Coming in 2027

AMD Ryzen™ Threadripper™ PRO, the world's best workstation processor, paired with AMD Instinct™ HBM accelerators.¹

Power Frontier-Class Models. Run Hundreds of Agents. Locally.

Whether you're running the most complex models available today—and tomorrow—or driving hundreds of agentic workflows, the AMD Threadripper™ Halo Station is built for it. With up to 2.6TB of combined memory and up to 16.4TB/s of total system memory bandwidth, it keeps frontier-class models loaded and responsive. Run your agents for as long as they need and keep what you'd otherwise spend on cloud inference.

Up to
2.6TB
Total system memory
Up to
3.4x
Total system memory (RDIMM and HBM3e aggregate) vs. Nvidia DGX Station²
Up to
16.4TB/s
Total system memory bandwidth

Specifications

What's Under the Hood

AMD Threadripper Halo Station with AMD Ryzen Threadripper PRO 9995WX processor and AMD Instinct class HBM workstation accelerators.

AMD Threadripper™ Halo Station
AMD Threadripper™ Halo Station
AMD
“Zen 5“
Architecture
96
Cores
192
Threads
Up to
576GB
HBM3e GPU Memory
Up to
2TB
RDIMM System Memory
Up to
2.6TB
Total System Memory
Up to
16TB/s
Aggregate GPU Memory Bandwidth
Up to
410GB/s
RDIMM Memory Bandwidth
Up to
16.4TB/s
Total System Memory Bandwidth

Frequently Asked Questions

AMD Threadripper™ Halo Station is a prototype system AMD is showing for the first time at IFA 2026 that brings datacenter-class AI compute into a single deskside workstation. It pairs an AMD Ryzen Threadripper PRO processor with AMD Instinct-class accelerators, delivering the memory capacity and bandwidth needed to train, fine-tune and run large models locally as well as run intensive agentic workflows—without a server room, a cloud queue or shared tenancy.

AMD Instinct accelerators have historically been built for data centers—rack-mounted, server-form-factor hardware. Today's AI workflows push memory and bandwidth demands past what a typical workstation GPU can handle, and those demands add up quickly on a cloud inference bill if you're running them all day. The AMD Threadripper™ Halo Station concept is designed to meet that shift head-on: frontier-class model capacity and the throughput to run it, locally, on a single desk, right as AI itself moves from single-prompt use to continuously running agents. That shift requires real advances in system design and integration—and AMD Threadripper Halo Station is proof of what's now possible. 

Individuals and small teams doing serious AI work who are currently constrained by cloud access or shared infrastructure—AI researchers, model developers, engineering and simulation users, and creators working with large generative models. Also organizations that need to keep sensitive data and IP on-premises and under their own control.

Footnotes

© 2026 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, AMD ROCm, Radeon, Ryzen, and combinations thereof are trademarks of Advanced Micro Devices, Inc.

  1. V-Ray Benchmark 6 (CPU, ksamples), Cinebench 2026 (nT), SPECapc Autodesk Maya 2024, SPECapc PTC Creo 9, SPECapc Solidworks 2025, SPECviewperf, Premiere Pro 26.3, DaVinci Resolve Studio 21, Keyshot Viewer 2026.1 benchmark, Corona Render (Rays/sec) benchmark 10, Autodesk Revit 2026, Chromium Compilation 151.0.7891.0, Unreal Engine 5.7.4 Compilation, and MATLAB R2026a benchmarks to compare the performance of the AMD Ryzen Threadripper PRO 9995WX processor configured with 8x 64GB DDR5 memory, Nvidia RTX PRO 6000 Blackwell graphics, 1TB SSD, Win 11 vs. a similarly configured BOXX workstation with the Intel® Xeon® 698X processor. Workstation manufacturers may vary configurations, yielding different results. Results may vary. SHP-76.
  2. Comparison as of August 2026 based on published manufacturer specifications. AMD Threadripper Halo Station configuration: 1x AMD Ryzen Threadripper PRO 9995WX processor (8-channel DDR5 RDIMM memory controller, up to 6400 MT/s) with 2TB RDIMM system memory; 4x AMD Instinct MI350P accelerators (144GB HBM3e memory each, 4 TB/s peak memory bandwidth each, 4096-bit memory interface). Compared to Nvidia DGX Station configuration: 1x NVIDIA GB300 Grace Blackwell Ultra Superchip with 252GB HBM3e GPU memory (7.1 TB/s peak bandwidth) and 496GB LPDDR5X CPU memory (396 GB/s peak bandwidth). Specs references: https://www.nvidia.com/en-us/products/workstations/dgx-station/; https://www.amd.com/en/products/processors/workstations/ryzen-threadripper/9000-wx-series/amd-ryzen-threadripper-pro-9995wx.html; https://www.amd.com/en/products/accelerators/instinct/mi350/mi350p.html. Workstation system manufacturers may vary configurations yielding different results. HALO-02.