AMD used IFA 2026 to unveil one of its most ambitious workstation concepts yet: the Halo Station, a machine built around some of the company's most powerful professional hardware. AMD is describing it as "the most powerful workstation in the world," and while that is a bold claim, the specification list certainly shows that the system is aimed far beyond ordinary desktop use.
At the centre of the Halo Station is the Ryzen Threadripper Pro 9995WX, paired with liquid-cooled Instinct MI350P accelerators, enormous amounts of system memory and enough high-bandwidth GPU memory to tackle extremely large AI workloads locally. This is not a gaming PC with workstation branding. It is a machine designed for AI development, scientific computing, professional visualisation and other workloads where even high-end desktops quickly run out of resources.
A 96-Core Threadripper Pro Sits at the Heart of the System
The Halo Station uses AMD's Ryzen Threadripper Pro 9995WX, a 96-core, 192-thread processor based on the Zen 5 architecture. The CPU runs at a 2.5GHz base clock and can boost up to 5.4GHz, while carrying a huge 384MB of L3 cache.
Its 350W TDP already puts it firmly into workstation territory, and the processor uses AMD's sTR5 platform with support for WRX90, TRX50 and Pro 695 motherboards. The chip itself is not new, but placing it inside a workstation designed around massive AI accelerators makes its role much more interesting.
Threadripper Pro processors are built for workloads that scale across large numbers of CPU cores, including 3D rendering, simulation, software compilation, engineering and content production. In the Halo Station, however, the CPU is only one part of a much larger compute platform.
Two Instinct MI350P Accelerators Come as Standard
The more dramatic part of the configuration is AMD's use of two liquid-cooled Instinct MI350P accelerators, with the system designed to support as many as four.
These are not conventional gaming GPUs. AMD's Instinct family is built specifically for data centres, AI training, scientific computing and high-performance workloads. The MI350P is based on the company's CDNA4 architecture and manufactured using TSMC's 3nm-class process.
Each accelerator contains 8,192 stream processors, 512 matrix cores and 129 compute units, with a peak engine clock of around 2.2GHz. The focus is raw parallel compute rather than gaming features such as ray tracing or consumer graphics performance.
The Memory Configuration Is Enormous
AMD says the Halo Station can be equipped with 2TB of DDR5 system memory, which already puts it far beyond even most high-end professional workstations.
The GPU side is equally extreme. A dual-accelerator configuration provides 288GB of HBM3E memory, while a fully configured four-accelerator system can reach 576GB.
That amount of high-bandwidth memory is particularly important for artificial intelligence workloads. Large language models often require enormous amounts of accelerator memory, especially when developers want to run them locally without splitting workloads across multiple external servers.
AMD claims the Halo Station can hold a trillion-parameter model within the system. Whether such a model could be run efficiently depends on factors such as precision, quantisation and workload configuration, but the claim illustrates the scale AMD is targeting.
This Is Clearly an AI Workstation First
The Halo Station is best understood as a local AI and high-performance computing platform rather than a traditional workstation. Its hardware configuration is designed for organisations that would normally consider rack-mounted servers or cloud GPU infrastructure.
For AI developers, having hundreds of gigabytes of HBM available locally could change the workflow significantly. Instead of constantly sending models and datasets to cloud services, teams could potentially run very large inference or development workloads inside their own environment.
That could be particularly attractive for organisations dealing with sensitive data, research workloads or proprietary models that they would prefer not to send outside their own infrastructure.
The Power Requirements Are Just as Extreme
The performance comes with an equally serious appetite for electricity. Each liquid-cooled Instinct MI350P can draw up to 600W, which means the two-accelerator configuration alone can consume as much as 1,200W before accounting for the CPU and the rest of the system.
Add the 350W Threadripper Pro processor, memory, storage, cooling and other components, and the Halo Station becomes a machine that needs substantial power delivery and thermal management.
A four-accelerator configuration would push those requirements considerably higher. This is not something designed to sit quietly under a typical office desk using a standard consumer power setup.
The liquid cooling is therefore not simply for aesthetics. It is essential for keeping this level of hardware operating reliably under sustained professional workloads.
Why AMD Is Bringing Server-Class Hardware to a Workstation
One of the more interesting aspects of the Halo Station is how much it blurs the line between workstation and server. Traditionally, extremely large AI models and GPU-heavy scientific workloads would be handled by dedicated server infrastructure.
The Halo Station attempts to bring a similar level of compute capability into a single workstation chassis. That could make powerful AI development more accessible to individual research teams, engineering departments or studios that need enormous local compute without building a full data-centre environment.
It also reflects the broader shift toward local AI. As models become more capable and hardware becomes more powerful, some organisations are looking for ways to reduce their dependency on cloud inference and training.
The Threadripper Pro Is Not Intended for Gaming
Despite the impressive specifications, AMD's Ryzen Threadripper Pro 9995WX is not designed primarily for gaming. Most modern games cannot take advantage of anywhere near 96 CPU cores, and a conventional Ryzen gaming processor would generally provide better value and efficiency for that purpose.
The Halo Station instead targets professional workloads that can scale across large amounts of CPU and GPU compute. Rendering, simulation, visual effects, AI development and scientific computing are far more appropriate workloads for this class of system.
The Instinct accelerators reinforce that point. These are compute-focused products built around matrix processing and large-scale parallel workloads rather than the features expected from Radeon gaming GPUs.
The Price Could Easily Reach Supercar Territory
AMD has not announced official pricing for the Halo Station, but the components already give a rough idea of how expensive the system could become.
Based on current market pricing for the Threadripper Pro processor, Instinct accelerators, massive DDR5 capacity and supporting workstation hardware, a base configuration could potentially cost more than RM400,000.
A fully loaded version with four Instinct accelerators and the maximum memory configuration could push the price closer to RM600,000 or beyond.
That makes it obvious who the target customer is. This is not a machine intended for enthusiasts looking for the fastest desktop possible. It is designed for businesses and research organisations where the hardware can justify its cost through productivity, model development or reduced reliance on external compute infrastructure.
Local AI Could Be the Biggest Selling Point
The strongest argument for a workstation like this may be the ability to keep large AI workloads entirely on-premise. Cloud GPU infrastructure is powerful, but large-scale usage can become extremely expensive over time.
For companies running AI inference continuously or handling sensitive proprietary models, investing in local hardware may eventually make financial or operational sense.
Local processing can also provide lower latency and greater control over data. That is especially important for industries such as healthcare, finance, engineering and research, where sending sensitive information to external services may introduce additional compliance or security concerns.
The Halo Station is essentially AMD's answer to the question: What if you want data-centre-class AI compute sitting next to your desk?
Final Thoughts
AMD's Halo Station is an extreme workstation built for an equally extreme class of workload. With a 96-core Ryzen Threadripper Pro 9995WX, dual Instinct MI350P accelerators, up to 2TB of DDR5 memory and as much as 576GB of HBM3E in a fully configured system, it sits much closer to a compact AI server than a conventional professional desktop.
Its ability to potentially accommodate trillion-parameter AI models makes it particularly interesting as organisations look for ways to perform more AI development and inference locally. The trade-off is equally obvious: enormous power consumption, complex cooling requirements and a likely price well into six figures.
For ordinary users, the Halo Station is complete overkill. For AI researchers, engineering teams and studios that routinely consume large amounts of compute, however, it could represent something much more useful: server-class performance without having to leave the workstation environment.
Comments 0