AMD unveiled the Threadripper Halo Station during its IFA 2026 keynote in Berlin, introducing a liquid-cooled developer tower pairing a 96-core Zen 5 processor with Instinct MI350P accelerators to run trillion-parameter AI models locally without cloud connections. Jack Huynh, AMD’s Senior VP and GM of their Computing and Graphics system, announced the Threadripper Halo Station as part of an AI-centric opening keynote for this year’s IFA tech convention in Berlin, describing it as a new class of workstation bringing supercomputer-class compute to individual users and developers.
Inside the 96-Core Zen 5 Architecture and Memory Specs
At the heart of AMD’s newly announced developer system is the flagship Ryzen Threadripper PRO 9995WX processor, a 96-core, 192-thread Zen 5 chip codenamed “Shimada Peak” as reported during the opening keynote. The workstation class processor operates with a boost frequency of up to 5.4 GHz, features 384 MB of L3 cache, and carries a thermal design power of 350W according to detailed hardware specifications. It is a 96 core Zen 5 chip with 8 channels of DDR5 memory and 128 lanes of PCIe Gen5 connectivity. Standalone chips for the 9995WX clock in at around $11,000 to $12,000.
Jack Huynh, AMD’s Senior VP and GM of Computing and Graphics, introduced the tower in Berlin as intended to augment AMD’s existing Ryzen AI Halo box, which was their initial Halo system based around the company’s Ryzen AI Max+ (Strix Halo) processors. In the case of rummaging through the hardware, the Threadripper Halo Station is essentially AMD going for the other end of the market, with a high-performance system that can offer far better performance than what the petite Ryzen AI Halo boxes can offer.
Liquid-Cooled Instinct MI350P Accelerators and Expansion Path
While the CPU manages system coordination, the heavy lifting in terms of AI workloads is provided by AMD’s Instinct MI350P accelerator cards. Launched a bit earlier this year, the MI350P is a cut-down version of AMD’s CDNA 4 architecture MI350X accelerator that is designed to fit in a PCIe form factor, featuring 128 CDNA 4 compute units built on TSMC N3. Each one of these cards comes with 144GB of HBM3E memory running at 4 TB/s, offering a bandwidth 14 times higher than any LPDDR5X variant. Each accelerator is rated for a Total Board Power (TBP) of up to 600 W.
At IFA 2026, AMD presented a configuration based on two MI350P PCIe cards for a total of 288 GB of HBM3E memory. However, AMD’s goal is to support up to four MI350P accelerators in the Halo Station, and AMD says there’s a “path to four,” opening up the possibility of two more accelerators bringing 576GB of HBM3E to the system. This amount of accelerator memory is enough, says AMD, to hold a trillion-parameter model entirely on-device. Individual MI350P accelerators are not sold by AMD on their own in traditional consumer channels, but the estimated price is somewhere around $20,000 per accelerator.
Thermal Engineering and Standalone Component Pricing Realities
Although AMD says it can support up to four accelerators, the workstation shown off at IFA only has room for two, both of which are liquid-cooled independently, as is the CPU. You’ll need plenty of power to feed the GPUs, as each accelerator is rated for up to 600W.
It’s essentially a server tray reconfigured into a tower, with an EPYC host replaced with a 96-core Threadripper. AMD didn’t share many details about the machine outside of the specs, though it appears to be a system design that AMD’s OEM partners will ultimately build and ship. Taking all of the components together, the street price should come out to over $100,000 with just the core components: memory, CPU, and dual GPUs. Configured higher, and with supporting storage, power, and cooling, the workstation could very easily climb over $150,000. AMD has yet to announce any partners supporting the machine, and complete system specifications, availability and pricing have not been announced.
Ecosystem Integration and Partner Outlook
The forthcoming dev system is a full tower system based on a pairing of a Threadripper Pro CPU and Instinct MI350P accelerators, with the purpose of offering a high-end, high-throughput system for local AI inference and development. Jack Huynh explicitly mentioned that the system on display featured a 96-core processor (and with no qualification of “up to”), which would be AMD’s flagship workstation Ryzen Threadripper PRO 9995WX processor.
While AMD showcased the initial tower design in Berlin, AMD has not yet announced any partners supporting the machine. Prospective buyers and enterprise developers await partner announcements and finalized configuration pricing from AMD’s hardware ecosystem in the coming months.
