threadripper Halo Station provides 4.6 PFLOPs of AI compute per MI350P GPU. The company notes this setup offers 3.4 times more total system memory than the competing NVIDIA DGX Station. The entire assembly is housed in a custom E-ATX chassis with a tempered glass side panel and relies on a liquid cooling solution to manage heat.
The device is currently on track for a 2027 release, according to AMD's projections. Pricing details have not been announced yet. This workstation targets users who need large amounts of local memory for AI model inference but lack access to or cannot afford dedicated datacenter facilities.
Source: TweakTown
Discussion
0 comments