On-PREM<\/p>\n
AI workstation promises to put up to 576 GB of HBM3e and 16 TB\/s of memory bandwidth on your desk<\/p>\n
Machine learning researchers have a new and expensive option for doing local AI. AMD just raided its spare parts bin and cobbled together a DGX Station rival boasting up to 576 GB of HBM3e memory that it\u2019s calling the Threadripper Halo.<\/p>\n
Announced at IFA 2026, the system, which is set to launch next year, marks the first time that AMD\u2019s Instinct accelerators have been offered in a workstation form factor.<\/p>\n
But much like the AI Halo we reviewed<\/a> earlier this year, the Threadripper Halo<\/a> isn\u2019t packing anything we haven\u2019t seen before. The liquid-cooled system is powered by a 96-core Threadripper PRO 9995WX, which you may recall made its debut<\/a> in 2025. The CPU is backed by up to 2 TB of DDR5 memory and up to 2.6 TB of combined system memory.<\/p>\n As you probably already guessed, AMD is positioning the box as a local AI workstation for researchers. From what we gather, the system can be had with up to four of the company\u2019s PCIe-based Instinct GPUs \u2014 though the systems displayed at IFA only showed two installed.<\/p>\n Launched back in May, the MI350P<\/a> is essentially half of an MI350X that\u2019s been crammed into a PCIe form factor. Each accelerator packs 144 GB of speedy HBM3e good for 4 TB\/s of memory bandwidth, and up to 4.6 petaFLOPS of FP4 compute. But we\u2019ll note that it’s in its 600 W configuration. When deployed in the Threadripper Halo, we expect AMD will be running the cards in the more sedate 450 W config.<\/p>\n Combined, the Threadripper Halo promises up to 576 GB of HBM3e and 16 TB\/s of memory bandwidth, enabling the system to run models exceeding a trillion parameters in size (at four-bit precision) entirely in GPU memory. By offloading some of the model to system memory, the system should be able to run the largest open weights models, like Moonshot.AI\u2019s 2.8 trillion-parameter Kimi K3<\/a>.<\/p>\n The ability to run large frontier-class models on your desk won\u2019t come cheap and may not be available in every market, at least not without an electrical service upgrade. AMD hasn\u2019t shared pricing, but as configured, we expect it to sell for somewhere between $100,000 and $150,000.<\/p>\n A quad MI350P system is going to push the limits of a standard North American power outlet unless AMD either underclocks the cards from 600 W to 300 W or it mandates a 20 amp circuit. This might explain why the system AMD showed off at IFA only had two.\u00a0<\/p>\n At this price and power level, the system is well positioned to compete with Nvidia\u2019s own high-end AI workstation, the DGX Station. Announced at GTC in 2025, the DGX Station<\/a> features a 252 GB B300 GPU and a 72-core Grace CPU along with 496 GB of LPDDR5x memory, and it retails for around $100,000<\/a> if you can find one in stock.<\/p>\n The House of Zen boasts that the system offers up to 3.4x the total system memory and more than twice the memory bandwidth of the DGX Station, which should give an edge in LLM inference, assuming tensor parallel operations don\u2019t bottleneck on the CPU\u2019s PCIe bus.<\/p>\n AMD says the Threadripper Halo will be available starting next year. \u00ae<\/p>\n","protected":false},"excerpt":{"rendered":"On-PREM AI workstation promises to put up to 576 GB of HBM3e and 16 TB\/s of memory bandwidth…\n","protected":false},"author":2,"featured_media":760285,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,733,4308,86,56,54,55],"class_list":["post-760284","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-technology","tag-uk","tag-united-kingdom","tag-unitedkingdom"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/760284","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/comments?post=760284"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/760284\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/760285"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=760284"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=760284"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=760284"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}