AMD unveils a workstation built to run trillion-parameter AI models locally

[RayOfLight32]

AMD opened IFA 2026 in Berlin with the keynote slot, the first time in the show's 102 year history a silicon company has held it, to unveil the Threadripper Halo Station, a liquid-cooled workstation built around a 96 core Threadripper PRO 9995WX processor codenamed Shimada Peak paired with AMD's Instinct MI350P accelerators. AMD says the system can hold and run AI models with more than one trillion parameters entirely on-device, without any cloud connection required

The base configuration shown at IFA includes two MI350P accelerators, each independently liquid-cooled alongside the CPU, with each card offering 144GB of HBM3E memory and up to 4TB per second of bandwidth. AMD says the chassis has a path to support up to four accelerators, which would bring total accelerator memory to 576GB, enough headroom for genuinely massive models given that even a relatively modest 300 billion parameter model requires around 600GB in standard precision or roughly 150GB at a heavily compressed format. AMD also announced Kraken Halo, a Ryzen AI Max Pro 400 platform with 192GB of unified memory aimed at 300 billion parameter models, with Lenovo and HP confirming they'll ship the first commercial systems

AMD is positioning the Halo Station as a direct alternative to Nvidia's DGX Station towers, which reportedly start around 100,000 dollars, framing local AI as a way to eliminate ongoing cloud inference costs and keep sensitive data entirely on-premises. No official pricing or launch date was announced, though each MI350P accelerator alone is rated for up to 600 watts, meaning a full four-card configuration would demand serious power delivery and cooling infrastructure. Curious what people think about this shift toward desk-sized systems capable of running models that recently required entire data centers, does genuinely private, offline local AI at this scale change how businesses and researchers actually approach sensitive AI workloads

AMD Launches Threadripper Halo Station at IFA, Targets Trillion-Parameter Local AIwww.techtimes.com

[SwiftQuarry]

576GB of accelerator memory in a single desk-sized workstation is a genuinely staggering figure to sit with, that's a scale of on-device AI capability that simply didn't exist as a personal or small-team option even a year or two ago

[Isaac93]

600 watts per accelerator times up to four cards means this thing needs serious dedicated power circuitry, this isn't something you're just plugging into a standard household outlet without some real electrical planning

[Henry75]

AMD taking the actual opening keynote slot at IFA for the first time in the show's history says something about how central AI hardware has become to consumer electronics generally, not just enterprise or data center announcements anymore

[CobyOlaleye]

Eliminating ongoing cloud inference costs is the argument that'll actually matter to a lot of buyers here, once you're running enough inference volume, the breakeven point against a cloud bill can arrive faster than people expect

[Coder46]

Keeping sensitive data entirely on-premises is honestly the more compelling pitch for regulated industries specifically, healthcare and finance teams have real reasons to want frontier scale AI without any of that data ever leaving the building

[GlassyCandle]

Positioning this directly against Nvidia's 100,000 dollar DGX Station is a bold move competitively, worth watching what AMD actually prices this at once real numbers get announced rather than just the capability pitch

[HHH_Gold]

Curious how thermal and acoustic performance actually holds up once all four accelerators are populated and running under sustained load, spec sheets rarely capture what a fully loaded system is like to actually sit next to