Archive · 5 Sep 2026
AMD Unveils Halo Station With 576GB HBM3E for Local AI
AMD announced a very powerful computer called Threadripper Halo Station.
It is designed to sit beside a person instead of inside a large data center.
The machine can use up to four special AI chips.
Together, those chips can provide 576GB of very fast memory.
AMD says this could let people run AI models with more than one trillion parameters without using the cloud.
This may help groups handling private medical, defense, or financial information.
It could also help researchers avoid waiting for cloud computers.
However, AMD has not said how much the machine will cost or when it will be available.
AMD announced Threadripper Halo Station at its IFA 2026 keynote as a high-end AI workstation.
The system uses a 96-core Threadripper PRO 9995WX processor and supports up to 2TB of DDR5 memory.
It can include up to four Instinct MI350P accelerators with 144GB of HBM3E each, totaling 576GB.
AMD says the workstation is designed to run AI models exceeding one trillion parameters locally without cloud access.
AMD has not disclosed pricing, availability, or complete specifications, leaving its economic case untested.
- Who
- Advanced Micro Devices, Inc. (AMD) announced the workstation.
- What
- AMD announced Threadripper Halo Station, a desk-side workstation supporting up to four Instinct MI350P accelerators and 576GB of HBM3E.
- When
- At AMD's IFA 2026 keynote.
- Where
- As a desk-side workstation intended for local operation; it was announced during the IFA 2026 keynote.
- Why
- To run extremely large AI models locally, support privacy- or jurisdiction-sensitive workloads, and offer an alternative to cloud computing and NVIDIA's DGX Station.
AMD's Hardware Case
Practical Reservations
Local AI performance
AMD's Hardware Case
AMD presents the workstation's large HBM3E capacity and high memory bandwidth as enabling trillion-parameter AI models to run locally at usable speeds.
Practical Reservations
The article notes that loading a model is not enough; practical performance will also depend heavily on software.
Alternative to NVIDIA
AMD's Hardware Case
AMD is positioning Halo Station against NVIDIA's DGX Station and argues that supercomputer-class computing could be available to individual developers and teams.
Practical Reservations
NVIDIA's software ecosystem, particularly CUDA, remains a major advantage, so comparable hardware may not be enough to attract many teams.
Ownership versus cloud rental
AMD's Hardware Case
A locally owned system could benefit organizations with private data, cloud scheduling bottlenecks, or consistently high inference usage.
Practical Reservations
Without pricing and availability, it is impossible to determine whether ownership will be cheaper than cloud subscriptions; the fully configured system is expected to be expensive.
- Processor
- 96-core, 192-thread Threadripper PRO 9995WX based on Zen 5, with boost speeds up to 5.4GHz.
- System memory
- Supports up to 2TB of DDR5 memory.
- AI accelerators
- Supports up to four Instinct MI350P accelerators.
- GPU memory
- Each accelerator has 144GB of HBM3E; four cards provide 576GB total.
- Presented configuration
- AMD showed a two-card configuration providing 288GB of HBM3E.
- Intended workload
- Local AI inference for models exceeding one trillion parameters.
- Unreleased details
- AMD has not announced pricing, availability, or complete system specifications.