AMD's Threadripper Halo Station Packs 96 Cores and Two MI350P Accelerators Into a Liquid-Cooled Desktop

At IFA 2026 in Berlin, AMD on September 4 pulled the wraps off a machine that stretches the phrase "personal computer" close to its breaking point. The Threadripper Halo Station is a fully liquid-cooled AI workstation pairing a 96-core Ryzen Threadripper PRO host with two Instinct MI350P data center accelerators, pitched as a deskside system for running local AI models with more than one trillion parameters. The company called it "the most powerful workstation in the world," a superlative delivered during the opening keynote of a show where consumer and local-first AI hardware is a defining theme.

The system shown in Berlin is best understood as a server tray stood up inside a tower chassis. Rather than a conventional desktop graphics card, the Halo Station uses the same Instinct silicon AMD sells into cloud and enterprise data centers, bolted onto its workstation CPU line. AMD said on September 4 that the design is a prototype, and it declined to name a price, a sales date, or any OEM partners during the presentation. What it did describe is a concrete roadmap: two accelerators today, with a planned path to four.

Key Facts

At the core of the Halo Station sits the Ryzen Threadripper PRO 9995WX, the only current PRO processor with 96 Zen 5 cores and 192 threads. It boosts to 5.4 GHz, carries 384 MB of L3 cache, and is rated at a 350 W TDP. The chip exposes eight DDR5 memory channels and 128 usable PCIe 5.0 lanes, and the workstation takes the platform to its ceiling with up to 2 TB of DDR5 system memory. Igor's Lab reported on September 5 that the core count, memory channels, and PCIe lane budget line up precisely with the 9995WX, making the host chip easy to identify even though AMD mostly called it a 96-core Threadripper PRO.

The AI heavy lifting falls to two Instinct MI350P accelerators built on the CDNA 4 architecture. Each card packs 128 compute units fabricated on TSMC N3, 144 GB of HBM3E, and up to 4 TB/s of memory bandwidth, giving the demonstrated configuration a combined 288 GB of high-speed accelerator memory. Tom's Hardware reported on September 4 that each accelerator is rated for a board power of up to 600 W, so the two GPUs alone can draw 1,200 W before the CPU, memory, and storage are counted. Every hot component, CPU and accelerators alike, is liquid cooled, which is why AMD has largely abandoned conventional workstation fans here.

The more striking number is the one AMD says is coming. The company described a "path to four" accelerators, a configuration that would hold 576 GB of HBM3E in total. For context on what all that memory is for, Igor's Lab estimated on September 5 that a dense model with one trillion parameters needs roughly 2 TB of memory for weights alone at 16-bit precision, about 1 TB at 8-bit precision, and close to 500 GB at 4-bit precision, before any working data is accounted for. Under that arithmetic, a four-GPU Halo Station could keep a quantized trillion-parameter model substantially inside fast accelerator memory, with the 2 TB of DDR5 as overflow.

Analysis

The biggest question hanging over the Halo Station is whether local trillion-parameter inference is a genuine workload or a marketing target. The bigger picture here is that AMD is not really selling a desktop at all; it is selling a miniature data center that happens to fit beside a desk. Tom's Hardware reported on September 4 that the machine is essentially a server tray reconfigured into a tower, and the power figures reinforce that reading. A four-GPU build could push the accelerators alone toward 2.4 kW before the host CPU and memory are counted, a load that strains the 16-amp circuits common in European offices. By any ordinary definition, that is not a consumer machine.

It is also worth separating the memory story from the compute story. A dense trillion-parameter model cannot fit in 576 GB of HBM3E at 16-bit precision, so the claim that it can run such models depends heavily on quantization and on how weights are spread between accelerator memory and the 2 TB of DDR5. Igor's Lab noted on September 5 that with only the two accelerators present today, 288 GB of HBM3E forces frequent partitioning of the largest models, which is precisely why the planned four-GPU configuration matters. What this really means is that the Halo Station, at least in its demonstrated two-GPU form, is less a finished answer than a proof of what AMD's workstation platform can hold, with the four-GPU version doing most of the work that the marketing promises.

Strategically, the announcement draws a direct line to Nvidia's DGX Station concept, and the contrast is instructive. AMD is stitching together its Threadripper PRO host, its Instinct accelerators, and ROCm, while Nvidia sells integrated Grace and Blackwell systems wrapped in CUDA. Tom's Hardware reported on September 4 that AMD has not named any OEM partners, has not set a price, and has not announced a release date, which leaves the Halo Station as a reference point rather than a purchasable product. The component math from that report is sobering: the 9995WX alone can cost around $11,000 to $12,000, each MI350P is estimated near $20,000, and 2 TB of DDR5 sells for roughly $50,000, which is how a street price clears $100,000 before storage and cooling are considered.

Why It Matters

The Halo Station matters beyond AMD's own product line because it marks how thoroughly the boundary between workstation and rack server is dissolving. For years, running frontier-scale models meant renting time in a data center or buying multi-rack clusters. AMD said on September 4 that a single deskside system, using the same Instinct accelerators found in cloud deployments, can hold and run models that until recently were firmly in data center territory. The company framed it as a new class of workstation that brings supercomputer-class compute to individual developers.

The timing is strategic as well. IFA 2026 has become a battleground for local AI hardware, with PC makers pushing NPU-equipped laptops and Nvidia showing edge silicon and partner devices aimed at on-device inference. AMD's counterargument is capacity: rather than squeezing smaller models onto a laptop, it offers a machine with the memory and bandwidth to skip the cloud for the largest open-weight models. That positions AMD against Nvidia not only in the AI PC category but at the edge of the data center, where the DGX Station has defined the premium local-AI workstation segment. The bandwidth figure helps explain the approach; each MI350P moves data at up to 4 TB/s, a rate AMD says is roughly 14 times what any LPDDR5X system memory can deliver.

Yet the caveats are real. AMD's "most powerful workstation in the world" label is a manufacturer claim without published comparative benchmarks, as Igor's Lab reported on September 5. The absence of pricing, availability, and partners means the product exists only as a demonstration, and the estimated cost puts it far beyond the budgets of the individual developers AMD name-checked from the stage. For all the enthusiasm around local AI, the Halo Station may end up mattering more as a signal of AMD's direction than as a machine that many people actually buy.

Next Up

The immediate question is who will actually build and sell the Halo Station. Tom's Hardware reported on September 4 that the design appears to be a system AMD's OEM partners will ultimately manufacture and ship, and that none have been announced. The obvious candidates are the workstation vendors that already build around Threadripper PRO, such as Lenovo with its ThinkStation line, Dell, and the system integrators that serve the AI development market. A fully loaded Lenovo ThinkStation P8, a comparable Threadripper PRO workstation with 2 TB of DDR5 and dual Blackwell accelerators, currently lists around $334,000, which offers a sense of the price bracket this machine is entering.

The chassis shown at IFA has room for only two liquid-cooled accelerators, so it is unclear whether the expansion to four cards will arrive in the same enclosure or require larger OEM systems. Igor's Lab raised that exact question on September 5, noting that the four-GPU path could demand a bigger chassis and more aggressive cooling than the tower on display. Watch for partner announcements in the coming months, along with any word on whether 576 GB is a distant roadmap item or a near-term option.

On the software side, the bet is on ROCm, AMD's open-source GPU compute stack, which must deliver a smooth experience for the PyTorch and vLLM workloads developers run if the hardware is to matter. AMD has spent years narrowing the software gap with CUDA, and the Halo Station is partly an argument that ROCm has matured enough to stand beside Nvidia for serious local work. With IFA running through September 8, more local-AI announcements are likely before the show closes, and the Halo Station gives AMD a centerpiece to measure them against. Whether it becomes a shipping product or remains a statement of intent, it has already changed the conversation about what a workstation can hold.