AMD’s Threadripper Halo Station: Spec Sheet, HBM Math, and the DGX Station Bet It Is Making

AMD opened IFA 2026 on September 4 with a deskside box aimed at people who already argue about HBM budgets: the Threadripper Halo Station. Jack Huynh called it the most powerful workstation in the world and said it can run models with more than a trillion parameters locally. I am not buying the slogan. I am buying the bill of materials — and the architectural bet against NVIDIA’s coherent DGX Station class.

This is a create-only read of what reporters on the floor documented today. AMD has not published a full SKU sheet, price, or availability window. Treat every “path to four” line as design intent until OEM configs land.

Spec sheet (show unit + platform class)

Primary coverage I used: ServeTheHome, StorageReview, TechPowerUp, Tom’s Hardware.

What the HBM math actually says

A trillion-parameter claim is a capacity claim until someone publishes tokens/sec and a harness. Rough weight floors:

So the four-card 576 GB story is the one that makes the “trillion-parameter local” line coherent at aggressive quantization. The two-card 288 GB show unit is a different product for most dense weights — closer to “very large MoE / heavy quant / multi-model deskside” than a full 1T FP8 resident set. StorageReview makes the same cut: ~500 GB for 1T at 4-bit, which wants the four-card pool, with the 2 TB host DDR5 and NVMe behind it for everything else.

Bandwidth is the other half. At ~4 TB/s HBM3E per MI350P, one card is in a different class from LPDDR5X “AI PC” unified memory. That is why this box sits above AMD’s Ryzen AI Halo / Max ladder and next to rack-adjacent Instinct iron, not next to a Strix/Kraken mini PC.

The DGX Station fight

NVIDIA’s recent deskside answer (GB300-class DGX Station / MSI XpertStation WS300 in StorageReview’s framing) bets on one big accelerator + Grace CPU sharing a large coherent memory pool over NVLink-C2C. AMD’s answer is the opposite:

• Discrete PCIe MI350P cards instead of a single coherent GPU+CPU package
• x86 Threadripper PRO host instead of Arm Grace (toolchain and driver story many of us already run)
• Scale by adding cards (2 → 4) instead of buying a denser single module
• Split pools: HBM on the Instincts, big DDR5 on the host — not one unified fabric

If your workload loves a single address space and NVLink-C2C bandwidth, the NVIDIA coherent pool still wins on paper. If your workload is ROCm + multi-GPU PCIe serving, or you refuse an Arm deskside toolchain, Halo Station is the first AMD-branded deskside that speaks that dialect with MI350-class HBM.

Nobody has published a public Halo Station tok/s table yet. Until then, the honest comparison is architectural, not a leaderboard.

Power, noise, and what was missing on stage

Two 600 W accelerators plus a ~350 W CPU is on the order of 1.5 kW of silicon before the rest of the box. Huynh’s liquid-cooling pitch (“stay super quiet”) is not marketing fluff at that TDP — it is the only way this stays a desk product instead of a lab cart. The show unit used per-device closed loops, not a shared custom loop, which ServeTheHome reads as COTS assembly more than a sealed appliance. Fine for a reference design. Awkward if you expected a sealed DGX-like brick with one service contract.

Also missing today:

• Named OEM / channel (direct AMD vs system builder)
• Interconnect detail between the four MI350Ps (PCIe topology, any Infinity Fabric / XGMI story)
• Storage layout, NIC (show unit had an SFP-class card — good, unspecified)
• Software image (ROCm version pin, Windows vs Linux, whether Project Zenith-class tooling shows up here or only on Ryzen AI Halo)
• Price. Street math on TR PRO 9995WX + 2–4× MI350P is “enterprise quote,” not Newegg.

How I would actually evaluate one

1. Demand the four-card BOM in writing. If the ship SKU is two-card only, re-rate the product as a 288 GB HBM workstation, not a 1T weight box.
2. Measure multi-GPU serving, not a single-card demo. vLLM / SGLang / llama.cpp tensor-parallel across PCIe MI350Ps is the workload that decides whether “path to four” matters.
3. A/B against a coherent DGX Station class box on the same model and quant. Same tokenizer, same batch, same context. Record tok/s, TTFT, and PCIe traffic. Coherence vs discrete is the real question.
4. Budget wall power and cooling. 1.5 kW-class silicon needs a circuit plan before a purchase order.
5. Pin ROCm. Deskside Instinct only helps if your stack’s day-0 wheel matches the driver AMD ships on the image.

Where it sits in AMD’s ladder

That ladder is the strategic point. AMD is no longer answering NVIDIA only in the rack. It is answering Spark-class and Station-class deskside products with named HBM capacity.

Bottom line

Threadripper Halo Station is the first AMD deskside I would put on the same evaluation sheet as a GB300-class DGX Station — not because the keynote said “trillion parameters,” but because 144 GB HBM3E × N at MI350P bandwidth is a real capacity and bandwidth story for local and air-gapped inference. The open questions are boring and decisive: four-card ship config, multi-GPU interconnect behavior, ROCm image, and price. Until those land, treat today as a serious reference design announcement, not a buy.

I will revisit when an OEM posts a configure-to-order page with a wattage rating and a four-card option you can actually order.

Anpoo Sivanadi, Staff Software Engineer