AMD Unveils Threadripper Halo Station with 96 Cores for Local AI Development

AMD unveiled the Threadripper Halo Station at the IFA 2026 opening keynote in Berlin. Built around a 96-core Zen 5 processor and liquid-cooled accelerators, the high-end developer workstation is designed to run trillion-parameter AI models locally without cloud connectivity.

Stepping firmly onto the high end of the developer workstation market, AMD used its opening keynote at the IFA tech convention in Berlin to reveal the Threadripper Halo Station. The tower system pairs a flagship Zen 5 workstation processor with up to four discrete artificial intelligence accelerators, aiming to deliver supercomputer-grade compute directly to individual engineers and developers.

Jack Huynh, AMD’s senior vice president and general manager of Computing and Graphics, introduced the new hardware as part of an AI-centric opening keynote at the Berlin event. According to the company, the development system is intended to sit alongside and augment AMD’s existing Ryzen AI Halo box, a smaller mini-PC form factor powered by Ryzen AI Max+ Strix Halo processors. While those compact boxes target mainstream desktop AI use, the new Threadripper system tackles the heavier end of the market.

Zen 5 Architecture and Processor Specifications

At the center of the workstation sits a 96-core processor from the Threadripper Pro lineup. Huynh explicitly confirmed that the showcased demo unit utilized the flagship Threadripper Pro 9995WX, codenamed “Shimada Peak”. This Zen 5 chip packs 96 cores and 192 threads, features a boost frequency reaching 5.4 GHz, and incorporates eight channels of DDR5 memory alongside 128 lanes of PCIe Gen5 connectivity.

The motherboard powering the demonstration unit was identified as ASUS’s Pro WS WRX90E-SAGE SE, a WRX90 platform designed specifically for Threadripper Pro 7000 and 9000 series processors. Out of the box, the system supports up to 2 TB of DDR5 system memory.

Instinct MI350P Accelerators and Liquid Cooling

While the processor handles general system demands, the bulk of the heavy lifting for artificial intelligence workloads falls to AMD’s Instinct accelerator cards. The Halo Station utilizes the Instinct MI350P, a cut-down version of the CDNA 4 architecture MI350X accelerator configured specifically to fit within a PCIe form factor.

Each MI350P card comes equipped with 144 GB of HBM3E memory running at 4 TB/s, delivering a bandwidth fourteen times higher than any LPDDR5X variant, according to details shared at IFA 2026. Each card also packs enough matrix cores to provide 2.3 PFLOPS of MXFP8 performance.

AMD Unveils Threadripper Halo Station with 96 Cores for Local AI Development — Photo: ServeTheHome

The base configuration presented by AMD features two MI350P cards, combining for a total of 288 GB of GPU memory. The company noted that customers have an upgrade path to install up to four cards, bringing total accelerator memory capacity to 576 GB. That footprint is specifically sized to hold a trillion-parameter AI model entirely on-device.

Thermal management required a distinct engineering approach. Standard MI350P accelerators are built for standardized servers with forced air cooling and thermal design power ratings exceeding 450 W. Because desktop tower workstations lack server-grade airflow, AMD equipped the Halo Station with independent closed-loop liquid cooling for every major processor, including individual liquid-cooled loops for each accelerator card and the central CPU.

Networking Upgrades and Open Questions

The development system also resolves connectivity limitations found in smaller form factor devices. Reviewers noted that previous mini-PC developer boxes lacked high-speed networking options beyond a 10Gb Ethernet port. The IFA showcase unit featured an unidentified SFP networking card, providing high-throughput data ingress capabilities necessary for feeding massive datasets into local machine learning pipelines.

Despite the hardware preview, AMD has not yet disclosed complete system specifications, pricing, availability dates, or the software stack that will accompany the developer station.