AMD Unveils High-End Workstation Prototype Amid AI Hardware Race
BERLIN — AMD debuted the Threadripper Halo Station at the IFA trade show Sept. 4, signaling a strategic push into the high-end AI workstation market. The liquid-cooled system features a 96-core Ryzen Threadripper PRO 9995WX and is designed to compete with top-tier hardware from Apple and Nvidia.
The prototype’s most striking specification is its capacity for massive memory expansion. While the unit demonstrated at the show utilized two Instinct MI350P accelerators providing 288GB of HBM3E memory, AMD stated the architecture supports up to four cards. A full four-card configuration would provide 576GB of HBM3E, directly challenging the 512GB capacity of Apple’s upcoming M5 Ultra Mac Studio.
Industry analysts note that the 576GB threshold is a critical benchmark for local AI development, as a one-trillion-parameter model at four bits requires approximately 500GB of raw data. However, the distinction between theoretical capacity and available hardware remains sharp. AMD has listed the Halo Station as a prototype with a projected 2027 release date, offering no official pricing or confirmed OEM partners.
In contrast, Apple’s 512GB Mac Studio is scheduled for release in late October, though pricing has not been disclosed. Currently, the highest available unified-memory configuration for Apple users is 256GB, retailing for $10,799.
Despite the impressive specifications, the Halo Station faces significant cost hurdles. Based on September 4 street prices, a rough bill of materials suggests a 96-core Threadripper PRO 9995WX costs between $11,000 and $12,000, while two MI350P accelerators cost approximately $40,000. The most significant expense is the 2TB of DDR5 RDIMM memory, estimated at $50,000—surpassing the cost of the processor itself.
Technical experts caution that the “576GB” figure often cited in headlines describes a theoretical maximum rather than the demonstrated system. Furthermore, the combined memory pool of HBM and DDR5 operates across different address spaces and speeds, meaning the total aggregate memory may not function as a single unified pool in the way Apple’s silicon architecture does.
The unveiling underscores a broader national and global trend toward “edge AI,” where corporations and researchers seek to run massive large language models (LLMs) on local hardware rather than relying exclusively on cloud-based clusters.
