AMD’s Threadripper Halo: A High-End Local AI Workstation for Researchers
AMD unveils a powerful AI workstation with up to 576 GB of HBM3e memory.
1Introduction to Threadripper Halo
AMD has introduced the Threadripper Halo, a high-performance AI workstation designed for researchers. Announced at IFA 2026, this system is AMD’s first workstation featuring Instinct accelerators. It boasts up to 576 GB of HBM3e memory and is powered by a 96-core Threadripper PRO 9995WX CPU. The workstation is aimed at running large AI models locally, offering significant computational power for advanced research.
2Key Specifications
The Threadripper Halo includes a liquid-cooled 96-core Threadripper PRO 9995WX CPU, up to 2 TB of DDR5 memory, and up to 2.6 TB of combined system memory. It supports up to four PCIe-based Instinct GPUs, each with 144 GB of HBM3e memory. The system promises up to 576 GB of HBM3e memory and 16 TB/s of memory bandwidth, enabling it to handle models with over a trillion parameters at four-bit precision.
3Performance Capabilities
With its high memory bandwidth and GPU configuration, the Threadripper Halo can run large frontier-class AI models entirely in GPU memory. It is capable of handling models like Moonshot.AI’s 2.8 trillion-parameter Kimi K3. The system’s performance is expected to surpass Nvidia’s DGX Station in LLM inference tasks, thanks to its superior memory bandwidth and system memory capacity.
4Power and Cooling Requirements
The Threadripper Halo operates at a power consumption of up to 600 W per GPU, though AMD plans to run the cards at a more efficient 450 W configuration in this system. This high power requirement may necessitate a 20 amp circuit, limiting its compatibility with standard North American power outlets. The system’s liquid cooling ensures optimal performance under heavy workloads.
5Market Positioning
AMD positions the Threadripper Halo as a direct competitor to Nvidia’s DGX Station, which retails for around $100,000. The Threadripper Halo offers up to 3.4x the total system memory and more than twice the memory bandwidth of the DGX Station. This makes it a compelling option for researchers seeking high-end AI workstations for local model training and inference.
6Availability and Pricing
The Threadripper Halo is set to launch next year, with pricing expected to range between $100,000 and $150,000. AMD has not yet disclosed exact pricing details, but the system’s high-end specifications suggest it will cater to a niche market of researchers and institutions with substantial budgets for AI infrastructure.
7Comparison with Nvidia’s DGX Station
The Threadripper Halo outperforms Nvidia’s DGX Station in several key areas, including memory bandwidth and system memory capacity. While the DGX Station features a 252 GB B300 GPU and 72-core Grace CPU, the Threadripper Halo’s 96-core CPU and up to four Instinct GPUs provide a significant edge in computational power. This makes it a strong contender in the high-end AI workstation market.
8Significance for AI Research
The Threadripper Halo represents a significant advancement in local AI research capabilities. Its ability to handle trillion-parameter models locally reduces reliance on cloud-based solutions, offering researchers greater control and faster processing times. This workstation is poised to accelerate breakthroughs in AI and machine learning, particularly in fields requiring extensive computational resources.