AMD unveils Threadripper Halo AI workstation, aiming to rival Nvidia's DGX Station

The Threadripper Halo is built around a 96-core Threadripper PRO 9995WX CPU, backed by up to 2 TB of DDR5 memory, and can be configured with up to four of AMD's PCIe-based Instinct MI350P GPUs. Each MI350P accelerator packs 144 GB of HBM3e memory, providing up to 4 TB/s of memory bandwidth per card. Combined, the system promises up to 576 GB of HBM3e and 16 TB/s of memory bandwidth, enabling it to run models exceeding a trillion parameters (at four-bit precision) entirely in GPU memory. By offloading some of the model to system memory, the workstation should be able to handle the largest open-weights models, such as Moonshot.AI's 2.8 trillion-parameter Kimi K3.

The liquid-cooled system comes at a significant power cost. In its maximum 600 W configuration per GPU, a quad-MI350P system would push the limits of a standard North American power outlet. AMD is expected to run the cards at a more sedate 450 W — or potentially 300 W — which may explain why the demonstration unit at IFA showed only two GPUs installed. AMD has not shared pricing, but industry estimates place the full system between $100,000 and $150,000.

The Threadripper Halo is directly positioned against Nvidia's DGX Station, a high-end AI workstation announced at GTC in 2025 that features a 252 GB B300 GPU and a 72-core Grace CPU, retailing for around $100,000. AMD claims its system offers up to 3.4 times the total system memory and more than twice the memory bandwidth of the DGX Station, potentially giving an edge in large language model inference — provided tensor parallel operations do not bottleneck on the CPU's PCIe bus.

AMD says the Threadripper Halo will be available starting next year. The system's high power demands and price point mean it will likely find a niche among well-funded research labs and enterprises seeking local AI computation without relying on cloud infrastructure.

Why This Matters

The Threadripper Halo brings enterprise-grade AI acceleration to a local workstation, reducing reliance on cloud-based inference and data transfer.
It intensifies competition between AMD and Nvidia in the high-performance AI hardware market, potentially driving down costs or increasing performance options for researchers.
The ability to run trillion-parameter models locally could accelerate research that requires handling sensitive data or low-latency inference.

Background

The workstation AI market has been dominated by Nvidia's DGX Station, which was announced in 2025 and offers a B300 GPU with 252 GB of memory. AMD's entry leverages its existing chip designs — the Threadripper PRO 9995WX CPU launched in 2025, and the MI350P accelerator debuted in May 2026 as a PCIe version of its larger MI350X. By combining these components, AMD aims to offer a system with vastly more memory and bandwidth than Nvidia's offering, though at a higher power consumption and likely similar cost.

Key Perspectives

Machine learning researchers: Gain access to a powerful local system that can handle the largest open-weights models without cloud dependencies. However, the $100,000–$150,000 price tag and power requirements limit accessibility to well-funded institutions. AMD: Positions itself as a serious contender in the AI workstation space, leveraging its Instinct accelerator line and existing CPU portfolio. The company claims significant memory and bandwidth advantages over Nvidia's incumbent DGX Station. Critics/Skeptics: The system's power demands may require electrical service upgrades, and the high price could limit adoption. The performance advantage claimed over the DGX Station depends on efficient tensor parallel scaling across four GPUs, which may introduce PCIe bus bottlenecks.

What to Watch

Actual system pricing and availability when launched next year.
Power configuration details — whether AMD will offer a 300 W, 450 W, or full 600 W mode for the GPUs.
Performance benchmarks comparing the Threadripper Halo to Nvidia's DGX Station in real-world LLM inference and training tasks.
Market reception and whether Nvidia responds with a spec upgrade or price adjustment.

ZOTPAPER