MPN: Threadripper Halo Station
AMD Threadripper Halo Station
96 Cores · 576 GB HBM3e · 16 TB/s

Datacenter-class AI compute in a single deskside tower · Ryzen Threadripper PRO 9995WX + up to 4× Instinct MI350P · trillion-parameter models fully resident in GPU memory

Threadripper Halo Station

A deskside tower that AMD describes as bringing datacenter-class AI compute into a single system.

Instinct MI350P accelerator

The workstation-class Instinct card that makes the Halo Station possible.

Local trillion-parameter inference

The workload the Halo Station is engineered for.

Specifications

Vendor

AMD

Product

AMD Threadripper Halo Station

Type

Deskside AI developer workstation

CPU

AMD Ryzen Threadripper PRO 9995WX
Zen 5, 96 cores / 192 threads

CPU Memory Channel

8-channel DDR5 RDIMM
up to 6,400 MT/s

System Memory

Up to 2 TB DDR5 RDIMM

Accelerators

2× or 4× AMD Instinct MI350P

GPU Memory

144 GB HBM3e per card

Total GPU Memory

Up to 576 GB HBM3e

GPU Memory Bandwidth

4 TB/s peak per card
16 TB/s aggregate

Memory Interface

4,096-bit per card

Model Capability

Trillion-parameter models at 4-bit
fully resident in GPU memory

Form Factor

Liquid-cooled tower workstation

Announced

IFA 2026, Berlin — 4 September 2026

Status

Prototype system
Launch expected 2027

Comparable System

NVIDIA DGX Station (252 GB HBM3e + 496 GB LPDDR5X)

Price

Quote

Overview

The AMD Threadripper Halo Station is a deskside AI workstation announced by AMD at IFA 2026 in Berlin — the company’s second Halo-branded developer system, sitting well above the Ryzen AI Halo box. Where that machine is built around a Ryzen AI Max+ (Strix Halo) processor, the Threadripper Halo Station pairs a Ryzen Threadripper PRO 9995WX with discrete AMD Instinct MI350P accelerators — the first time AMD has offered Instinct silicon in a workstation form factor.

The configuration is deliberately datacenter-grade: a Zen 5 96-core / 192-thread Threadripper PRO 9995WX with an 8-channel DDR5 RDIMM controller running up to 6,400 MT/s and up to 2 TB of system memory, alongside two or four Instinct MI350P accelerators at 144 GB of HBM3e each on a 4,096-bit interface with 4 TB/s of peak bandwidth per card. Fully populated, that is up to 576 GB of HBM3e and 16 TB/s of aggregate GPU memory bandwidth in a liquid-cooled tower — enough to hold a trillion-parameter model at four-bit precision entirely in GPU memory, and to run larger open-weight models such as 2.8-trillion-parameter systems when part of the model is offloaded to system memory.

AMD positions the system for researchers and developers who are currently queueing for shared cloud GPU capacity, explicitly framing it as a DGX Station competitor. The company is showing it as a prototype; availability is expected in 2027. Contact QS Compute for configuration and lead-time information as it approaches release.

Need AMD Threadripper Halo Station?

Contact QS Compute for availability, configuration, and volume pricing.

Request Quote