AMD Halo Station: Separate the Memory Pools
HARDWARE & DESIGN
AMD Halo Station: Separate the Memory Pools
Updated October 17, 2026: manufacturer specifications were checked again against AMD’s Halo Station page. Memory sums and bandwidth peaks are not measured inference performance. The four-bit example below now distinguishes raw weights from the complete runtime requirement. Third-party stage photographs have been removed.
A large total memory figure is not one uniform accelerator pool. Halo Station combines HBM and system memory with different paths to the processors. Before using the trillion-parameter example to plan a purchase, check where the weights would sit and what the interconnect must move.
In this article
So here is the short answer up front. This is AMD's DGX Station rival: a 96-core Threadripper PRO plus up to four Instinct MI350P cards, up to 2.6TB of combined memory and up to 16.4TB/s of total bandwidth, aiming at trillion-parameter models run locally. It is a prototype, it ships through OEM partners sometime in 2027, and there is no official price or date yet. Everything below carries that prototype caveat, and I will redo this post the moment real OEM pricing and independent numbers land.
The IFA prototype
September 4, 2026, Berlin, IFA opening keynote. Huynh, AMD's SVP and GM of Computing and Graphics, framed it as personal AI moving from single prompts to agents that run all day. The Halo Station is the big brother of the Ryzen AI Halo mini-PC box from earlier this year, aimed at the other end of the market: researchers and small teams who are currently renting cloud GPUs by the hour and hating the bill.
AMD's own page calls it a prototype outright and says "Coming in 2027". No partners named yet, no SKU list, no OS confirmed. ServeTheHome's read matches mine: the demo box is assembled from commercial parts, closed-loop radiators per processor rather than a shared loop, an ASUS Pro WS WRX90E-SAGE SE board and an unidentified SFP networking card inside. That is encouraging in one way, since COTS hardware can ship fast, and it also means the final OEM boxes may look different from the stage unit.
Components in the demo configuration
| Part | Halo Station spec |
|---|---|
| CPU | Threadripper PRO 9995WX, Zen 5 "Shimada Peak", 96 cores / 192 threads, up to 5.4 GHz boost, 384MB L3, 350W |
| Accelerators | 2x Instinct MI350P base (288GB HBM3e), path to 4x (576GB). Each card: 144GB HBM3e at 4 TB/s, 128 CDNA4 CUs, up to 600W board power, configurable down to 450W |
| System memory | Up to 2TB DDR5 RDIMM, 8-channel, up to DDR5-6400 |
| Combined memory | Up to 2.6TB (2TB DDR5 + 576GB HBM3e, 4-card config) |
| Bandwidth | Up to 16 TB/s GPU aggregate + up to 410 GB/s RDIMM = up to 16.4 TB/s total |
| Cooling | Liquid cooling on CPU and each accelerator (first liquid-cooled MI350P seen; launch cards were air-cooled for servers) |
| Platform | 128x PCIe Gen5 lanes, WRX90 board in demo unit, SFP networking card |
| Software | To be confirmed. Linux-based stack likely; Windows possible. Microsoft showed its Project Zenith dev toolkit on stage |
The CPU is last year's flagship workstation chip, and that is fine. The 9995WX brings 8 channels of DDR5 and 128 PCIe Gen5 lanes, which is exactly what a four-accelerator box needs. The newsworthy part is the MI350P in liquid-cooled form. The card launched in May as a cut-down CDNA4 part for PCIe servers, air-cooled, and a tower with no forced-air server chassis needed something else. Hence the water blocks.
Separate HBM and system-memory pools
Here is where I put it next to the boxes I already covered on this blog, using the same numbers as those posts so nothing drifts. The chart below stacks system RAM under accelerator HBM for all four.
The headline math: Halo at 2,624GB vs DGX Station at 748GB. That divides out to about 3.5x, and AMD claims 3.4x on its page, so the arithmetic roughly checks against their own footnote config. What the chart also shows is the shape of the two designs. Halo piles up a huge 2TB slab of slower DDR5 next to the fast HBM, while DGX Station keeps a smaller 496GB LPDDR5X pool but wires it coherently with the GPU over a 900 GB/s link. Notebookcheck made the same point I would: these are different memory architectures, so raw gigabytes flatter the Halo side. For pure weight capacity the Halo wins by a mile; for how smoothly a model straddling both pools actually runs, coherence matters and AMD has not shown that part yet.
Peak bandwidth versus measured inference
Advertised memory bandwidth is an upper-bound input, not a token-speed measurement. Adding four cards’ local bandwidth and system RAM bandwidth does not create one uniformly accessible memory pool. A real comparison needs the same model, precision, context, batch size and software, plus measurements of communication between the cards. These peak specifications do not establish a speed ranking against DGX Station or Mac Studio.
The raw trillion-parameter calculation
The raw arithmetic for one trillion parameters at four bits is 1,000,000,000,000 × 4 ÷ 8 = 500,000,000,000 bytes: 500 decimal GB, about 465.7 GiB. Against the advertised 576 GB HBM total, that leaves 76 nominal GB before quantization metadata, KV cache, runtime buffers and communication overhead. It does not prove the model will run, and it does not promise any context length or throughput. The two-card demo’s 288 GB is below that raw weight requirement. CPU RAM offload is another execution path with different performance constraints.
Price estimates and power requirements
AMD announced no price, and anyone stating one as fact is guessing. The press estimates cluster at $100,000 to $150,000, and Tom's Hardware showed the component-level reasoning: a 9995WX streets around $11,000 to $12,000, each MI350P is estimated near $20,000, and 2TB of DDR5 costs about $50,000 right now in the middle of a RAM shortage. Add chassis, cooling, storage, power and OEM margin and six figures is not a prediction, it is arithmetic. Power is the sleeper issue. Four cards at up to 600W each plus a 350W CPU plus memory and pumps pushes past 3kW at full load by Notebookcheck's estimate, which is server-room territory and possibly a 20-amp circuit conversation in North America. That may be why the stage unit showed two cards. A two-card box sips less, costs far less, and still out-memories everything short of a DGX.
Workloads that need the configuration
AMD's FAQ names researchers, model developers, simulation users, creators with big generative models, and shops that must keep IP on premises. That reads right to me, with one addition: teams currently paying cloud inference all day. AMD's page says it out loud, keep what you would otherwise spend on cloud inference, and for a lab burning thousands a month on rented HBM, a deskside box starts looking like capital expenditure with a payback schedule rather than a toy. The flip side is the software question, which is open. Instinct's server stack is real, but a workstation box lives or dies on day-one drivers, ROCm stability on the exact OEM config, and whether Windows is a first-class option. Microsoft's stage cameo suggests Windows matters here, yet nothing is confirmed, and I would not pre-order anything until an OEM puts an OS and a support page next to the price.
The maximum configuration’s 576GB HBM and 2TB DDR5 need to be read as separate pools. The demo showed two accelerators, while the largest numbers require four. OEM price, supported software, communication performance and the power installation were unresolved at the analysis date; component-cost estimates do not establish demand for a finished product.
Halo Station targets a larger memory budget than the smaller local machines linked here, but HBM and system RAM still need separate accounting. The next useful evidence is an OEM configuration with a complete price, supported operating system and matched inference test. The interconnect is central to that test because moving weights can erase a capacity advantage.
Reference: Xiaomi AI Cube
Reference: Mac mini RAM question
Reference: M5 Ultra Mac Studio
Reference: 2026 value ranking
Sources: AMD Halo Station product page (specs, footnotes HALO-02/SHP-76); ServeTheHome IFA report (COTS build, board, NIC, MI350P background); The Register (price/power estimates, Kimi K3 offload note, PCIe bottleneck flag); Tom's Hardware (component price math, 9995WX street pricing, Lenovo P8 analogy); TechPowerUp (keynote details, 5.4 GHz boost, 600W TBP); Notebookcheck (3kW estimate, coherent-memory caveat, GB300 salvage context). Photos: Tom's Hardware, TechPowerUp. Charts: author's redraws of vendor-reported specs. Related on this blog: 2026 local AI value ranking, M5 Ultra Mac Studio, Qwen3.8 35B-A3B Mac mini, Xiaomi AI Cube.
Comments
Post a Comment